fix(windows/gamestream): create the virtual display on Windows (Moonlight black screen)

The GameStream video path (open_gs_virtual_source) ran the Linux compositor-
detection state machine on every platform. On Windows detect_active_session()
returns None and vdisplay::detect() bails ("could not detect compositor ...
XDG_CURRENT_DESKTOP=''"), killing the video thread right after RTSP PLAY — so a
Moonlight client paired, negotiated, then black-screened and dropped.

The native punktfunk/1 path already guards this (resolve_compositor returns a
placeholder Compositor on Windows, since vdisplay::open ignores the compositor
arg there and always uses the pf-vdisplay IddCx backend). Mirror that guard in
the GameStream path: short-circuit to a placeholder on Windows, keep the Linux
session detection (apply_session_env/apply_input_env) under cfg(not(windows)).

Validated live: Moonlight -> this box now creates the pf-vdisplay virtual
monitor, attaches the IDD-push ring, and NVENC streams 5120x1440@240.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.cargo/audit.toml

@@ -13,11 +13,25 @@
 
 [advisories]
 ignore = [
-    # rsa "Marvin Attack" — a timing sidechannel in RSA *decryption* (PKCS#1 v1.5 padding oracle).
+    # rsa "Marvin Attack" (RUSTSEC-2023-0071): a timing side-channel in the rsa crate's variable-time
-    # There is NO fixed rsa release (the constant-time rewrite is still unreleased upstream), and rsa
+    # modular exponentiation of the SECRET exponent. IMPORTANT — this affects the RSA private-key op in
-    # is required for GameStream/Moonlight pairing. Crucially, the host uses rsa ONLY for PKCS#1 v1.5
+    # general, INCLUDING signing (m^d mod n), which the host DOES perform (gamestream/pairing.rs
-    # SIGNING / VERIFYING (gamestream/cert.rs + gamestream/pairing.rs: SigningKey / VerifyingKey /
+    # `signing_key.sign(&serversecret)`). It is NOT, as an earlier version of this note wrongly claimed,
-    # Signer / Verifier) — it never performs RSA decryption, which is the operation Marvin targets.
+    # limited to decryption — so "the vulnerable path isn't exercised" is false; signing exercises it.
-    # So the vulnerable code path is not exercised. Revisit if a fixed rsa ships or we add RSA decrypt.
+    # We accept it because the attack is not practically reachable here, NOT because the path is unused:
+    #   * No RSA decryption / PKCS#1v1.5 padding oracle exists anywhere (every `decrypt` in the tree is
+    #     AES/AES-GCM), so the classic Bleichenbacher/Marvin chosen-ciphertext oracle is absent.
+    #   * The only signed message (`serversecret`) is HOST-generated random, never attacker-chosen — so
+    #     there's no adaptive chosen-input probing (the lever remote RSA-timing key recovery needs); and
+    #     signing is gated behind the operator-entered pairing PIN, ONE signature per ceremony (a
+    #     repeated phase-3 is rejected — gamestream/pairing.rs — to deny a passive timing-sample harvester).
+    #   * GameStream is OFF by default (bare `serve` is native-only); the secure native QUIC plane uses
+    #     rustls' constant-time backend, NOT the rsa crate. RSA is touched only on the opt-in,
+    #     trusted-LAN GameStream/Moonlight pairing handshake. Moonlight mandates RSA-2048, so the
+    #     GameStream identity cannot move to Ed25519/ECDSA (only the native identity could, and it
+    #     already avoids the rsa crate).
+    # There is NO fixed rsa release (the constant-time rewrite is still unreleased upstream). Revisit if:
+    # a constant-time rsa ships (then drop this), the host ever signs an attacker-chosen message with
+    # this key, or any RSA decryption / key-transport using the private key is added.
     "RUSTSEC-2023-0071",
 ]

.gitea/workflows/release.yml

@@ -207,10 +207,20 @@ jobs:
         # (Config/Punktfunk-macOS.entitlements) — mandatory for the Mac App Store.
         continue-on-error: true
         run: |
-          # Separate archive from the Developer ID one above: App Store needs a profile-signed
+          # Separate archive from the Developer ID one above: App Store needs a signed, entitled
-          # archive (manual signing), not the unsigned-then-codesign DMG path. Same App-Manager
+          # archive that -exportArchive can re-sign for distribution, not the unsigned-then-codesign
-          # ASC-key constraint as iOS/tvOS — MANUAL signing, NOT -allowProvisioningUpdates
+          # DMG path. Archive with AUTOMATIC signing (development). Why not a manually-specified
-          # (cloud signing the key can't do). Quit Xcode so it can't prune the dropped profile.
+          # profile (as this step used to do): the in-app license screens added a SwiftPM resource
+          # bundle (PunktfunkKit_PunktfunkKit), and a resource bundle is a product type that cannot
+          # carry a provisioning profile — a global PROVISIONING_PROFILE_SPECIFIER (here) or an
+          # sdk-scoped one (iOS/tvOS) lands on it and fails the archive ("does not support
+          # provisioning profiles"). Automatic signing assigns a profile only to the app and leaves
+          # the resource bundle (and the macOS-host macro plugins) alone, and bakes the sandbox
+          # entitlements in. No -allowProvisioningUpdates → it stays OFFLINE and never cloud-signs
+          # (the App-Manager ASC key can't), so the runner must have a macOS *development* profile
+          # for io.unom.punktfunk installed. DISTRIBUTION signing happens in the export step below
+          # (manual, via the plist). Quit Xcode so it can't prune the manually-installed App Store
+          # distribution profile that export needs.
           osascript -e 'tell application "Xcode" to quit' >/dev/null 2>&1 || true
           pkill -x Xcode 2>/dev/null || true
           PROFILE="Punktfunk macOS App Store Distribution"
@@ -218,11 +228,10 @@ jobs:
             -project "$PROJECT" -scheme Punktfunk \
             -destination 'generic/platform=macOS' \
             -archivePath "$RUNNER_TEMP/Punktfunk-macos-appstore.xcarchive" \
+            -skipMacroValidation -skipPackagePluginValidation \
             MARKETING_VERSION="$VERSION" CURRENT_PROJECT_VERSION="$BUILD_NUM" \
-            CODE_SIGN_STYLE=Manual \
+            CODE_SIGN_STYLE=Automatic \
-            CODE_SIGN_IDENTITY="Apple Distribution" \
+            DEVELOPMENT_TEAM="$TEAM_ID"
-            DEVELOPMENT_TEAM="$TEAM_ID" \
-            PROVISIONING_PROFILE_SPECIFIER="$PROFILE"
           cat > "$RUNNER_TEMP/export-macos-appstore.plist" <<EOF
           <?xml version="1.0" encoding="UTF-8"?>
           <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
@@ -252,35 +261,27 @@ jobs:
         # Best-effort until the App Store Connect app record for io.unom.punktfunk exists.
         continue-on-error: true
         run: |
-          # MANUAL App Store signing: the local (valid) Apple Distribution identity + the App
+          # Archive with AUTOMATIC signing (development) — see the macOS App Store step for the full
-          # Store provisioning profile. NOT -allowProvisioningUpdates — with an App-Manager-role
+          # rationale. The SwiftPM resource bundle (PunktfunkKit_PunktfunkKit, added with the in-app
-          # ASC key that forces Xcode's CLOUD-managed signing, which the role can't do ("Cloud
+          # license screens) builds for iphoneos, so even the sdk-scoped PROVISIONING_PROFILE_SPECIFIER
-          # signing permission error"). The profile must be installed on the runner under
+          # this step used to set matched it and failed the archive ("does not support provisioning
-          # ~/Library/Developer/Xcode/UserData/Provisioning Profiles/ (install it once with
+          # profiles"). Automatic signing profiles only the app and leaves the resource bundle (and
-          # Xcode.app quit, or it prunes the manually-dropped distribution profile).
+          # the macOS-host macro plugins) alone. No -allowProvisioningUpdates → OFFLINE, never
-          # A running Xcode.app prunes unrecognized profiles from that dir — quit it so the App
+          # cloud-signs (the App-Manager ASC key can't), so the runner needs an iOS *development*
-          # Store profile survives this build; headless xcodebuild doesn't need the GUI app.
+          # profile for io.unom.punktfunk installed. DISTRIBUTION signing is the export step below
+          # (manual, via the plist). A running Xcode.app prunes unrecognized profiles — quit it so the
+          # manually-installed App Store distribution profile survives for export.
           osascript -e 'tell application "Xcode" to quit' >/dev/null 2>&1 || true
           pkill -x Xcode 2>/dev/null || true
           PROFILE="Punktfunk iOS App Store Distribution"
-          # Scope signing to the iOS device SDK via an xcconfig — see the tvOS step below for the
-          # full rationale. A global (CLI) profile specifier would also be forced onto the shared
-          # macOS-host SwiftPM macro plugins, which reject it and fail the archive; [sdk=iphoneos*]
-          # in an xcconfig lands it on the app/framework slices only.
-          SIGN_XCCONFIG="$RUNNER_TEMP/sign-ios.xcconfig"
-          cat > "$SIGN_XCCONFIG" <<XCCONF
-          CODE_SIGN_STYLE = Manual
-          DEVELOPMENT_TEAM = $TEAM_ID
-          CODE_SIGN_IDENTITY[sdk=iphoneos*] = Apple Distribution
-          PROVISIONING_PROFILE_SPECIFIER[sdk=iphoneos*] = $PROFILE
-          XCCONF
           DEVELOPER_DIR="$XCODE_DEV_DIR" xcodebuild archive \
             -project "$PROJECT" -scheme Punktfunk-iOS \
             -destination 'generic/platform=iOS' \
             -archivePath "$RUNNER_TEMP/Punktfunk-ios.xcarchive" \
             -skipMacroValidation -skipPackagePluginValidation \
-            -xcconfig "$SIGN_XCCONFIG" \
+            MARKETING_VERSION="$VERSION" CURRENT_PROJECT_VERSION="$BUILD_NUM" \
-            MARKETING_VERSION="$VERSION" CURRENT_PROJECT_VERSION="$BUILD_NUM"
+            CODE_SIGN_STYLE=Automatic \
+            DEVELOPMENT_TEAM="$TEAM_ID"
           cat > "$RUNNER_TEMP/export-appstore.plist" <<EOF
           <?xml version="1.0" encoding="UTF-8"?>
           <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
@@ -312,33 +313,24 @@ jobs:
         # on the runner (xcodebuild -downloadPlatform tvOS).
         continue-on-error: true
         run: |
-          # Same manual App Store signing as iOS (the App-Manager ASC key can't cloud-sign).
+          # Archive with AUTOMATIC signing (development) — see the macOS App Store step. The SwiftPM
+          # resource bundle (PunktfunkKit_PunktfunkKit) builds for appletvos and rejected the
+          # sdk-scoped profile this step used to set; Automatic signing profiles only the app and
+          # leaves the resource bundle + the macOS-host macro plugins (OnceMacro/SwizzlingMacro/
+          # AssociationMacro) alone. No -allowProvisioningUpdates → OFFLINE, never cloud-signs (the
+          # App-Manager ASC key can't), so the runner needs a tvOS *development* profile for
+          # io.unom.punktfunk installed. DISTRIBUTION signing is the export step below (manual, plist).
           osascript -e 'tell application "Xcode" to quit' >/dev/null 2>&1 || true
           pkill -x Xcode 2>/dev/null || true
           PROFILE="Punktfunk tvOS App Store Distribution"
-          # Scope signing to the tvOS device SDK via an xcconfig. A global (CLI) profile specifier
-          # hits EVERY target, including the shared SwiftPM macro plugins (OnceMacro/SwizzlingMacro/
-          # AssociationMacro) which build for the macOS host and reject a provisioning profile
-          # ("<macro> does not support provisioning profiles"), failing the archive. Conditionals
-          # work only in an xcconfig (xcodebuild mis-parses a CLI "SETTING[sdk=..]=val"), and a
-          # command-line -xcconfig outranks target settings, so [sdk=appletvos*] puts the profile on
-          # the app/framework slices only — the macosx-host macros get nothing. (The macOS archive
-          # above is immune: its host-SDK macros are CODE_SIGNING_ALLOWED=NO, so a global specifier
-          # is ignored there.)
-          SIGN_XCCONFIG="$RUNNER_TEMP/sign-tvos.xcconfig"
-          cat > "$SIGN_XCCONFIG" <<XCCONF
-          CODE_SIGN_STYLE = Manual
-          DEVELOPMENT_TEAM = $TEAM_ID
-          CODE_SIGN_IDENTITY[sdk=appletvos*] = Apple Distribution
-          PROVISIONING_PROFILE_SPECIFIER[sdk=appletvos*] = $PROFILE
-          XCCONF
           DEVELOPER_DIR="$XCODE_DEV_DIR" xcodebuild archive \
             -project "$PROJECT" -scheme Punktfunk-tvOS \
             -destination 'generic/platform=tvOS' \
             -archivePath "$RUNNER_TEMP/Punktfunk-tvos.xcarchive" \
             -skipMacroValidation -skipPackagePluginValidation \
-            -xcconfig "$SIGN_XCCONFIG" \
+            MARKETING_VERSION="$VERSION" CURRENT_PROJECT_VERSION="$BUILD_NUM" \
-            MARKETING_VERSION="$VERSION" CURRENT_PROJECT_VERSION="$BUILD_NUM"
+            CODE_SIGN_STYLE=Automatic \
+            DEVELOPMENT_TEAM="$TEAM_ID"
           cat > "$RUNNER_TEMP/export-tvos.plist" <<EOF
           <?xml version="1.0" encoding="UTF-8"?>
           <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">

CLAUDE.md

@@ -144,11 +144,25 @@ Low-latency desktop/game streaming stack, Linux-first, with a shared Rust protoc
    `test-loopback.sh` (Swift client vs synthetic punktfunk1-hosts on loopback — runs on macOS;
    includes the pairing ceremony + `--require-pairing` gate),
    `RemoteFirstLightTests` (full pipeline over the LAN). See
-   [`clients/apple/README.md`](clients/apple/README.md). **Stage 2 presenter**
+   [`clients/apple/README.md`](clients/apple/README.md). **Stage 2 presenter is now the DEFAULT**
-   (`VTDecompressionSession` + `CAMetalLayer`) is built and live-validated on glass behind the opt-in
+   (stage-1 is the Metal-unavailable / DEBUG fallback): explicit `VTDecompressionSession` decode →
-   `punktfunk.presenter` flag (~11 ms p50 capture→present), to become the default after a few
+   `CAMetalLayer`, presented from the hosting view's **main-runloop `CADisplayLink`** (`renderTick` pops
-   resolution/HDR checks. Next: make stage 2 the default, glass-to-glass numbers via
+   the newest ready frame per vsync; macOS `displaySyncEnabled = false` is the real fullscreen-judder fix,
-   `tools/latency-probe`, iOS/iPadOS/tvOS variants.
+   ~11 ms p50). *(An off-main `CAMetalDisplayLink` and an off-main blocking-render present thread were
+   both tried and reverted — both measured slower on macOS and iPad.)* **HDR fixed**
+   (`design/apple-stage2-presenter.md`): the "too bright" bug was a missing reference-white anchor — the
+   fix keeps the PQ-passthrough shader and adds `CAEDRMetadata.hdr10(…, opticalOutputScale: 203)` +
+   `wantsExtendedDynamicRangeContent` on the layer (on all platforms; the old `#if os(macOS)` guard left
+   iOS/tvOS EDR half-engaged), routing the 0xCE mastering metadata to the layer (via `setHdrMeta`) instead
+   of a never-composited source buffer. **Mid-session SDR↔HDR** is handled: `render` reconciles the layer
+   per-frame from the decoded `frame.isHDR` (per-mode pixel format `bgra8`/`rgba16Float`), so a game
+   entering HDR mid-stream just reconfigures (last 0xCE grade cached + re-applied; pump drains 0xCE
+   unconditionally). **4:4:4 added**: decode format is a 2×2 `(chroma, HDR)` matrix
+   (`420v/x420/444v/x444`, all biplanar so the shaders are unchanged), advertised (`VIDEO_CAP_444`) only
+   behind a **hardware-required `VTDecompressionSession` probe** (`Stage444Probe`, validated on M3) with a
+   Settings opt-out + a bounded pump backstop for an undecodable 4:4:4 session. *HDR brightness + 4:4:4
+   still need on-glass validation (Windows-HDR / `PUNKTFUNK_444` host).* Next: glass-to-glass numbers via
+   `tools/latency-probe`.
    **Linux stage 1 done, first light 2026-06-12** (`clients/linux`, binary
    `punktfunk-client`): GTK4/libadwaita shell linking `punktfunk-core` directly (no C ABI;
    `NativeClient` is now `Sync` — mutexed plane receivers), mDNS host list, TOFU + SPAKE2

Cargo.lock

@@ -2331,6 +2331,17 @@ version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "521739c6d2bac4aa25192232afe6841231376b2b26d4d9fae5ecf8ca5772e441"
 
+[[package]]
+name = "num-derive"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed3955f1a9c7c0c15e092f9c887db08b1fc683305fdf6eb6684f22555355e202"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
 [[package]]
 name = "num-integer"
 version = "0.1.46"
@@ -2839,12 +2850,14 @@ dependencies = [
  "tracing",
  "tracing-subscriber",
  "ureq",
+ "usbip-sim",
  "utoipa",
  "utoipa-axum",
  "utoipa-scalar",
  "wasapi",
  "wayland-backend",
  "wayland-client",
+ "wayland-protocols",
  "wayland-protocols-misc",
  "wayland-protocols-wlr",
  "wayland-scanner",
@@ -4236,6 +4249,17 @@ dependencies = [
  "serde",
 ]
 
+[[package]]
+name = "usbip-sim"
+version = "0.8.0"
+dependencies = [
+ "log",
+ "num-derive",
+ "num-traits",
+ "serde",
+ "tokio",
+]
+
 [[package]]
 name = "utf8_iter"
 version = "1.0.4"

Cargo.toml

@@ -3,6 +3,7 @@ resolver = "2"
 members = [
     "crates/punktfunk-core",
     "crates/punktfunk-host",
+    "crates/punktfunk-host/vendor/usbip-sim",
     "crates/pf-driver-proto",
     "clients/probe",
     "clients/linux",
@@ -11,6 +12,8 @@ members = [
     "tools/latency-probe",
     "tools/loss-harness",
 ]
+# Standalone PoC (built on its own; pulls usbip/tokio/libusb we don't want in the workspace).
+exclude = ["packaging/linux/steam-deck-gadget/usbip-poc"]
 
 [workspace.package]
 version = "0.3.0"

api/openapi.json

@@ -10,7 +10,7 @@
       "name": "MIT OR Apache-2.0",
       "identifier": "MIT OR Apache-2.0"
     },
-    "version": "0.0.1"
+    "version": "0.3.0"
   },
   "paths": {
     "/api/v1/clients": {
@@ -1354,6 +1354,14 @@
         "type": "object",
         "description": "Arm-native-pairing request body.",
         "properties": {
+          "fingerprint": {
+            "type": [
+              "string",
+              "null"
+            ],
+            "description": "Optional: bind the window to ONE device fingerprint (hex SHA-256, e.g. from a pending knock).\nWhen set, only a pairing attempt from that fingerprint consumes the window — so an unpaired\nLAN peer can neither pair nor burn a window armed for a specific device (security-review #9).\nOmit for an unbound window (any device may use the PIN — trusted-LAN only).",
+            "example": "9f86d081884c7d659a2feaa0c55ad015a3bf4f1b2b0b822cd15d6c15b0f00a08"
+          },
           "ttl_secs": {
             "type": [
               "integer",

clients/android/app/src/main/kotlin/io/unom/punktfunk/ConnectScreen.kt

@@ -175,9 +175,9 @@ fun ConnectScreen(settings: Settings, onConnected: (Long) -> Unit) {
         status = "Connecting to $targetHost:$targetPort…"
         discovery.stop() // free the Wi-Fi radio before the stream session
         scope.launch {
-            // Advertise HDR only when this device's display can present it (else the host sends a
+            // Advertise HDR only when the user enabled it AND this device's display can present it
-            // proper SDR stream rather than PQ the panel would mis-tone-map).
+            // (else the host sends a proper SDR stream rather than PQ the panel would mis-tone-map).
-            val hdrEnabled = displaySupportsHdr(context)
+            val hdrEnabled = settings.hdrEnabled && displaySupportsHdr(context)
             // "Automatic" resolves to a concrete pad type from the connected controller's VID/PID
             // (Android exposes no controller-type enum) — parity with the Linux/Apple clients. An
             // explicit choice is passed through unchanged.
@@ -224,7 +224,7 @@ fun ConnectScreen(settings: Settings, onConnected: (Long) -> Unit) {
         status = null
         discovery.stop() // free the Wi-Fi radio before the (parked) stream session
         scope.launch {
-            val hdrEnabled = displaySupportsHdr(context)
+            val hdrEnabled = settings.hdrEnabled && displaySupportsHdr(context)
             val gamepadPref = Gamepad.resolvePref(settings.gamepad)
             // Pin the advertised fingerprint for a discovered host (defence against an impostor while
             // we wait); a manually-typed host has none, so trust-on-first-use.

clients/android/app/src/main/kotlin/io/unom/punktfunk/Settings.kt

@@ -14,6 +14,13 @@ data class Settings(
     val height: Int = 0,
     val hz: Int = 0,
     val bitrateKbps: Int = 0,
+    /**
+     * Advertise HDR (10-bit BT.2020 PQ) to the host. Default on, but only *effective* on a panel that
+     * can actually present HDR10 (see [displaySupportsHdr]) — on an SDR display HDR is never
+     * advertised regardless, so the host sends a proper 8-bit BT.709 stream rather than PQ the panel
+     * would mis-tone-map. Turning this off forces SDR even on a capable panel.
+     */
+    val hdrEnabled: Boolean = true,
     val compositor: Int = 0,
     val gamepad: Int = 0,
     /** Requested audio channel count: 2 (stereo), 6 (5.1) or 8 (7.1). The host clamps to what it
@@ -40,6 +47,7 @@ class SettingsStore(context: Context) {
         height = prefs.getInt(K_H, 0),
         hz = prefs.getInt(K_HZ, 0),
         bitrateKbps = prefs.getInt(K_BITRATE, 0),
+        hdrEnabled = prefs.getBoolean(K_HDR, true),
         compositor = prefs.getInt(K_COMPOSITOR, 0),
         gamepad = prefs.getInt(K_GAMEPAD, 0),
         audioChannels = prefs.getInt(K_AUDIO_CH, 2),
@@ -54,6 +62,7 @@ class SettingsStore(context: Context) {
             .putInt(K_H, s.height)
             .putInt(K_HZ, s.hz)
             .putInt(K_BITRATE, s.bitrateKbps)
+            .putBoolean(K_HDR, s.hdrEnabled)
             .putInt(K_COMPOSITOR, s.compositor)
             .putInt(K_GAMEPAD, s.gamepad)
             .putInt(K_AUDIO_CH, s.audioChannels)
@@ -68,6 +77,7 @@ class SettingsStore(context: Context) {
         const val K_H = "height"
         const val K_HZ = "hz"
         const val K_BITRATE = "bitrate_kbps"
+        const val K_HDR = "hdr_enabled"
         const val K_COMPOSITOR = "compositor"
         const val K_GAMEPAD = "gamepad"
         const val K_AUDIO_CH = "audio_channels"

clients/android/app/src/main/kotlin/io/unom/punktfunk/SettingsScreen.kt

@@ -94,6 +94,22 @@ fun SettingsScreen(initial: Settings, onChange: (Settings) -> Unit, onBack: () -
                 options = BITRATE_OPTIONS,
                 selected = s.bitrateKbps,
             ) { kbps -> update(s.copy(bitrateKbps = kbps)) }
+
+            // HDR is only meaningful on a panel that can present HDR10; on an SDR display the toggle
+            // is disabled (and HDR is never advertised regardless) so the host doesn't send PQ the
+            // panel would mis-tone-map. The capability is fixed for the device, so read it once.
+            val hdrCapable = remember { displaySupportsHdr(context) }
+            ToggleRow(
+                title = "HDR",
+                subtitle = if (hdrCapable) {
+                    "Stream 10-bit HDR (BT.2020 PQ) when the host supports it"
+                } else {
+                    "This display can't present HDR10 — streams stay SDR"
+                },
+                checked = s.hdrEnabled && hdrCapable,
+                enabled = hdrCapable,
+                onCheckedChange = { on -> update(s.copy(hdrEnabled = on)) },
+            )
         }
 
         SettingsGroup("Host") {
@@ -181,24 +197,31 @@ private fun SettingsGroup(title: String, content: @Composable ColumnScope.() ->
     }
 }
 
-/** A title + subtitle on the left, a Switch on the right. */
+/** A title + subtitle on the left, a Switch on the right. [enabled] greys out the whole row. */
 @Composable
 private fun ToggleRow(
     title: String,
     subtitle: String,
     checked: Boolean,
     onCheckedChange: (Boolean) -> Unit,
+    enabled: Boolean = true,
 ) {
+    // Dim the labels when disabled so the row reads as inactive (the Switch dims itself).
+    val labelAlpha = if (enabled) 1f else 0.38f
     Row(modifier = Modifier.fillMaxWidth(), verticalAlignment = Alignment.CenterVertically) {
         Column(Modifier.weight(1f)) {
-            Text(title, style = MaterialTheme.typography.bodyLarge)
+            Text(
+                title,
+                style = MaterialTheme.typography.bodyLarge,
+                color = MaterialTheme.colorScheme.onSurface.copy(alpha = labelAlpha),
+            )
             Text(
                 subtitle,
                 style = MaterialTheme.typography.bodySmall,
-                color = MaterialTheme.colorScheme.onSurfaceVariant,
+                color = MaterialTheme.colorScheme.onSurfaceVariant.copy(alpha = labelAlpha),
             )
         }
-        Switch(checked = checked, onCheckedChange = onCheckedChange)
+        Switch(checked = checked, onCheckedChange = onCheckedChange, enabled = enabled)
     }
 }
 

clients/android/app/src/main/kotlin/io/unom/punktfunk/StreamScreen.kt

@@ -1,6 +1,7 @@
 package io.unom.punktfunk
 
 import android.Manifest
+import android.content.pm.ActivityInfo
 import android.content.pm.PackageManager
 import android.view.SurfaceHolder
 import android.view.SurfaceView
@@ -102,6 +103,13 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
             it.systemBarsBehavior = WindowInsetsControllerCompat.BEHAVIOR_SHOW_TRANSIENT_BARS_BY_SWIPE
             it.hide(WindowInsetsCompat.Type.systemBars())
         }
+        // Lock to landscape while streaming — the host streams a landscape desktop, so pin the device
+        // there (either landscape direction is fine) and stop it rotating to portrait mid-session. The
+        // activity declares configChanges=orientation, so this re-lays out the surface in place without
+        // recreating the activity (no stream restart). On TV (fixed landscape) it's a harmless no-op.
+        // The prior request is captured and restored on the way out.
+        val priorOrientation = activity?.requestedOrientation
+        activity?.requestedOrientation = ActivityInfo.SCREEN_ORIENTATION_SENSOR_LANDSCAPE
         activity?.streamHandle = handle // route hardware keys to this session
         activity?.axisMapper = Gamepad.AxisMapper(handle) // route joystick axes
         // Host→client feedback (rumble + DualSense lightbar/LEDs); poll threads stopped before close.
@@ -114,6 +122,9 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
             activity?.streamHandle = 0L
             controller?.show(WindowInsetsCompat.Type.systemBars())
             window?.clearFlags(WindowManager.LayoutParams.FLAG_KEEP_SCREEN_ON)
+            // Release the landscape lock so the rest of the app follows the device/system again.
+            activity?.requestedOrientation =
+                priorOrientation ?: ActivityInfo.SCREEN_ORIENTATION_UNSPECIFIED
             // Leaving the stream: stop the mic + audio + decode threads and tear down the session.
             NativeBridge.nativeStopMic(handle)
             NativeBridge.nativeStopAudio(handle)
@@ -314,9 +325,11 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
 }
 
 /**
- * The live stats overlay — mirrors the Apple client's HUD. Reads the 10-double layout from
+ * The live stats overlay — mirrors the Apple client's HUD. Reads the 14-double layout from
  * [NativeBridge.nativeVideoStats]:
- * `[fps, mbps, latP50Ms, latP95Ms, latValid, skew, w, h, hz, dropped]`.
+ * `[fps, mbps, latP50Ms, latP95Ms, latValid, skew, w, h, hz, dropped, bitDepth, colorPrimaries,
+ * colorTransfer, chromaFormatIdc]`. The trailing four (present on a current native lib) describe the
+ * negotiated video feed and render as a codec/depth/colour/chroma line; older layouts just omit it.
  */
 @Composable
 internal fun StatsOverlay(s: DoubleArray, modifier: Modifier = Modifier) {
@@ -338,6 +351,14 @@ internal fun StatsOverlay(s: DoubleArray, modifier: Modifier = Modifier) {
             fontFamily = FontFamily.Monospace,
             fontSize = 12.sp,
         )
+        videoFeedLine(s)?.let { feed ->
+            Text(
+                feed,
+                color = Color.White,
+                fontFamily = FontFamily.Monospace,
+                fontSize = 12.sp,
+            )
+        }
         if (latValid) {
             val tag = if (skew) "" else " (same-host)"
             Text(
@@ -357,3 +378,31 @@ internal fun StatsOverlay(s: DoubleArray, modifier: Modifier = Modifier) {
         }
     }
 }
+
+/**
+ * Format the negotiated video-feed descriptor from the trailing four stats doubles
+ * `[bitDepth, colorPrimaries, colorTransfer, chromaFormatIdc]`, e.g.
+ * `HEVC · 10-bit · HDR (BT.2020 PQ) · 4:2:0`. Returns `null` on a pre-video-feed layout (< 14 doubles)
+ * so the overlay simply omits the line. The codes are CICP / H.273: transfer 16 = PQ, 18 = HLG (else
+ * SDR); primaries 9 = BT.2020, 1 = BT.709; chroma_format_idc 1 = 4:2:0, 2 = 4:2:2, 3 = 4:4:4. The
+ * Android decoder is always HEVC (`video/hevc`).
+ */
+private fun videoFeedLine(s: DoubleArray): String? {
+    if (s.size < 14) return null
+    val bitDepth = s[10].toInt()
+    val primaries = s[11].toInt()
+    val transfer = s[12].toInt()
+    val chromaIdc = s[13].toInt()
+    val depthLabel = if (bitDepth > 0) "$bitDepth-bit" else "8-bit"
+    val (dynamicRange, colorSpace) = when (transfer) {
+        16 -> "HDR" to "BT.2020 PQ"
+        18 -> "HDR" to "BT.2020 HLG"
+        else -> "SDR" to if (primaries == 9) "BT.2020" else "BT.709"
+    }
+    val chromaLabel = when (chromaIdc) {
+        3 -> "4:4:4"
+        2 -> "4:2:2"
+        else -> "4:2:0"
+    }
+    return "HEVC · $depthLabel · $dynamicRange ($colorSpace) · $chromaLabel"
+}

clients/android/app/src/test/kotlin/io/unom/punktfunk/screenshots/ShotScenes.kt

@@ -186,9 +186,11 @@ internal fun StreamScene() {
                 Brush.linearGradient(listOf(Color(0xFF2A1E5C), Color(0xFF0E1B3D), Color(0xFF06122B))),
             ),
     ) {
-        // [fps, mbps, latP50, latP95, latValid, skew, w, h, hz, dropped]
+        // [fps, mbps, latP50, latP95, latValid, skew, w, h, hz, dropped,
+        //  bitDepth, colorPrimaries, colorTransfer, chromaFormatIdc] — the last four = a 10-bit
+        //  BT.2020 PQ (HDR) 4:2:0 feed, so the HUD renders its video-feed line.
         StatsOverlay(
-            doubleArrayOf(238.0, 921.4, 1.3, 2.1, 1.0, 1.0, 5120.0, 1440.0, 240.0, 0.0),
+            doubleArrayOf(238.0, 921.4, 1.3, 2.1, 1.0, 1.0, 5120.0, 1440.0, 240.0, 0.0, 10.0, 9.0, 16.0, 1.0),
             Modifier.align(Alignment.TopStart).padding(12.dp),
         )
     }

clients/android/kit/src/main/kotlin/io/unom/punktfunk/kit/Gamepad.kt

@@ -50,15 +50,25 @@ object Gamepad {
     const val PREF_DUALSENSE = 2
     const val PREF_XBOXONE = 3
     const val PREF_DUALSHOCK4 = 4
+    const val PREF_STEAMCONTROLLER = 5
+    const val PREF_STEAMDECK = 6
 
     // USB vendor ids of the controllers we can identify by VID/PID.
     private const val VID_SONY = 0x054C
     private const val VID_MICROSOFT = 0x045E
+    private const val VID_VALVE = 0x28DE
 
     // Sony product ids. DualSense (PS5) and DualShock 4 (PS4) map to distinct host pad types.
     private val PID_DUALSENSE = setOf(0x0CE6, 0x0DF2)
     private val PID_DUALSHOCK4 = setOf(0x05C4, 0x09CC)
 
+    // Valve: Steam Deck built-in controller (0x1205); classic Steam Controller wired (0x1102) /
+    // dongle (0x1142). The host builds the virtual hid-steam pad; rich-input capture (paddles /
+    // trackpads / gyro) is out of scope on Android (no rich-input plane yet), so only the standard
+    // buttons + sticks reach the host for now — parity with the desktop type resolution.
+    private val PID_STEAMDECK = setOf(0x1205)
+    private val PID_STEAMCONTROLLER = setOf(0x1102, 0x1142)
+
     // Microsoft Xbox One / Series product ids (wired + the common Bluetooth/dongle revisions). All
     // behave like Xbox 360 on the host minus the glyph identity, so they share one pref byte.
     private val PID_XBOXONE = setOf(
@@ -82,6 +92,8 @@ object Gamepad {
             vid == VID_SONY && pid in PID_DUALSENSE -> PREF_DUALSENSE
             vid == VID_SONY && pid in PID_DUALSHOCK4 -> PREF_DUALSHOCK4
             vid == VID_MICROSOFT && pid in PID_XBOXONE -> PREF_XBOXONE
+            vid == VID_VALVE && pid in PID_STEAMDECK -> PREF_STEAMDECK
+            vid == VID_VALVE && pid in PID_STEAMCONTROLLER -> PREF_STEAMCONTROLLER
             else -> PREF_XBOX360
         }
     }

clients/android/kit/src/main/kotlin/io/unom/punktfunk/kit/NativeBridge.kt

@@ -103,9 +103,12 @@ object NativeBridge {
 
     /**
      * Drain ~1 s of live decode stats for the on-stream HUD, or `null` when no decode thread runs.
-     * Returns 10 doubles:
+     * Returns 14 doubles:
-     * `[fps, mbps, latP50Ms, latP95Ms, latValid, skewCorrected, width, height, refreshHz, framesDropped]`
+     * `[fps, mbps, latP50Ms, latP95Ms, latValid, skewCorrected, width, height, refreshHz, framesDropped,
-     * (the two flags are 1.0/0.0). Poll ~1 Hz; each call resets the measurement window.
+     * bitDepth, colorPrimaries, colorTransfer, chromaFormatIdc]`
+     * (the two flags are 1.0/0.0; the trailing four describe the negotiated video feed — bit depth
+     * 8/10, CICP primaries/transfer, and the HEVC chroma_format_idc 1=4:2:0 / 3=4:4:4). Poll ~1 Hz;
+     * each call resets the measurement window.
      */
     external fun nativeVideoStats(handle: Long): DoubleArray?
 

clients/android/native/src/feedback.rs

@@ -114,6 +114,11 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeNextHidout(
                 out[2..n].copy_from_slice(&effect);
                 n
             }
+            HidOutput::TrackpadHaptic { .. } => {
+                // Steam Controller trackpad-coil haptics — no Android equivalent; drop it (motor
+                // rumble already rides the universal 0xCA plane).
+                return -1;
+            }
         };
         n as jint
     })

clients/android/native/src/session.rs

@@ -409,11 +409,13 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeStopVideo(
 }
 
 /// `NativeBridge.nativeVideoStats(handle): DoubleArray?` — drain ~1 s of decode stats for the HUD.
-/// Returns 10 doubles
+/// Returns 14 doubles
-/// `[fps, mbps, latP50Ms, latP95Ms, latValid, skewCorrected, width, height, refreshHz, framesDropped]`
+/// `[fps, mbps, latP50Ms, latP95Ms, latValid, skewCorrected, width, height, refreshHz, framesDropped,
-/// (the two flags are 1.0/0.0), or `null` when no decode thread is running. Poll ~1 Hz from the UI;
+/// bitDepth, colorPrimaries, colorTransfer, chromaFormatIdc]`
-/// each call resets the measurement window. Not android-gated — pure `jni` + connector reads, so it
+/// (the two flags are 1.0/0.0; the trailing four describe the negotiated video feed — see below), or
-/// links on the host build too (Kotlin only ever calls it on device).
+/// `null` when no decode thread is running. Poll ~1 Hz from the UI; each call resets the measurement
+/// window. Not android-gated — pure `jni` + connector reads, so it links on the host build too
+/// (Kotlin only ever calls it on device).
 #[no_mangle]
 pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeVideoStats(
     env: JNIEnv,
@@ -431,7 +433,8 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeVideoStats(
             None => return std::ptr::null_mut(), // not streaming → no stats
         };
         let mode = h.client.mode();
-        let buf: [f64; 10] = [
+        let color = h.client.color;
+        let buf: [f64; 14] = [
             snap.fps,
             snap.mbps,
             snap.lat_p50_ms,
@@ -442,6 +445,14 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeVideoStats(
             mode.height as f64,
             mode.refresh_hz as f64,
             h.client.frames_dropped() as f64,
+            // Video-feed properties the host resolved at the handshake (Welcome): encode bit depth
+            // (8 / 10), the CICP colour primaries + transfer code points (Kotlin maps these to a
+            // colour-space / HDR label — transfer 16 = PQ, 18 = HLG ⇒ HDR), and the HEVC
+            // chroma_format_idc (1 = 4:2:0, 3 = 4:4:4). Static for the session unless renegotiated.
+            h.client.bit_depth as f64,
+            color.primaries as f64,
+            color.transfer as f64,
+            h.client.chroma_format as f64,
         ];
         let arr = match env.new_double_array(buf.len() as jsize) {
             Ok(a) => a,

clients/apple/Package.swift

@@ -24,6 +24,9 @@ let package = Package(
                 .copy("Resources/THIRD-PARTY-NOTICES.txt"),
                 .copy("Resources/LICENSE-MIT.txt"),
                 .copy("Resources/LICENSE-APACHE.txt"),
+                // Geist (SIL OFL 1.1) — the brand typeface, shared with punktfunk-website.
+                // Registered with Core Text at first use; see BrandFont.swift.
+                .copy("Resources/Fonts"),
             ],
             linkerSettings: [
                 // Rust staticlib system deps.

clients/apple/Punktfunk.xcodeproj/project.pbxproj

@@ -364,7 +364,7 @@
 				ENABLE_HARDENED_RUNTIME = YES;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_FILE = Config/Info.plist;
-				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunkempfänger";
+				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunk";
 				INFOPLIST_KEY_LSApplicationCategoryType = "public.app-category.utilities";
 				INFOPLIST_KEY_NSHumanReadableCopyright = "";
 				INFOPLIST_KEY_NSLocalNetworkUsageDescription = "Punktfunk connects directly to your punktfunk host on the local network to stream video, audio, and input.";
@@ -398,7 +398,7 @@
 				ENABLE_HARDENED_RUNTIME = YES;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_FILE = Config/Info.plist;
-				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunkempfänger";
+				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunk";
 				INFOPLIST_KEY_LSApplicationCategoryType = "public.app-category.utilities";
 				INFOPLIST_KEY_NSHumanReadableCopyright = "";
 				INFOPLIST_KEY_NSLocalNetworkUsageDescription = "Punktfunk connects directly to your punktfunk host on the local network to stream video, audio, and input.";
@@ -429,7 +429,7 @@
 				CURRENT_PROJECT_VERSION = 1;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_FILE = Config/Info.plist;
-				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunkempfänger";
+				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunk";
 				INFOPLIST_KEY_LSApplicationCategoryType = "public.app-category.utilities";
 				INFOPLIST_KEY_NSLocalNetworkUsageDescription = "Punktfunk connects directly to your punktfunk host on the local network to stream video, audio, and input.";
 				INFOPLIST_KEY_NSMicrophoneUsageDescription = "Your microphone is streamed to the connected punktfunk host, where it appears as a virtual microphone.";
@@ -468,7 +468,7 @@
 				CURRENT_PROJECT_VERSION = 1;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_FILE = Config/Info.plist;
-				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunkempfänger";
+				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunk";
 				INFOPLIST_KEY_LSApplicationCategoryType = "public.app-category.utilities";
 				INFOPLIST_KEY_NSLocalNetworkUsageDescription = "Punktfunk connects directly to your punktfunk host on the local network to stream video, audio, and input.";
 				INFOPLIST_KEY_NSMicrophoneUsageDescription = "Your microphone is streamed to the connected punktfunk host, where it appears as a virtual microphone.";
@@ -506,7 +506,7 @@
 				CURRENT_PROJECT_VERSION = 1;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_FILE = Config/Info.plist;
-				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunkempfänger";
+				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunk";
 				INFOPLIST_KEY_LSApplicationCategoryType = "public.app-category.utilities";
 				INFOPLIST_KEY_NSLocalNetworkUsageDescription = "Punktfunk connects directly to your punktfunk host on the local network to stream video, audio, and input.";
 				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
@@ -536,7 +536,7 @@
 				CURRENT_PROJECT_VERSION = 1;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_FILE = Config/Info.plist;
-				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunkempfänger";
+				INFOPLIST_KEY_CFBundleDisplayName = "Punktfunk";
 				INFOPLIST_KEY_LSApplicationCategoryType = "public.app-category.utilities";
 				INFOPLIST_KEY_NSLocalNetworkUsageDescription = "Punktfunk connects directly to your punktfunk host on the local network to stream video, audio, and input.";
 				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;

clients/apple/Sources/PunktfunkClient/AcknowledgementsView.swift

@@ -10,12 +10,17 @@ struct AcknowledgementsView: View {
 
     var body: some View {
         ScrollView {
+            // Top-level LazyVStack so the third-party-notices chunks (Licenses.thirdPartyNoticesChunks,
+            // ~885 KB total) load lazily as they scroll into view — a single Text that large overshoots
+            // the text-rendering height limit (blank below the limit + very slow). spacing 0 keeps the
+            // notice chunks visually continuous; the header block carries its own spacing + bottom pad.
+            LazyVStack(alignment: .leading, spacing: 0) {
                 VStack(alignment: .leading, spacing: 18) {
                     Text("punktfunk")
-                    .font(.title2).bold()
+                        .font(.geist(22, .bold, relativeTo: .title2))
                     if let version {
                         Text("Version \(version)")
-                        .font(.caption)
+                            .font(.geist(12, relativeTo: .caption))
                             .foregroundStyle(.secondary)
                     }
                     Text(Licenses.appLicense)
@@ -24,19 +29,41 @@ struct AcknowledgementsView: View {
 
                     Divider()
 
+                    Text("Bundled font")
+                        .font(.geist(17, .semibold, relativeTo: .headline))
+                    Text("punktfunk ships the Geist typeface (Geist Sans), "
+                        + "© The Geist Project Authors / Vercel, used under the SIL Open Font "
+                        + "License 1.1.")
+                        .font(.geist(12, relativeTo: .caption))
+                        .foregroundStyle(.secondary)
+                    if !Licenses.fontLicense.isEmpty {
+                        Text(Licenses.fontLicense)
+                            .font(.caption2.monospaced())
+                            .modifier(SelectableText())
+                    }
+
+                    Divider()
+
                     Text("Third-party software")
-                    .font(.headline)
+                        .font(.geist(17, .semibold, relativeTo: .headline))
                     Text(
                         "punktfunk uses the open-source components below, each under its own license. "
                             + "On some platforms FFmpeg is additionally bundled under the LGPL v2.1+ "
                             + "(dynamically linked, replaceable)."
                     )
-                .font(.caption)
+                    .font(.geist(12, relativeTo: .caption))
                     .foregroundStyle(.secondary)
-                Text(Licenses.thirdPartyNotices)
+                }
+                .frame(maxWidth: .infinity, alignment: .leading)
+                .padding(.bottom, 18)
+
+                ForEach(Licenses.thirdPartyNoticesChunks.indices, id: \.self) { i in
+                    Text(Licenses.thirdPartyNoticesChunks[i])
                         .font(.caption2.monospaced())
+                        .frame(maxWidth: .infinity, alignment: .leading)
                         .modifier(SelectableText())
                 }
+            }
             .frame(maxWidth: 900, alignment: .leading)
             .frame(maxWidth: .infinity, alignment: .leading)
             .padding()

clients/apple/Sources/PunktfunkClient/AddHostSheet.swift

@@ -80,6 +80,11 @@ struct AddHostSheet: View {
             }
             #if !os(tvOS)
         .formStyle(.grouped)
+        #endif
+            #if os(iOS)
+            // The detent below is sized to fit all 3 rows + the action button exactly, so the
+            // Form must NOT scroll/bounce inside it — lock it. (iOS 16+; safe at iOS 17.)
+            .scrollDisabled(true)
             #endif
             #if os(macOS)
             // macOS: UNCHANGED — Cancel + Spacer + Add in an HStack, both wired to the
@@ -120,8 +125,8 @@ struct AddHostSheet: View {
         // Form + the full-width action row, instead of the half-screen .medium it used to rest
         // at. A single fixed detent is enough: the system keeps the content above the keyboard
         // when Address/Port is focused, and on iPadOS this renders as a short bottom sheet (not a
-        // centered formSheet card). If Dynamic Type grows the rows past this height the Form just
+        // centered formSheet card). The Form itself is .scrollDisabled (above) so it can't
-        // scrolls inside the detent — nothing is clipped. (.height(_:) is iOS 16+, safe at iOS 17.)
+        // bounce/scroll inside this fixed detent. (.height(_:) is iOS 16+, safe at iOS 17.)
         .presentationDetents([.height(320)])
         .presentationDragIndicator(.visible)
         #endif

clients/apple/Sources/PunktfunkClient/BrandTheme.swift

@@ -0,0 +1,39 @@
+// App-wide brand chrome. SwiftUI has no single switch to put a custom font on every navigation
+// title, so the iOS large/inline nav titles are themed through UINavigationBar's appearance proxy
+// (set once at launch). Backgrounds are left at the system defaults — transparent at the scroll
+// edge (the large title floats on the content), blurred once scrolled — so only the typeface
+// changes: Geist, matching the cards and the website.
+
+#if os(iOS)
+import PunktfunkKit
+import UIKit
+
+enum BrandTheme {
+    static func apply() {
+        BrandFont.registerIfNeeded()
+
+        let scrollEdge = UINavigationBarAppearance()
+        scrollEdge.configureWithTransparentBackground()
+        applyFonts(to: scrollEdge)
+
+        let standard = UINavigationBarAppearance()
+        standard.configureWithDefaultBackground()
+        applyFonts(to: standard)
+
+        let proxy = UINavigationBar.appearance()
+        proxy.scrollEdgeAppearance = scrollEdge
+        proxy.standardAppearance = standard
+        proxy.compactAppearance = standard
+    }
+
+    /// Override only the title fonts; leave colors/backgrounds at the configured defaults.
+    private static func applyFonts(to appearance: UINavigationBarAppearance) {
+        if let large = UIFont(name: "Geist-Bold", size: 34) {
+            appearance.largeTitleTextAttributes[.font] = large
+        }
+        if let inline = UIFont(name: "Geist-SemiBold", size: 17) {
+            appearance.titleTextAttributes[.font] = inline
+        }
+    }
+}
+#endif

clients/apple/Sources/PunktfunkClient/ContentView.swift

@@ -28,6 +28,7 @@ struct ContentView: View {
     @AppStorage(DefaultsKey.gamepadType) private var gamepadType = 0
     @AppStorage(DefaultsKey.bitrateKbps) private var bitrateKbps = 0
     @AppStorage(DefaultsKey.audioChannels) private var audioChannels = 2
+    @AppStorage(DefaultsKey.hdrEnabled) private var hdrEnabled = true
     @AppStorage(DefaultsKey.fullscreenWhileStreaming) private var fullscreenWhileStreaming = true
     @AppStorage(DefaultsKey.hudEnabled) private var hudEnabled = true
     @AppStorage(DefaultsKey.hudPlacement) private var hudPlacement = HUDPlacement.topTrailing.rawValue
@@ -68,15 +69,19 @@ struct ContentView: View {
                 // A session actually started — remember it on the card ("Connected … ago"
                 // plus the accent ring on the most recent host).
                 guard let host = model.activeHost else { break }
-                store.markConnected(host.id)
                 // Delegated approval just succeeded: the operator let this device in, so pin the
                 // host's observed fingerprint and remember it as paired — future connects are then
                 // silent (rule 1), exactly like after a PIN/TOFU success. Dismisses the wait prompt.
-                if awaitingApproval?.host.id == host.id {
+                let approvedFingerprint = awaitingApproval?.host.id == host.id
-                    if let fp = model.connection?.hostFingerprint {
+                    ? model.connection?.hostFingerprint : nil
-                        store.pin(host.id, fingerprint: fp)
+                if awaitingApproval?.host.id == host.id { awaitingApproval = nil }
-                    }
+                // Persist on the next runloop tick: HostStore is an ObservableObject, and mutating
-                    awaitingApproval = nil
+                // its @Published from inside .onChange (a view-update callback) trips SwiftUI's
+                // "Publishing changes from within view updates". A one-tick delay is imperceptible.
+                let store = store
+                DispatchQueue.main.async {
+                    store.markConnected(host.id)
+                    if let approvedFingerprint { store.pin(host.id, fingerprint: approvedFingerprint) }
                 }
             case .idle:
                 // The delegated-approval connect failed, timed out, or was cancelled — drop the
@@ -333,6 +338,7 @@ struct ContentView: View {
                     rawValue: UInt32(clamping: gamepadType)) ?? .auto),
             bitrateKbps: UInt32(clamping: bitrateKbps),
             audioChannels: UInt8(clamping: audioChannels),
+            hdrEnabled: hdrEnabled,
             launchID: launchID,
             allowTofu: allowTofu,
             requestAccess: requestAccess)
@@ -475,6 +481,7 @@ struct ContentView: View {
             gamepad: pad,
             bitrateKbps: bitrate,
             audioChannels: UInt8(clamping: audioChannels),
+            hdrEnabled: hdrEnabled,
             autoTrust: true)
     }
 }

clients/apple/Sources/PunktfunkClient/ControllerTestView.swift

@@ -54,7 +54,7 @@ struct ControllerTestView: View {
     var body: some View {
         VStack(spacing: 0) {
             HStack {
-                Text("Test Controller").font(.headline)
+                Text("Test Controller").font(.geist(17, .semibold, relativeTo: .headline))
                 Spacer()
                 Button("Done") { dismiss() }.keyboardShortcut(.cancelAction)
             }
@@ -99,8 +99,8 @@ struct ControllerTestView: View {
                 .font(.title2)
                 .foregroundStyle(.secondary)
             VStack(alignment: .leading, spacing: 2) {
-                Text(c.name).font(.headline)
+                Text(c.name).font(.geist(17, .semibold, relativeTo: .headline))
-                Text(c.productCategory).font(.caption).foregroundStyle(.secondary)
+                Text(c.productCategory).font(.geist(12, relativeTo: .caption)).foregroundStyle(.secondary)
             }
             Spacer()
         }
@@ -209,7 +209,7 @@ struct ControllerTestView: View {
     ) -> some View {
         VStack(alignment: .leading, spacing: 4) {
             Text("Touchpad\(tp.button.isPressed ? " — click" : "")")
-                .font(.caption2).foregroundStyle(.secondary)
+                .font(.geist(11, relativeTo: .caption2)).foregroundStyle(.secondary)
             ZStack {
                 RoundedRectangle(cornerRadius: 8).stroke(Color.secondary.opacity(0.3))
                 fingerDot(tp.primary, color: .accentColor)
@@ -230,7 +230,7 @@ struct ControllerTestView: View {
     private func motionReadout(_ m: GCMotion) -> some View {
         let a = Self.totalAccel(m)
         return VStack(alignment: .leading, spacing: 2) {
-            Text("Motion").font(.caption2).foregroundStyle(.secondary)
+            Text("Motion").font(.geist(11, relativeTo: .caption2)).foregroundStyle(.secondary)
             Text(String(format: "gyro  %+.2f %+.2f %+.2f",
                         m.rotationRate.x, m.rotationRate.y, m.rotationRate.z))
                 .font(.caption2.monospaced())
@@ -254,11 +254,11 @@ struct ControllerTestView: View {
                 Toggle("Heavy motor (left)", isOn: $heavyOn)
                 Toggle("Light motor (right)", isOn: $lightOn)
                 Label("Backend: \(tester.rumbleBackend)", systemImage: "waveform")
-                    .font(.caption).foregroundStyle(.secondary)
+                    .font(.geist(12, relativeTo: .caption)).foregroundStyle(.secondary)
                 Text("Toggle a motor to feel it. The host maps a game's low/high-frequency "
                     + "rumble onto these two. A DualSense is driven over raw HID (CoreHaptics "
                     + "can't reach its motors on macOS).")
-                    .font(.caption).foregroundStyle(.secondary)
+                    .font(.geist(12, relativeTo: .caption)).foregroundStyle(.secondary)
             }
             .onChange(of: heavyOn) { _, _ in applyRumble() }
             .onChange(of: lightOn) { _, _ in applyRumble() }
@@ -289,11 +289,11 @@ struct ControllerTestView: View {
                         }
                     }
                     Text("Pick an effect, then pull L2/R2 to feel the resistance.")
-                        .font(.caption).foregroundStyle(.secondary)
+                        .font(.geist(12, relativeTo: .caption)).foregroundStyle(.secondary)
                 }
             } else {
                 Text("Adaptive triggers need a DualSense.")
-                    .font(.caption).foregroundStyle(.secondary)
+                    .font(.geist(12, relativeTo: .caption)).foregroundStyle(.secondary)
             }
         }
     }
@@ -348,7 +348,7 @@ struct ControllerTestView: View {
         _ title: String, @ViewBuilder _ content: () -> Content
     ) -> some View {
         VStack(alignment: .leading, spacing: 10) {
-            Text(title).font(.subheadline.weight(.semibold))
+            Text(title).font(.geist(15, .semibold, relativeTo: .subheadline))
             content()
         }
         .frame(maxWidth: .infinity, alignment: .leading)

clients/apple/Sources/PunktfunkClient/HomeView.swift

@@ -127,14 +127,13 @@ struct HomeView: View {
             AddHostSheet { store.add($0) }
         }
         #if os(iOS)
+        // SettingsView owns its own NavigationSplitView (sidebar + detail) and Done button, so it
+        // is presented directly — wrapping it in a NavigationStack here would nest a split view in
+        // a stack (double title bars). `settingsSheetSizing()` widens the sheet on iPad for the
+        // two-column layout.
         .sheet(isPresented: $showSettings) {
-            NavigationStack {
             SettingsView()
-                    .navigationTitle("Settings")
+                .settingsSheetSizing()
-                    .toolbar {
-                        Button("Done") { showSettings = false }
-                    }
-            }
         }
         #endif
         #endif
@@ -172,7 +171,7 @@ struct HomeView: View {
     private var discoveredSection: some View {
         VStack(alignment: .leading, spacing: 10) {
             Label("On this network", systemImage: "antenna.radiowaves.left.and.right")
-                .font(.headline)
+                .font(.geist(15, .semibold, relativeTo: .headline))
                 .foregroundStyle(.secondary)
                 .padding(.horizontal)
             LazyVGrid(columns: gridColumns, spacing: gridSpacing) {
@@ -249,8 +248,10 @@ struct HomeView: View {
     /// the width so the cards stay edge-aligned with the title and bars — sized touch-first: one
     /// column on iPhone portrait, 3–4 generous cards on iPad.
     private var gridColumns: [GridItem] {
+        // Wider than before: the monogram card is a horizontal module (tile + address line), so
+        // it needs room for a monospaced "IP:port" without truncating.
         #if os(macOS)
-        [GridItem(.adaptive(minimum: 180, maximum: 240), spacing: 16)]
+        [GridItem(.adaptive(minimum: 250, maximum: 320), spacing: 16)]
         #elseif os(tvOS)
         [GridItem(.adaptive(minimum: 320), spacing: 48)]
         #else

clients/apple/Sources/PunktfunkClient/HostCards.swift

@@ -1,26 +1,75 @@
 // The host grid's cards: a saved host (tap to connect, context menu) and an mDNS-discovered
-// host (tap to save + connect). Both share the same platform-tuned sizing.
+// host (tap to save + connect). Both share the "monogram module" look — a squared brand-purple
+// monogram tile + a left-aligned bold Geist name over monospaced technical metadata
+// (address, status), framed by a hairline panel border. Industrial, not soft.
 
 import PunktfunkKit
 import SwiftUI
 
-/// Shared host-card sizing — touch-first on iOS, compact on macOS/tvOS.
+/// Shared host-card sizing — touch-first on iOS, compact on macOS, roomy on tvOS.
 private struct CardMetrics {
-    let iconSize: CGFloat
+    let tile: CGFloat       // monogram tile side
-    let iconBox: CGFloat
+    let monogram: CGFloat   // monogram letter point size
-    let cardPadding: CGFloat
+    let name: CGFloat       // host-name point size
-    let nameFont: Font
+    let meta: CGFloat       // address (mono) point size
+    let status: CGFloat     // status-label (mono) point size
+    let padding: CGFloat
+    let spacing: CGFloat    // tile ↔ text gap
+    let radius: CGFloat
 
     static var current: CardMetrics {
         #if os(iOS)
-        CardMetrics(iconSize: 56, iconBox: 76, cardPadding: 28, nameFont: .title3.weight(.semibold))
+        CardMetrics(tile: 54, monogram: 26, name: 19, meta: 13, status: 11,
+                    padding: 16, spacing: 14, radius: 12)
+        #elseif os(tvOS)
+        CardMetrics(tile: 64, monogram: 32, name: 24, meta: 16, status: 14,
+                    padding: 18, spacing: 18, radius: 14)
         #else
-        CardMetrics(iconSize: 42, iconBox: 56, cardPadding: 18, nameFont: .headline)
+        CardMetrics(tile: 44, monogram: 21, name: 15, meta: 12, status: 10.5,
+                    padding: 13, spacing: 12, radius: 10)
         #endif
     }
 }
 
-/// A saved host. The accent ring marks the most-recently-connected one; the context menu
+/// First letter of a host name, uppercased — the monogram glyph. Falls back to a bullet.
+private func monogram(_ name: String) -> String {
+    guard let first = name.trimmingCharacters(in: .whitespacesAndNewlines).first else { return "•" }
+    return String(first).uppercased()
+}
+
+/// The squared monogram tile. `filled` = a solid brand-purple chip (saved hosts); otherwise a
+/// tinted outline (discovered hosts). Shows a spinner in place of the glyph while connecting.
+private func monogramTile(_ letter: String, m: CardMetrics, connecting: Bool, filled: Bool) -> some View {
+    let shape = RoundedRectangle(cornerRadius: m.radius - 3, style: .continuous)
+    return ZStack {
+        shape.fill(filled
+            ? AnyShapeStyle(LinearGradient(
+                colors: [Color.brand, Color.brand.opacity(0.72)],
+                startPoint: .top, endPoint: .bottom))
+            : AnyShapeStyle(Color.brand.opacity(0.14)))
+        if connecting {
+            ProgressView().tint(filled ? .white : Color.brand)
+        } else {
+            // Fixed size (not Dynamic Type): the glyph is pinned inside a fixed tile, so it must
+            // not scale up and spill out at large accessibility text sizes. minimumScaleFactor +
+            // the clip below are belt-and-suspenders for an unusually wide glyph.
+            Text(letter)
+                .font(.geistFixed(m.monogram, .bold))
+                .minimumScaleFactor(0.5)
+                .lineLimit(1)
+                .foregroundStyle(filled ? Color.white : Color.brand)
+        }
+    }
+    .frame(width: m.tile, height: m.tile)
+    .clipShape(shape)
+    .overlay {
+        if !filled {
+            shape.strokeBorder(Color.brand.opacity(0.45), lineWidth: 1)
+        }
+    }
+}
+
+/// A saved host. A left accent bar marks the most-recently-connected one; the context menu
 /// pairs / speed-tests / forgets / removes. Disabled while a session is busy.
 struct HostCardView: View {
     let host: StoredHost
@@ -41,66 +90,44 @@ struct HostCardView: View {
     var body: some View {
         let m = CardMetrics.current
         return Button(action: onConnect) {
-            VStack(spacing: 10) {
+            HStack(spacing: m.spacing) {
-                ZStack {
+                monogramTile(monogram(host.displayName), m: m, connecting: isConnecting, filled: true)
-                    Image(systemName: "play.display")
+                VStack(alignment: .leading, spacing: 4) {
-                        .font(.system(size: m.iconSize, weight: .light))
-                        .foregroundStyle(.tint)
-                        .opacity(isConnecting ? 0.3 : 1)
-                    if isConnecting {
-                        ProgressView()
-                    }
-                }
-                .frame(height: m.iconBox)
-                VStack(spacing: 2) {
-                    HStack(spacing: 6) {
-                        // Presence dot: green = advertising on the LAN now; grey = not seen.
-                        Circle()
-                            .fill(isOnline ? Color.green : Color.secondary.opacity(0.35))
-                            .frame(width: 7, height: 7)
-                            .accessibilityLabel(isOnline ? "Online" : "Offline")
                     Text(host.displayName)
-                            .font(m.nameFont)
+                        .font(.geist(m.name, .bold, relativeTo: .title3))
+                        .foregroundStyle(.primary)
                         .lineLimit(1)
-                    }
-                    HStack(spacing: 4) {
-                        if host.pinnedSHA256 != nil {
-                            Image(systemName: "lock.fill")
-                                .font(.system(size: 9))
-                                .foregroundStyle(.secondary)
-                        }
                     Text("\(host.address):\(String(host.port))")
-                            .font(.caption)
+                        .font(.geist(m.meta, relativeTo: .caption))
                         .foregroundStyle(.secondary)
                         .lineLimit(1)
+                    statusRow(m)
                 }
-                    if let last = host.lastConnected {
+                Spacer(minLength: 0)
-                        Text("Connected \(last, format: .relative(presentation: .named))")
-                            .font(.caption2)
-                            .foregroundStyle(.tertiary)
-                            .lineLimit(1)
             }
-                }
+            .padding(m.padding)
-            }
+            .frame(maxWidth: .infinity, alignment: .leading)
-            .frame(maxWidth: .infinity)
-            .padding(.vertical, m.cardPadding)
-            .padding(.horizontal, 12)
             #if !os(tvOS)
-            // tvOS: the .card button style owns platter + focus motion — extra chrome
+            // tvOS: the .card button style owns platter + focus motion; extra chrome mutes it.
-            // inside it mutes the grow/tilt. Material + accent ring are for pointer UIs.
+            // Elsewhere: a flat material panel with a hairline border (industrial, not a soft blob),
-            // Deliberately .regularMaterial, not Liquid Glass: HIG keeps glass off content
+            // and a brand accent bar down the leading edge for the most-recent host.
-            // tiles (it flattens hierarchy over an opaque grid) — see GlassStyle.swift.
+            .background(.regularMaterial)
-            .background(.regularMaterial, in: RoundedRectangle(cornerRadius: 14))
+            .overlay(alignment: .leading) {
-            .overlay {
                 if isMostRecent {
-                    RoundedRectangle(cornerRadius: 14)
+                    Rectangle().fill(Color.brand).frame(width: 3)
-                        .strokeBorder(Color.accentColor.opacity(0.35), lineWidth: 1.5)
                 }
             }
+            .clipShape(RoundedRectangle(cornerRadius: m.radius, style: .continuous))
+            .overlay {
+                RoundedRectangle(cornerRadius: m.radius, style: .continuous)
+                    .strokeBorder(.quaternary, lineWidth: 1)
+            }
             #endif
         }
         #if os(tvOS)
         .buttonStyle(.card)
+        #elseif os(iOS)
+        .buttonStyle(HostCardButtonStyle(cornerRadius: m.radius))
         #else
         .buttonStyle(.plain)
         #endif
@@ -119,10 +146,31 @@ struct HostCardView: View {
             Button("Remove", role: .destructive, action: onRemove)
         }
     }
+
+    /// Technical status line: a square presence pip + monospaced ONLINE/OFFLINE, and PAIRED when a
+    /// certificate is pinned (the lock state, spelled out).
+    @ViewBuilder private func statusRow(_ m: CardMetrics) -> some View {
+        HStack(spacing: 6) {
+            RoundedRectangle(cornerRadius: 1.5)
+                .fill(isOnline ? Color.green : Color.secondary.opacity(0.4))
+                .frame(width: 6, height: 6)
+                // The state is spelled out in the adjacent text, so the pip is decorative —
+                // otherwise VoiceOver reads the status twice ("Online, ONLINE …").
+                .accessibilityHidden(true)
+            Text(isOnline ? "ONLINE" : "OFFLINE")
+            if host.pinnedSHA256 != nil {
+                Text("· PAIRED")
+            }
+        }
+        .font(.geist(m.status, .medium, relativeTo: .caption2))
+        .tracking(0.8)
+        .foregroundStyle(.secondary)
+        .lineLimit(1)
+    }
 }
 
-/// A host found on the LAN but not yet saved. A dashed ring distinguishes it from saved cards;
+/// A host found on the LAN but not yet saved. A tinted-outline monogram + dashed panel border
-/// tapping saves it and connects (or pairs, if the host requires it).
+/// distinguish it from saved cards; tapping saves it and connects (or pairs, if required).
 struct DiscoveredCardView: View {
     let discovered: DiscoveredHost
     let isBusy: Bool
@@ -131,47 +179,77 @@ struct DiscoveredCardView: View {
     var body: some View {
         let m = CardMetrics.current
         return Button(action: onConnect) {
-            VStack(spacing: 10) {
+            HStack(spacing: m.spacing) {
-                Image(systemName: "play.display")
+                monogramTile(monogram(discovered.name), m: m, connecting: false, filled: false)
-                    .font(.system(size: m.iconSize, weight: .light))
+                VStack(alignment: .leading, spacing: 4) {
-                    .foregroundStyle(.tint)
-                    .frame(height: m.iconBox)
-                VStack(spacing: 2) {
                     Text(discovered.name)
-                        .font(m.nameFont)
+                        .font(.geist(m.name, .bold, relativeTo: .title3))
+                        .foregroundStyle(.primary)
                         .lineLimit(1)
-                    HStack(spacing: 4) {
-                        Image(systemName: discovered.requiresPairing ? "lock.fill" : "wifi")
-                            .font(.system(size: 9))
-                            .foregroundStyle(.secondary)
                     Text("\(discovered.host):\(String(discovered.port))")
-                            .font(.caption)
+                        .font(.geist(m.meta, relativeTo: .caption))
+                        .foregroundStyle(.secondary)
+                        .lineLimit(1)
+                    HStack(spacing: 6) {
+                        Image(systemName: discovered.requiresPairing
+                            ? "lock.fill" : "antenna.radiowaves.left.and.right")
+                            .font(.system(size: m.status))
+                            .accessibilityHidden(true) // decorative; the adjacent text says the state
+                        Text(discovered.requiresPairing ? "PAIRING REQUIRED" : "DISCOVERED")
+                    }
+                    .font(.geist(m.status, .medium, relativeTo: .caption2))
+                    .tracking(0.8)
                     .foregroundStyle(.secondary)
                     .lineLimit(1)
                 }
-                    Text(discovered.requiresPairing ? "Pairing required" : "Discovered")
+                Spacer(minLength: 0)
-                        .font(.caption2)
-                        .foregroundStyle(.tertiary)
             }
-            }
+            .padding(m.padding)
-            .frame(maxWidth: .infinity)
+            .frame(maxWidth: .infinity, alignment: .leading)
-            .padding(.vertical, m.cardPadding)
-            .padding(.horizontal, 12)
             #if !os(tvOS)
-            .background(.regularMaterial, in: RoundedRectangle(cornerRadius: 14))
+            .background(.regularMaterial)
+            .clipShape(RoundedRectangle(cornerRadius: m.radius, style: .continuous))
             .overlay {
-                RoundedRectangle(cornerRadius: 14)
+                RoundedRectangle(cornerRadius: m.radius, style: .continuous)
                     .strokeBorder(
-                        Color.secondary.opacity(0.25),
+                        Color.secondary.opacity(0.3),
                         style: StrokeStyle(lineWidth: 1, dash: [4, 3]))
             }
             #endif
         }
         #if os(tvOS)
         .buttonStyle(.card)
+        #elseif os(iOS)
+        .buttonStyle(HostCardButtonStyle(cornerRadius: m.radius))
         #else
         .buttonStyle(.plain)
         #endif
         .disabled(isBusy)
     }
 }
+
+#if os(iOS)
+/// The iOS host-card press/hover treatment, one style for both idioms:
+/// - iPhone: a subtle scale-down on press + a light impact haptic on press-down. (`hoverEffect` is
+///   inert without a pointer.)
+/// - iPad: the system pointer "magnet" — the cursor morphs into a highlight that conforms to the
+///   card's rounded rect on hover. (`sensoryFeedback` is inert without a Taptic Engine, and the
+///   press scale doubles as click feedback.)
+struct HostCardButtonStyle: ButtonStyle {
+    var cornerRadius: CGFloat
+
+    func makeBody(configuration: Configuration) -> some View {
+        configuration.label
+            .scaleEffect(configuration.isPressed ? 0.96 : 1)
+            .animation(.spring(response: 0.3, dampingFraction: 0.65), value: configuration.isPressed)
+            // Conform the pointer highlight to the card's rounded rect, not its square bounds.
+            .contentShape(.hoverEffect, RoundedRectangle(cornerRadius: cornerRadius, style: .continuous))
+            .hoverEffect(.highlight)
+            // Light tap on press-down (nil on release so it fires once, on touch). No haptic
+            // hardware on iPad → silently ignored there.
+            .sensoryFeedback(trigger: configuration.isPressed) { _, pressed in
+                pressed ? .impact(weight: .light) : nil
+            }
+    }
+}
+#endif

clients/apple/Sources/PunktfunkClient/LibraryView.swift

@@ -146,7 +146,7 @@ private struct GameCard: View {
                 .clipShape(RoundedRectangle(cornerRadius: 10, style: .continuous))
                 .overlay(alignment: .topLeading) { storeBadge }
             Text(game.title)
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
                 .lineLimit(2)
                 .foregroundStyle(.secondary)
         }
@@ -154,7 +154,7 @@ private struct GameCard: View {
 
     private var storeBadge: some View {
         Text(game.isCustom ? "Custom" : "Steam")
-            .font(.caption2.weight(.semibold))
+            .font(.geist(11, .semibold, relativeTo: .caption2))
             .padding(.horizontal, 6)
             .padding(.vertical, 3)
             .background(.ultraThinMaterial, in: Capsule())
@@ -193,7 +193,7 @@ private struct PosterImage: View {
         ZStack {
             Rectangle().fill(.quaternary)
             Text(title)
-                .font(.headline)
+                .font(.geist(17, .semibold, relativeTo: .headline))
                 .multilineTextAlignment(.center)
                 .foregroundStyle(.secondary)
                 .padding(8)

clients/apple/Sources/PunktfunkClient/PairSheet.swift

@@ -48,7 +48,7 @@ struct PairSheet: View {
                 + "(http://<host>:3000 → Pairing). "
                 + "Pairing verifies both sides at once — no fingerprint comparison "
                 + "needed.")
-                .font(.callout)
+                .font(.geist(16, relativeTo: .callout))
                 .foregroundStyle(.secondary)
                 .multilineTextAlignment(.center)
             TVFieldRow(
@@ -59,7 +59,7 @@ struct PairSheet: View {
             ) { editing = .clientName }
             if let errorText {
                 Text(errorText)
-                    .font(.callout)
+                    .font(.geist(16, relativeTo: .callout))
                     .foregroundStyle(.red)
             }
             HStack(spacing: 32) {
@@ -121,13 +121,13 @@ struct PairSheet: View {
                         + "(http://<host>:3000 → Pairing). "
                         + "Pairing verifies both sides at once — no fingerprint "
                         + "comparison needed.")
-                        .font(.caption)
+                        .font(.geist(12, relativeTo: .caption))
                         .foregroundStyle(.secondary)
                 }
                 if let errorText {
                     Section {
                         Text(errorText)
-                            .font(.callout)
+                            .font(.geist(16, relativeTo: .callout))
                             .foregroundStyle(.red)
                     }
                 }

clients/apple/Sources/PunktfunkClient/PunktfunkClientApp.swift

@@ -12,8 +12,19 @@ struct PunktfunkClientApp: App {
     @NSApplicationDelegateAdaptor(AppDelegate.self) private var appDelegate
     #endif
 
+    init() {
+        #if os(iOS)
+        // Put Geist on the navigation titles before any bar is built.
+        BrandTheme.apply()
+        #endif
+    }
+
     var body: some Scene {
         WindowGroup("Punktfunk") {
+            // Pin the whole app's tint to the brand purple explicitly — the asset-catalog accent
+            // resolution is environment/timing-sensitive and can fall back to system blue. Wraps the
+            // screenshot harness too, so captured screens are on-brand.
+            Group {
                 #if DEBUG
                 // PUNKTFUNK_SHOT_SCENE=<name> → show that single mock-populated screen full-bleed for
                 // the App Store screenshot capture (tools/screenshots.sh). Normal launch otherwise;
@@ -27,6 +38,11 @@ struct PunktfunkClientApp: App {
                 ContentView()
                 #endif
             }
+            .tint(.brand)
+            // Geist Sans is the app's typeface. This sets the default for unstyled text and the
+            // form row labels; views that pick an explicit size/weight use `.geist(…)` directly.
+            .font(.geist(17, relativeTo: .body))
+        }
         // The Stream menu (Disconnect ⌘D, Show/Hide Statistics ⌘⇧S) — a real menu bar on
         // macOS, hardware-keyboard shortcuts on iPad. tvOS has neither.
         #if !os(tvOS)
@@ -34,7 +50,10 @@ struct PunktfunkClientApp: App {
         #endif
         #if os(macOS)
         Settings {
+            // A separate scene — `.tint` does not cross scene boundaries, so re-apply the brand
+            // tint here or the Preferences window falls back to the (unreliable) asset accent.
             SettingsView()
+                .tint(.brand)
         }
         #endif
     }

clients/apple/Sources/PunktfunkClient/Screenshots/ScreenshotScenes.swift

@@ -103,11 +103,11 @@ private struct ShotSettings: View {
                 .shadow(radius: 40, y: 16)
         }
         #elseif os(iOS)
-        NavigationStack {
+        // SettingsView owns its NavigationSplitView (sidebar + detail) and Done button, so it is
-            SettingsView()
+        // rendered directly — a wrapping NavigationStack would nest a split view in a stack. Open
-                .navigationTitle("Settings")
+        // on General so the shot lands on real controls (iPad: sidebar + General detail; iPhone:
-                .navigationBarTitleDisplayMode(.inline)
+        // the General page) instead of the bare category list.
-        }
+        SettingsView(initialCategory: .general)
         #else
         NavigationStack { SettingsView() }
         #endif
@@ -175,10 +175,10 @@ private struct ShotHUD: View {
                 .foregroundStyle(.secondary)
             #if os(macOS)
             Text("⌘⎋ releases the mouse")
-                .font(.caption2).foregroundStyle(.secondary)
+                .font(.geist(11, relativeTo: .caption2)).foregroundStyle(.secondary)
             #elseif os(tvOS)
             Text("Press Menu to disconnect")
-                .font(.caption).foregroundStyle(.secondary)
+                .font(.geist(12, relativeTo: .caption)).foregroundStyle(.secondary)
             #endif
         }
         .padding(10)
@@ -259,7 +259,7 @@ private struct ShotDesktopFrame: View {
             HStack(spacing: 8) {
                 Image(systemName: "gamecontroller.fill")
                 Text("Streaming from Battlestation")
-                    .font(.system(.callout, weight: .semibold))
+                    .font(.geist(16, .semibold, relativeTo: .callout))
             }
             .padding(.horizontal, 14).padding(.vertical, 9)
             .glassBackground(Capsule())

clients/apple/Sources/PunktfunkClient/SessionModel.swift

@@ -129,6 +129,8 @@ final class SessionModel: ObservableObject {
             #endif
         }()
         let hdrCapable = hdrEnabled && displayHDR
+        // 4:4:4 opt-out (default on); the hardware-decode probe below is the real gate.
+        let want444 = (UserDefaults.standard.object(forKey: DefaultsKey.enable444) as? Bool) ?? true
         Task.detached(priority: .userInitiated) {
             // PunktfunkConnection.init blocks on the QUIC handshake — keep it off the main
             // actor. The persistent identity is presented on every connect so a paired
@@ -138,9 +140,21 @@ final class SessionModel: ObservableObject {
             // Advertise 10-bit + HDR10 when enabled: the host upgrades to a BT.2020 PQ Main10 stream
             // only for actual HDR content (its own gate); the VideoToolbox/Metal present path is
             // HDR-capable (P010 + itur_2100_PQ + EDR). 0 keeps the 8-bit BT.709 SDR stream.
-            let videoCaps: UInt8 = hdrCapable
+            var videoCaps: UInt8 = hdrCapable
                 ? (PunktfunkConnection.videoCap10Bit | PunktfunkConnection.videoCapHDR)
                 : 0
+            // Advertise full-chroma 4:4:4 only when allowed AND this device can HARDWARE-decode it
+            // (software 4:4:4 is too slow for real-time). The host content-gates depth, so an
+            // HDR-advertised session can still receive an 8-bit 4:4:4 stream (SDR content) — require
+            // BOTH depths there. Otherwise a no-op (the host emits 4:4:4 only if it too opted in);
+            // `chromaFormat` on the connection reflects what was actually resolved.
+            let canDecode444 =
+                hdrCapable
+                ? (Stage444Probe.hwDecode444_8bit && Stage444Probe.hwDecode444_10bit)
+                : Stage444Probe.hwDecode444_8bit
+            if want444, canDecode444 {
+                videoCaps |= PunktfunkConnection.videoCap444
+            }
             let result = Result { try PunktfunkConnection(
                 host: host.address, port: host.port,
                 width: width, height: height, refreshHz: hz,

clients/apple/Sources/PunktfunkClient/SettingsView.swift

@@ -1,10 +1,12 @@
 // App settings. The host creates a native virtual output at exactly the chosen size/refresh —
 // there is no scaling anywhere in the pipeline.
 //
-// Navigation differs per platform: macOS uses a tabbed preferences window (the sections had
+// Navigation differs per platform, but all three group the same categories (General, Display,
-// outgrown one scrolling pane); iOS uses a single grouped Form; tvOS uses a focus-native
+// Audio, Controllers, Advanced, About): macOS uses a tabbed preferences window; iOS/iPadOS uses
-// pushed-picker layout. The individual sections (`streamModeSection`, `audioSection`, …) are
+// an adaptive NavigationSplitView — a category sidebar + detail pane on iPad, auto-collapsing to
-// shared across all three so a setting is defined exactly once.
+// a hierarchical push list on iPhone (the system Settings idiom on each); tvOS uses a
+// focus-native pushed-picker layout. The individual sections (`streamModeSection`,
+// `audioSection`, …) are shared across all three so a setting is defined exactly once.
 
 #if os(macOS)
 import AppKit
@@ -21,7 +23,9 @@ struct SettingsView: View {
     @AppStorage(DefaultsKey.compositor) private var compositor = 0
     @AppStorage(DefaultsKey.gamepadType) private var gamepadType = 0
     @AppStorage(DefaultsKey.bitrateKbps) private var bitrateKbps = 0
-    @AppStorage(DefaultsKey.presenter) private var presenter = "stage1"
+    @AppStorage(DefaultsKey.presenter) private var presenter = "stage2"
+    @AppStorage(DefaultsKey.hdrEnabled) private var hdrEnabled = true
+    @AppStorage(DefaultsKey.enable444) private var enable444 = true
     @AppStorage(DefaultsKey.libraryEnabled) private var libraryEnabled = false
     @AppStorage(DefaultsKey.fullscreenWhileStreaming) private var fullscreenWhileStreaming = true
     @AppStorage(DefaultsKey.micEnabled) private var micEnabled = true
@@ -32,6 +36,22 @@ struct SettingsView: View {
     #if DEBUG && !os(tvOS)
     @State private var showControllerTest = false
     #endif
+    #if os(iOS)
+    @AppStorage(DefaultsKey.pointerCapture) private var pointerCapture = true
+    // The sidebar selection drives the detail pane on iPad and the pushed sub-page on iPhone.
+    // Width class decides the initial value: nil on iPhone (show the category list first),
+    // General on iPad (a two-column layout should never open with an empty detail).
+    @Environment(\.horizontalSizeClass) private var horizontalSizeClass
+    @State private var settingsSelection: SettingsCategory?
+    // Tracked so the detail can show its own Done whenever the sidebar (and its Done) is off screen
+    // — not just on iPhone, but on any iPad layout that collapses the sidebar to an overlay. Starts
+    // .doubleColumn so iPad reliably opens with the sidebar (and its Done) visible.
+    @State private var columnVisibility: NavigationSplitViewVisibility = .doubleColumn
+    // Sticky once the wheel lands on "Custom…", so editing a width/height that briefly equals a
+    // preset doesn't snap the wheel back off Custom. A stored non-preset value reads as custom even
+    // when this is false (see `isCustomResolution`), so it survives relaunches without persisting.
+    @State private var customMode = false
+    #endif
     #if os(macOS)
     @AppStorage(DefaultsKey.speakerUID) private var speakerUID = ""
     @AppStorage(DefaultsKey.micUID) private var micUID = ""
@@ -39,6 +59,15 @@ struct SettingsView: View {
     @State private var inputDevices: [AudioDevice] = []
     #endif
 
+    #if os(iOS)
+    /// `initialCategory` is nil in the app (the list opens un-selected on iPhone; iPad lands on
+    /// General via `onAppear`). The screenshot harness passes an explicit category so the captured
+    /// shot opens on a real settings page (a populated detail) rather than the bare category list.
+    init(initialCategory: SettingsCategory? = nil) {
+        _settingsSelection = State(initialValue: initialCategory)
+    }
+    #endif
+
     var body: some View {
         #if os(tvOS)
         // Native tv pattern: no inline text entry (typing numbers with a remote is
@@ -66,6 +95,7 @@ struct SettingsView: View {
 
             Form {
                 presenterSection
+                hdrSection
                 windowSection
                 statisticsSection
             }
@@ -106,29 +136,116 @@ struct SettingsView: View {
     }
     #endif
 
-    // MARK: - iOS: one grouped Form
+    // MARK: - iOS / iPadOS: adaptive split view
 
     #if os(iOS)
     private var iosBody: some View {
-        Form {
+        NavigationSplitView(columnVisibility: $columnVisibility) {
-            streamModeSection
+            List(selection: $settingsSelection) {
-            audioSection
+                ForEach(SettingsCategory.allCases) { category in
-            compositorSection
+                    // On iPhone the split view collapses to a push list, but a selection List
-            presenterSection
+                    // draws no disclosure indicator of its own — add one in compact width for the
-            statisticsSection
+                    // expected drill-in affordance. On iPad the selected row highlights instead, so
-            experimentalSection
+                    // the chevron is omitted there.
-            controllersSection
+                    HStack {
-            Section {
+                        Label(category.title, systemImage: category.symbol)
-                NavigationLink("Acknowledgements") { AcknowledgementsView() }
+                        if horizontalSizeClass == .compact {
+                            Spacer()
+                            Image(systemName: "chevron.forward")
+                                .font(.footnote.weight(.semibold))
+                                .foregroundStyle(.tertiary)
+                                // Purely a drill-in affordance — the row's button trait already
+                                // conveys "opens"; keep it out of the VoiceOver announcement.
+                                .accessibilityHidden(true)
+                        }
+                    }
+                    .tag(category)
+                }
+            }
+            .navigationTitle("Settings")
+            .toolbar {
+                ToolbarItem(placement: .confirmationAction) {
+                    Button("Done") { dismiss() }
+                }
+            }
+        } detail: {
+            // NavigationSplitView hosts the detail in its own navigation context (its title bar),
+            // so no inner NavigationStack — that would double the bar on iPad. On iPhone the split
+            // view collapses to one stack and pushes this when a row is tapped. `?? .general` only
+            // backs the brief pre-selection window; the list never auto-pushes on a nil selection.
+            settingsDetail(settingsSelection ?? .general)
+                // Keep a Done on the detail whenever the sidebar (and its Done) isn't on screen: the
+                // iPhone push, or any iPad layout that collapsed the sidebar to an overlay. When the
+                // sidebar is showing, its Done is the only one — so this stays hidden to avoid two.
+                .toolbar {
+                    if horizontalSizeClass == .compact || columnVisibility == .detailOnly {
+                        ToolbarItem(placement: .confirmationAction) {
+                            Button("Done") { dismiss() }
+                        }
+                    }
                 }
         }
-        .formStyle(.grouped)
         .onAppear {
+            if horizontalSizeClass == .regular, settingsSelection == nil {
+                settingsSelection = .general
+            }
             gamepads.refresh()
             gamepads.startDiscovery()
         }
+        // A regular→regular launch sets the default above; this catches a compact→regular change
+        // (e.g. an iPad leaving narrow split-screen multitasking) so the detail pane fills in.
+        .onChange(of: horizontalSizeClass) { _, newValue in
+            if newValue == .regular, settingsSelection == nil {
+                settingsSelection = .general
+            }
+        }
         .onDisappear { gamepads.stopDiscovery() }
     }
+
+    @ViewBuilder
+    private func settingsDetail(_ category: SettingsCategory) -> some View {
+        switch category {
+        case .general:
+            Form {
+                streamModeSection
+                pointerSection
+                compositorSection
+            }
+            .formStyle(.grouped)
+            .navigationTitle("General")
+            .navigationBarTitleDisplayMode(.inline)
+        case .display:
+            Form {
+                presenterSection
+                hdrSection
+                statisticsSection
+            }
+            .formStyle(.grouped)
+            .navigationTitle("Display")
+            .navigationBarTitleDisplayMode(.inline)
+        case .audio:
+            Form { audioSection }
+                .formStyle(.grouped)
+                .navigationTitle("Audio")
+                .navigationBarTitleDisplayMode(.inline)
+        case .controllers:
+            Form { controllersSection }
+                .formStyle(.grouped)
+                .navigationTitle("Controllers")
+                .navigationBarTitleDisplayMode(.inline)
+        case .advanced:
+            Form { experimentalSection }
+                .formStyle(.grouped)
+                .navigationTitle("Advanced")
+                .navigationBarTitleDisplayMode(.inline)
+        case .about:
+            // Already a full scrollable view that sets its own "Acknowledgements" title; pin the
+            // display mode inline to match the five sibling detail pages (it would otherwise inherit
+            // the large title from the "Settings" sidebar root).
+            AcknowledgementsView()
+                .navigationBarTitleDisplayMode(.inline)
+        }
+    }
     #endif
 
     // MARK: - tvOS
@@ -156,6 +273,10 @@ struct SettingsView: View {
         Binding(get: { hudEnabled ? "on" : "off" }, set: { hudEnabled = $0 == "on" })
     }
 
+    private var hdrEnabledTag: Binding<String> {
+        Binding(get: { hdrEnabled ? "on" : "off" }, set: { hdrEnabled = $0 == "on" })
+    }
+
     private var tvBody: some View {
         let currentTag = "\(width)x\(height)x\(hz)"
         let bounds = UIScreen.main.nativeBounds
@@ -186,20 +307,25 @@ struct SettingsView: View {
                     selection: $audioChannels)
                 if bitrateKbps > 1_000_000 {
                     Label(Self.gigabitWarning, systemImage: "exclamationmark.triangle.fill")
-                        .font(.caption)
+                        .font(.geist(12, relativeTo: .caption))
                         .foregroundStyle(.orange)
                         .multilineTextAlignment(.center)
                 }
                 TVSelectionRow(
                     title: "Compositor", options: compositors, selection: $compositor)
+                #if DEBUG
                 TVSelectionRow(
-                    title: "Presenter",
+                    title: "Presenter (debug)",
-                    options: [("Stage 1 (default)", "stage1"), ("Stage 2 (experimental)", "stage2")],
+                    options: [("Stage 2 (default)", "stage2"), ("Stage 1 (debug)", "stage1")],
                     selection: $presenter)
+                #endif
+                TVSelectionRow(
+                    title: "10-bit HDR",
+                    options: [("On", "on"), ("Off", "off")], selection: hdrEnabledTag)
                 Text("The host creates a virtual output at exactly this mode — native "
                     + "resolution, no scaling. \(Self.bitrateFooter) A specific compositor "
                     + "is honored only if available on the host.")
-                    .font(.caption)
+                    .font(.geist(12, relativeTo: .caption))
                     .foregroundStyle(.secondary)
                     .multilineTextAlignment(.center)
                     .padding(.top, 8)
@@ -219,7 +345,7 @@ struct SettingsView: View {
                 TVSelectionRow(
                     title: "Controller type", options: Self.padTypes, selection: $gamepadType)
                 Text(Self.controllersFooter)
-                    .font(.caption)
+                    .font(.geist(12, relativeTo: .caption))
                     .foregroundStyle(.secondary)
                     .multilineTextAlignment(.center)
                     .padding(.top, 8)
@@ -243,6 +369,63 @@ struct SettingsView: View {
 
     @ViewBuilder private var streamModeSection: some View {
         Section {
+            #if os(iOS)
+            // Touch-first: a rotating wheel of common resolutions (this device's own mode first) and
+            // a segmented refresh-rate control — the same family as the Clock/Timer pickers. The host
+            // renders a virtual output at exactly the chosen mode, so these are real pixel sizes. The
+            // last wheel row, "Custom…", reveals width/height/refresh fields for an arbitrary mode.
+            VStack(alignment: .leading, spacing: 4) {
+                Text("Resolution")
+                    .font(.geist(15, relativeTo: .subheadline))
+                    .foregroundStyle(.secondary)
+                Picker("Resolution", selection: resolutionSelection) {
+                    ForEach(resolutionChoices, id: \.tag) { choice in
+                        Text(choice.label).tag(choice.tag)
+                    }
+                }
+                .labelsHidden()
+                .pickerStyle(.wheel)
+                .frame(maxHeight: 140)
+            }
+            if isCustomResolution {
+                // Arbitrary entry: type the exact width × height (and refresh) the host should drive.
+                HStack {
+                    TextField("Width", value: $width, format: .number.grouping(.never))
+                        .keyboardType(.numberPad)
+                    Text("×")
+                    TextField("Height", value: $height, format: .number.grouping(.never))
+                        .labelsHidden()
+                        .keyboardType(.numberPad)
+                }
+                // A row built from an HStack of TextFields otherwise insets its bottom separator to
+                // the inner content, clipping the hairline under "Width"; pin it to the cell edge.
+                .alignmentGuide(.listRowSeparatorLeading) { _ in 0 }
+                LabeledContent("Refresh rate") {
+                    TextField("Hz", value: $hz, format: .number.grouping(.never))
+                        .keyboardType(.numberPad)
+                        .multilineTextAlignment(.trailing)
+                }
+            } else if refreshChoices.count > 1 {
+                VStack(alignment: .leading, spacing: 6) {
+                    Text("Refresh rate")
+                        .font(.geist(15, relativeTo: .subheadline))
+                        .foregroundStyle(.secondary)
+                    Picker("Refresh rate", selection: $hz) {
+                        ForEach(refreshChoices, id: \.self) { rate in
+                            Text("\(rate) Hz").tag(rate)
+                        }
+                    }
+                    .labelsHidden()
+                    .pickerStyle(.segmented)
+                }
+            } else {
+                // A device with a single supported rate (e.g. 60 Hz) has nothing to pick.
+                LabeledContent("Refresh rate") {
+                    Text("\(hz) Hz").foregroundStyle(.secondary)
+                }
+            }
+            Button("Use this display's mode") { fillFromMainScreen() }
+            #elseif os(macOS)
             HStack {
                 TextField("Resolution", value: $width, format: .number.grouping(.never))
                 Text("×")
@@ -253,6 +436,7 @@ struct SettingsView: View {
             LabeledContent("") {
                 Button("Use this display's mode") { fillFromMainScreen() }
             }
+            #endif
             #if !os(tvOS)
             Toggle("Automatic bitrate", isOn: automaticBitrate)
             if bitrateKbps != 0 {
@@ -267,7 +451,7 @@ struct SettingsView: View {
                 }
                 if bitrateKbps > 1_000_000 {
                     Label(Self.gigabitWarning, systemImage: "exclamationmark.triangle.fill")
-                        .font(.caption)
+                        .font(.geist(12, relativeTo: .caption))
                         .foregroundStyle(.orange)
                 }
             }
@@ -277,11 +461,85 @@ struct SettingsView: View {
         } footer: {
             Text("The host creates a virtual output at exactly this mode — "
                 + "native resolution, no scaling. \(Self.bitrateFooter)")
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
                 .foregroundStyle(.secondary)
         }
     }
 
+    #if os(iOS)
+    // MARK: - Stream mode (iOS wheel)
+
+    /// Sentinel wheel tag for the "Custom…" row. Real tags are "WxH" (digits + "x"), so this can't
+    /// collide with a resolution.
+    private static let customResolutionTag = "custom"
+
+    /// 16:9 then ultrawide presets; the device's native mode is prepended at runtime.
+    private static let resolutionPresets: [(name: String, w: Int, h: Int)] = [
+        ("720p", 1280, 720),
+        ("1080p", 1920, 1080),
+        ("1440p", 2560, 1440),
+        ("4K", 3840, 2160),
+        ("Ultrawide 1080p", 2560, 1080),
+        ("Ultrawide 1440p", 3440, 1440),
+        ("Super ultrawide", 5120, 1440),
+    ]
+
+    /// The non-custom wheel rows: this device's native mode first, then the presets, deduped by
+    /// dimensions (native wins a tie).
+    private var resolutionModes: [(name: String, w: Int, h: Int)] {
+        let bounds = UIScreen.main.nativeBounds // portrait-oriented pixels
+        let native = (w: Int(max(bounds.width, bounds.height)), h: Int(min(bounds.width, bounds.height)))
+        let all = [(name: "This device", w: native.w, h: native.h)] + Self.resolutionPresets
+        var seen = Set<String>()
+        return all.filter { seen.insert("\($0.w)x\($0.h)").inserted }
+    }
+
+    /// Wheel rows: the resolution modes, then a "Custom…" row that reveals the numeric fields.
+    private var resolutionChoices: [(label: String, tag: String)] {
+        resolutionModes.map { (label: "\($0.name)  ·  \($0.w) × \($0.h)", tag: "\($0.w)x\($0.h)") }
+            + [(label: "Custom…", tag: Self.customResolutionTag)]
+    }
+
+    private var presetResolutionTags: Set<String> {
+        Set(resolutionModes.map { "\($0.w)x\($0.h)" })
+    }
+
+    /// True when the editable custom fields should show: the wheel is parked on "Custom…" (sticky),
+    /// or the stored size simply isn't one of the presets (e.g. a value synced from a Mac) — so a
+    /// non-preset mode stays editable across relaunches without a persisted flag.
+    private var isCustomResolution: Bool {
+        customMode || !presetResolutionTags.contains("\(width)x\(height)")
+    }
+
+    /// The wheel works in "WxH" tags so one selection drives both width and height; the custom
+    /// sentinel toggles `customMode` instead of writing a size.
+    private var resolutionSelection: Binding<String> {
+        Binding(
+            get: { isCustomResolution ? Self.customResolutionTag : "\(width)x\(height)" },
+            set: { tag in
+                if tag == Self.customResolutionTag {
+                    customMode = true
+                    return
+                }
+                customMode = false
+                let parts = tag.split(separator: "x").compactMap { Int($0) }
+                guard parts.count == 2 else { return }
+                width = parts[0]
+                height = parts[1]
+            })
+    }
+
+    /// Refresh rates the device can actually display (no point asking the host to render frames the
+    /// screen can't show), plus any stored custom value so it stays selectable.
+    private var refreshChoices: [Int] {
+        let maxHz = UIScreen.main.maximumFramesPerSecond
+        var rates = [60, 120, 240].filter { $0 <= maxHz }
+        if rates.isEmpty { rates = [maxHz] }
+        if !rates.contains(hz) { rates.append(hz) }
+        return rates.sorted()
+    }
+    #endif
+
     @ViewBuilder private var audioSection: some View {
         Section {
             Picker("Audio channels", selection: $audioChannels) {
@@ -321,11 +579,35 @@ struct SettingsView: View {
             Text("Host audio plays through the speaker; the microphone feeds the "
                 + "host's virtual mic. System default follows macOS device changes. "
                 + "Applies from the next session.")
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
                 .foregroundStyle(.secondary)
         }
     }
 
+    #if os(iOS)
+    /// iPad-only pointer-capture toggle: lock the mouse/trackpad for relative movement (games) vs
+    /// forward an absolute cursor position (desktop). Empty on iPhone (no hardware-pointer lock —
+    /// the mouse path there is always the absolute fallback).
+    @ViewBuilder private var pointerSection: some View {
+        if UIDevice.current.userInterfaceIdiom == .pad {
+            Section {
+                Toggle("Capture pointer for games", isOn: $pointerCapture)
+            } header: {
+                Text("Pointer")
+            } footer: {
+                Text("With a mouse or trackpad connected, lock the pointer and send relative "
+                    + "movement — the expected behavior for games (mouse-look). Turn this off for "
+                    + "desktop use to keep the pointer free and send its absolute position instead. "
+                    + "The lock needs the stream full-screen and frontmost; it falls back to the "
+                    + "absolute pointer automatically (Stage Manager, Slide Over). Finger touch is "
+                    + "unaffected. Applies from the next session.")
+                    .font(.geist(12, relativeTo: .caption))
+                    .foregroundStyle(.secondary)
+            }
+        }
+    }
+    #endif
+
     @ViewBuilder private var compositorSection: some View {
         Section {
             Picker("Compositor", selection: $compositor) {
@@ -341,7 +623,7 @@ struct SettingsView: View {
             Text("Which compositor drives the virtual output on the host. A specific "
                 + "choice is honored only if that backend is available there — "
                 + "otherwise the host falls back to auto-detection.")
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
                 .foregroundStyle(.secondary)
         }
     }
@@ -355,26 +637,50 @@ struct SettingsView: View {
         } footer: {
             Text("Take the window fullscreen when a session starts and restore it on the host "
                 + "list, so only the stream is fullscreen — not the picker.")
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
                 .foregroundStyle(.secondary)
         }
         #endif
     }
 
+    // Stage-2 (Metal/VTDecompressionSession) is the default and only user-visible presenter — it
+    // recovers from a wedged decoder, where stage-1's AVSampleBufferDisplayLayer freezes hard on a
+    // lost HEVC reference. Stage-1 is kept reachable as a DEBUG-only override for diagnostics, like
+    // the controller test. Empty in release builds (no presenter UI; stage-2 always).
     @ViewBuilder private var presenterSection: some View {
+        #if DEBUG
         Section {
             Picker("Presenter", selection: $presenter) {
-                Text("Stage 1 (default)").tag("stage1")
+                Text("Stage 2 (default)").tag("stage2")
-                Text("Stage 2 (experimental)").tag("stage2")
+                Text("Stage 1 (debug)").tag("stage1")
             }
         } header: {
-            Text("Video presenter")
+            Text("Video presenter · debug")
         } footer: {
-            Text("Stage 1 feeds compressed video to the system display layer (known-good). "
+            Text("Stage 2 (default) decodes explicitly and presents through Metal with a display "
-                + "Stage 2 decodes explicitly and presents through Metal with a display "
+                + "link — it adds a capture→present (glass-to-glass) latency line in the HUD and "
-                + "link — it adds a capture→present (glass-to-glass) latency line in the HUD "
+                + "self-recovers from decode stalls. Stage 1 feeds compressed video straight to the "
-                + "and shortens the present tail. Applies from the next session.")
+                + "system display layer; it freezes on a lost HEVC reference frame, so it's a debug "
-                .font(.caption)
+                + "fallback only. Applies from the next session.")
+                .font(.geist(12, relativeTo: .caption))
+                .foregroundStyle(.secondary)
+        }
+        #endif
+    }
+
+    @ViewBuilder private var hdrSection: some View {
+        Section {
+            Toggle("10-bit HDR", isOn: $hdrEnabled)
+            Toggle("Full chroma (4:4:4)", isOn: $enable444)
+        } header: {
+            Text("Video quality")
+        } footer: {
+            Text("HDR requests a 10-bit BT.2020 PQ (HDR10) stream — it only engages when the host is "
+                + "sending HDR content AND this display supports HDR. 4:4:4 requests full chroma "
+                + "(sharper text/UI, more bandwidth) — it only engages when this device can "
+                + "hardware-decode it AND the host opted in. Otherwise the stream stays 8-bit "
+                + "4:2:0 SDR. Applies from the next session.")
+                .font(.geist(12, relativeTo: .caption))
                 .foregroundStyle(.secondary)
         }
     }
@@ -392,7 +698,7 @@ struct SettingsView: View {
             Text("Statistics")
         } footer: {
             Text(Self.statisticsFooter)
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
                 .foregroundStyle(.secondary)
         }
     }
@@ -407,7 +713,7 @@ struct SettingsView: View {
                 + "(Steam + custom) via the host's management API; tap a title to launch it. "
                 + "The host must expose that API on the LAN with a token "
                 + "(serve --mgmt-bind 0.0.0.0 --mgmt-token …).")
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
                 .foregroundStyle(.secondary)
         }
     }
@@ -441,7 +747,7 @@ struct SettingsView: View {
             Text("Controllers")
         } footer: {
             Text(Self.controllersFooter)
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
                 .foregroundStyle(.secondary)
         }
     }
@@ -593,13 +899,13 @@ struct SettingsView: View {
                         }
                     }
                 }
-                .font(.caption2)
+                .font(.geist(11, relativeTo: .caption2))
                 .foregroundStyle(.secondary)
             }
             Spacer()
             if gamepads.active?.id == controller.id {
                 Text("In use")
-                    .font(.caption2.weight(.semibold))
+                    .font(.geist(11, .semibold, relativeTo: .caption2))
                     .padding(.horizontal, 8)
                     .padding(.vertical, 3)
                     .background(Capsule().fill(.green.opacity(0.2)))
@@ -621,6 +927,10 @@ struct SettingsView: View {
         width = Int(max(bounds.width, bounds.height))
         height = Int(min(bounds.width, bounds.height))
         hz = UIScreen.main.maximumFramesPerSecond
+        #if os(iOS)
+        // The native mode is the "This device" wheel row, so leave Custom mode if it was on.
+        customMode = false
+        #endif
         #endif
     }
 }
@@ -631,3 +941,52 @@ extension Double {
         Swift.min(Swift.max(self, lo), hi)
     }
 }
+
+#if os(iOS)
+/// The settings groups, mirroring the macOS preference tabs. On iPad each is a sidebar row that
+/// drives the detail pane; on iPhone the same list collapses to pushed sub-pages. Internal (not
+/// private) so the screenshot harness can open SettingsView on a specific category.
+enum SettingsCategory: String, CaseIterable, Identifiable {
+    case general, display, audio, controllers, advanced, about
+
+    var id: Self { self }
+
+    var title: String {
+        switch self {
+        case .general: return "General"
+        case .display: return "Display"
+        case .audio: return "Audio"
+        case .controllers: return "Controllers"
+        case .advanced: return "Advanced"
+        case .about: return "About"
+        }
+    }
+
+    var symbol: String {
+        switch self {
+        case .general: return "gearshape"
+        case .display: return "display"
+        case .audio: return "speaker.wave.2"
+        case .controllers: return "gamecontroller"
+        case .advanced: return "slider.horizontal.3"
+        case .about: return "info.circle"
+        }
+    }
+}
+
+extension View {
+    /// Present the settings sheet large on iPad so the NavigationSplitView has room for its
+    /// sidebar + detail — a default form sheet is too narrow and the split view would collapse to
+    /// the iPhone push list. No-op on iPhone (the standard sheet is already right) and on iOS 17
+    /// (no `presentationSizing` — it falls back to the default sheet, which still degrades cleanly
+    /// to the push list).
+    @ViewBuilder
+    func settingsSheetSizing() -> some View {
+        if UIDevice.current.userInterfaceIdiom == .pad, #available(iOS 18, *) {
+            presentationSizing(.page)
+        } else {
+            self
+        }
+    }
+}
+#endif

clients/apple/Sources/PunktfunkClient/SpeedTestSheet.swift

@@ -52,7 +52,7 @@ struct SpeedTestSheet: View {
     var body: some View {
         VStack(spacing: 20) {
             Label("Speed test — \(host.displayName)", systemImage: "gauge.with.needle")
-                .font(.headline)
+                .font(.geist(17, .semibold, relativeTo: .headline))
                 .foregroundStyle(.tint)
 
             switch phase {
@@ -73,7 +73,7 @@ struct SpeedTestSheet: View {
                 resultView(result)
             case .failed(let message):
                 Text(message)
-                    .font(.callout)
+                    .font(.geist(16, relativeTo: .callout))
                     .foregroundStyle(.red)
                     .multilineTextAlignment(.center)
             }
@@ -149,13 +149,13 @@ struct SpeedTestSheet: View {
             if let rec = Self.recommendedKbps(result) {
                 Text("Recommended bitrate: \(Self.mbpsLabel(kbps: rec)) "
                     + "(~70% of measured, headroom for encoder bursts).")
-                    .font(.caption)
+                    .font(.geist(12, relativeTo: .caption))
                     .foregroundStyle(.secondary)
                     .multilineTextAlignment(.center)
             } else {
                 Text("Too little data made it through to recommend a bitrate — "
                     + "check the network and retry.")
-                    .font(.caption)
+                    .font(.geist(12, relativeTo: .caption))
                     .foregroundStyle(.secondary)
                     .multilineTextAlignment(.center)
             }

clients/apple/Sources/PunktfunkClient/StreamHUDView.swift

@@ -69,19 +69,19 @@ struct StreamHUDView: View {
             Text(model.mouseCaptured
                 ? "⌘⎋ releases the mouse"
                 : "Click the stream to capture input")
-                .font(.caption2)
+                .font(.geist(11, relativeTo: .caption2))
                 .foregroundStyle(.secondary)
             // The client-side cursor (⌘⇧C) draws the local cursor over the stream instead of
             // capturing it — the only accurate cursor for gamescope, whose capture has none.
             Text("⌘⇧C toggles the on-screen cursor")
-                .font(.caption2)
+                .font(.geist(11, relativeTo: .caption2))
                 .foregroundStyle(.secondary)
             #elseif os(iOS)
             // Touch always plays directly; ⌘⎋ (hardware keyboard) toggles kb/mouse.
             Text(model.mouseCaptured
                 ? "⌘⎋ releases keyboard & mouse"
                 : "⌘⎋ captures keyboard & mouse")
-                .font(.caption2)
+                .font(.geist(11, relativeTo: .caption2))
                 .foregroundStyle(.secondary)
             #endif
             #if os(tvOS)
@@ -89,13 +89,13 @@ struct StreamHUDView: View {
             // A press (the focus engine consumes it before the host sees it). Disconnect is
             // the Siri Remote's Menu button (.onExitCommand on the stream) — just hint it.
             Text("Press Menu to disconnect")
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
                 .foregroundStyle(.secondary)
             #else
             // ⌘D lives on the app's Stream menu (so it still works when the HUD is hidden);
             // this button is the in-overlay, click-to-disconnect affordance.
             Button("Disconnect (⌘D)") { model.disconnect() }
-                .font(.caption)
+                .font(.geist(12, relativeTo: .caption))
             #endif
         }
         .padding(10)

clients/apple/Sources/PunktfunkClient/TrustCardView.swift

@@ -3,6 +3,7 @@
 // or drops this and runs the PIN pairing ceremony instead.
 
 import Foundation
+import PunktfunkKit
 import SwiftUI
 
 struct TrustCardView: View {
@@ -18,11 +19,11 @@ struct TrustCardView: View {
                 .font(.system(size: 36, weight: .light))
                 .foregroundStyle(.tint)
             Text("Verify \(hostName)")
-                .font(.title3.weight(.semibold))
+                .font(.geist(20, .semibold, relativeTo: .title3))
             Text("First connection. Compare this fingerprint with the one "
                 + "punktfunk-host logged at startup (\u{201C}clients pin this "
                 + "fingerprint\u{201D}):")
-                .font(.callout)
+                .font(.geist(16, relativeTo: .callout))
                 .foregroundStyle(.secondary)
                 .multilineTextAlignment(.center)
             Text(Self.format(fingerprint: fingerprint))
@@ -58,7 +59,7 @@ struct TrustCardView: View {
                 #else
                 .buttonStyle(.borderless)
                 #endif
-                .font(.callout)
+                .font(.geist(16, relativeTo: .callout))
         }
         .padding(28)
         .frame(maxWidth: 440)

clients/apple/Sources/PunktfunkKit/BrandFont.swift

@@ -0,0 +1,101 @@
+// Geist — the punktfunk brand typeface (the same family the website ships). Bundled as static
+// OTF weights in this kit's resources and registered with Core Text at first use, so it works
+// identically in the Xcode app and the `swift run` dev shell (Bundle.module resolves to the
+// package resource bundle in both). Geist Sans carries titles/UI; Geist Mono carries the technical
+// readouts — host addresses, status labels, the stream-stats HUD — for the industrial look.
+//
+// Licensed under the SIL Open Font License 1.1 (Resources/Fonts/Geist-OFL.txt).
+
+import CoreText
+import SwiftUI
+#if canImport(UIKit)
+import UIKit
+#elseif canImport(AppKit)
+import AppKit
+#endif
+
+public enum BrandFont {
+    public enum Weight {
+        case regular, medium, semibold, bold
+    }
+
+    /// PostScript names of the bundled faces (verified from each OTF's name table). Geist Sans only
+    /// — Geist Mono is intentionally not shipped; the app's typeface is Geist Sans throughout.
+    private static let sansFaces = ["Geist-Regular", "Geist-Medium", "Geist-SemiBold", "Geist-Bold"]
+
+    /// Registered exactly once per process — a static `let` initializer is run lazily and is
+    /// guaranteed thread-safe + run-at-most-once by the runtime.
+    private static let registered: Void = {
+        for face in sansFaces {
+            guard let url = Bundle.module.url(
+                forResource: face, withExtension: "otf", subdirectory: "Fonts") else {
+                #if DEBUG
+                print("BrandFont: bundled face \(face).otf not found — text will fall back to system")
+                #endif
+                continue
+            }
+            var error: Unmanaged<CFError>?
+            if !CTFontManagerRegisterFontsForURL(url as CFURL, .process, &error) {
+                #if DEBUG
+                let message = error?.takeRetainedValue().localizedDescription ?? "unknown error"
+                print("BrandFont: failed to register \(face): \(message)")
+                #endif
+            }
+        }
+    }()
+
+    /// Force registration before the first `Font.custom` lookup. Cheap to call repeatedly.
+    public static func registerIfNeeded() { _ = registered }
+
+    fileprivate static func sansFace(_ weight: Weight) -> String {
+        switch weight {
+        case .regular: return "Geist-Regular"
+        case .medium: return "Geist-Medium"
+        case .semibold: return "Geist-SemiBold"
+        case .bold: return "Geist-Bold"
+        }
+    }
+}
+
+public extension Color {
+    /// The punktfunk brand purple (the app-icon lens / website `--brand`). Defined explicitly,
+    /// independent of the asset-catalog accent — `Color.accentColor` resolution is environment- and
+    /// timing-sensitive (it can fall back to system blue), and the brand mark must never drift.
+    /// Light: #6656F2, Dark: #8678F5 (the lighter violet reads better on dark surfaces).
+    static let brand: Color = {
+        #if canImport(UIKit)
+        return Color(UIColor { traits in
+            traits.userInterfaceStyle == .dark
+                ? UIColor(red: 0x86 / 255, green: 0x78 / 255, blue: 0xF5 / 255, alpha: 1)
+                : UIColor(red: 0x66 / 255, green: 0x56 / 255, blue: 0xF2 / 255, alpha: 1)
+        })
+        #elseif canImport(AppKit)
+        return Color(NSColor(name: nil) { appearance in
+            appearance.bestMatch(from: [.aqua, .darkAqua]) == .darkAqua
+                ? NSColor(red: 0x86 / 255, green: 0x78 / 255, blue: 0xF5 / 255, alpha: 1)
+                : NSColor(red: 0x66 / 255, green: 0x56 / 255, blue: 0xF2 / 255, alpha: 1)
+        })
+        #else
+        // Non-Apple fallback: the light brand value, so all branches agree on a canonical color.
+        return Color(red: 0x66 / 255, green: 0x56 / 255, blue: 0xF2 / 255)
+        #endif
+    }()
+}
+
+public extension Font {
+    /// Geist Sans at an explicit point size, scaling with Dynamic Type relative to `textStyle`.
+    static func geist(
+        _ size: CGFloat, _ weight: BrandFont.Weight = .regular,
+        relativeTo textStyle: TextStyle = .body
+    ) -> Font {
+        BrandFont.registerIfNeeded()
+        return .custom(BrandFont.sansFace(weight), size: size, relativeTo: textStyle)
+    }
+
+    /// Geist Sans at a FIXED point size that does not scale with Dynamic Type — for glyphs pinned
+    /// inside a fixed-size container (e.g. the monogram tile), where a scaled letter would overflow.
+    static func geistFixed(_ size: CGFloat, _ weight: BrandFont.Weight = .regular) -> Font {
+        BrandFont.registerIfNeeded()
+        return .custom(BrandFont.sansFace(weight), fixedSize: size)
+    }
+}

clients/apple/Sources/PunktfunkKit/DefaultsKeys.swift

@@ -22,9 +22,22 @@ public enum DefaultsKey {
     public static let speakerUID = "punktfunk.speakerUID"
     public static let micUID = "punktfunk.micUID"
     public static let presenter = "punktfunk.presenter"
+    /// Request a 10-bit BT.2020 PQ (HDR10) stream. On by default; only takes effect when the host
+    /// has HDR content AND this display supports HDR — otherwise the stream stays 8-bit SDR.
+    public static let hdrEnabled = "punktfunk.hdrEnabled"
+    /// Request a full-chroma 4:4:4 stream when this device can HARDWARE-decode it (`Stage444Probe`).
+    /// On by default; only takes effect when the host also opted in to 4:4:4 (otherwise the stream
+    /// stays 4:2:0). Sharper text/UI at the cost of more bandwidth.
+    public static let enable444 = "punktfunk.enable444"
     public static let hosts = "punktfunk.hosts"
     /// Client-side cursor mode: "auto" (shown only in gamescope sessions), "always", "never".
     public static let cursorMode = "punktfunk.cursorMode"
+    /// iPad: capture the mouse/trackpad pointer (pointer lock → relative movement) for games,
+    /// rather than forwarding an absolute cursor position. On by default. Only meaningful on iPad
+    /// with a hardware mouse/trackpad; the system grants the lock only to a full-screen, frontmost
+    /// scene and silently falls back to the absolute pointer when it can't (Stage Manager / Slide
+    /// Over). Read by `StreamViewController.prefersPointerLocked`.
+    public static let pointerCapture = "punktfunk.pointerCapture"
     /// Experimental: show the host's game library (browsed over the management API). Off by default.
     public static let libraryEnabled = "punktfunk.libraryEnabled"
     /// macOS: take the window fullscreen while streaming and restore it on the host list. On by default.

clients/apple/Sources/PunktfunkKit/GamepadFeedback.swift

@@ -68,6 +68,14 @@ private final class RumbleRenderer: @unchecked Sendable {
     private var broken = false
     /// Last logged active/silent state — for a one-line transition log, not per-event spam.
     private var wasActive = false
+    // Backoff after an engine failure. A broken `gamecontrollerd.haptics` XPC connection (CoreHaptics
+    // -4811 "server connection broke") fails EVERY rebuild until the service relaunches — and that
+    // break fires neither stoppedHandler nor resetHandler, so without a cooldown the next rumble
+    // update immediately rebuilds into the same dead connection, flooding the log and never
+    // recovering. Delay the next setup() — growing 0.5→1→2→4 s on repeated failure — and clear it
+    // the moment a player runs cleanly (or the controller changes).
+    private var retryAfter = Date.distantPast
+    private var consecutiveFailures = 0
 
     /// CHHapticEvent sharpness = actuator frequency. A DualSense's voice-coil motors need a
     /// defined frequency to move at all — an intensity-only event (no sharpness) left them
@@ -91,6 +99,8 @@ private final class RumbleRenderer: @unchecked Sendable {
             self.closeHID()
             self.controller = c
             self.broken = false
+            self.consecutiveFailures = 0
+            self.retryAfter = .distantPast
             _ = self.openHIDIfDualSense(c)
             onBackend?(self.backendNote(for: c))
         }
@@ -108,7 +118,7 @@ private final class RumbleRenderer: @unchecked Sendable {
             // other pad (and for a DualSense whose HID device could not be opened).
             if self.hidRumble(low: lowAmp, high: highAmp) { return }
             guard !self.broken else { return }
-            if active, self.low == nil, self.high == nil {
+            if active, self.low == nil, self.high == nil, Date() >= self.retryAfter {
                 self.setup()
             }
             let ok: Bool
@@ -124,8 +134,15 @@ private final class RumbleRenderer: @unchecked Sendable {
             }
             // Rebuild on the next nonzero amplitude if an engine errored — and tear down OUTSIDE
             // the `inout` accesses above, so teardown() never mutates a motor that a `drive` call
-            // still holds an exclusive reference to.
+            // still holds an exclusive reference to. Back off so a broken XPC isn't re-hit every
-            if !ok { self.teardown() }
+            // update; once a player is actually running the path has recovered, so clear the backoff.
+            if !ok {
+                self.teardown()
+                self.scheduleRetryBackoff()
+            } else if self.low?.player != nil || self.high?.player != nil {
+                self.consecutiveFailures = 0
+                self.retryAfter = .distantPast
+            }
         }
     }
 
@@ -157,14 +174,29 @@ private final class RumbleRenderer: @unchecked Sendable {
             low = makeMotor(haptics, .default)
         }
         if low == nil, high == nil {
-            // Haptics present but no engine could be built right now (server busy / a transient
+            // Haptics present but no engine could be built right now (server busy / XPC broken). Do
-            // error). Do NOT latch broken — the next nonzero amplitude retries setup().
+            // NOT latch broken — back off and the next nonzero amplitude past the cooldown retries.
-            log.warning("rumble: haptics present but engine setup failed — will retry on next rumble")
+            log.warning("rumble: haptics present but engine setup failed — backing off, will retry")
+            scheduleRetryBackoff()
         }
     }
 
+    /// Push the next engine-build attempt out after a failure (capped exponential backoff), so a
+    /// broken `gamecontrollerd.haptics` connection gets time to relaunch instead of being re-hit on
+    /// every rumble update.
+    private func scheduleRetryBackoff() {
+        consecutiveFailures += 1
+        let shift = min(consecutiveFailures - 1, 4)
+        retryAfter = Date().addingTimeInterval(min(0.5 * Double(1 << shift), 4))
+    }
+
     private func makeMotor(_ haptics: GCDeviceHaptics, _ locality: GCHapticsLocality) -> Motor? {
         guard let engine = haptics.createEngine(withLocality: locality) else { return nil }
+        // A controller's motors carry no audio, so keep this engine OUT of the app's audio session
+        // (the default is to join it). Streaming keeps an AVAudioSession active the whole time;
+        // letting a haptics-only engine join it is a needless coupling that can get its
+        // gamecontrollerd XPC connection interrupted (the repeated -4811 server-connection breaks).
+        engine.playsHapticsOnly = true
         // The haptic server can stop or reset the engine out from under us — app backgrounding, an
         // audio-session interruption (a call, Siri, another audio app), or a server crash. Left
         // unhandled the players go dead and every later rumble throws, latching rumble off for the
@@ -338,29 +370,32 @@ public final class GamepadFeedback {
         // Hidout traffic (lightbar / player LEDs / triggers) only exists on a PlayStation-pad
         // session — a DualSense or a DualShock 4 (lightbar only). Block briefly on it there and
         // let rumble own the wait elsewhere; on an Xbox session it stays nonblocking.
-        let hasHidout = connection.resolvedGamepad == .dualSense
-            || connection.resolvedGamepad == .dualShock4
-        let hidTimeout: UInt32 = hasHidout ? 10 : 0
         let thread = Thread { [connection, flag, drainDone, weak self] in
             while !flag.isStopped {
                 do {
-                    if let r = try connection.nextRumble(timeoutMs: 10), r.pad == 0 {
+                    // Poll the feedback planes NON-BLOCKING. A blocking poll (timeoutMs > 0) holds
+                    // the connection's shared feedback lock for its whole wait; the video pump drains
+                    // HDR mastering metadata (nextHdrMeta) on the SAME lock every frame, so a blocking
+                    // poll here starved it and throttled HDR to ~1 fps (SDR, which never drains HDR
+                    // meta, was unaffected). Pacing with a short sleep OUTSIDE the lock (below) keeps
+                    // rumble/HID latency low while leaving the lock free between polls.
+                    if let r = try connection.nextRumble(timeoutMs: 0), r.pad == 0 {
                         self?.rumble.apply(low: r.low, high: r.high)
                     }
-                    // Drain a BOUNDED burst of hidout events: only the first poll waits,
+                    // Drain a BOUNDED burst of hidout events so sustained 0xCD traffic (a game writing
-                    // and the cap + stop check keep sustained 0xCD traffic (a game writing
+                    // per-frame LED/trigger reports) can't spin here or block stop() past one cycle.
-                    // per-frame LED/trigger reports) from starving the rumble poll above
-                    // or blocking stop() past one cycle.
                     var burst = 0
                     while burst < 64, !flag.isStopped,
-                          let ev = try connection.nextHidOutput(
+                          let ev = try connection.nextHidOutput(timeoutMs: 0) {
-                            timeoutMs: burst == 0 ? hidTimeout : 0) {
                         self?.render(ev)
                         burst += 1
                     }
                 } catch {
                     break // .closed (or fatal) — the session is over
                 }
+                // ~8 ms poll cadence (≈125 Hz), slept OUTSIDE the feedback lock — low rumble/HID
+                // latency without holding the lock the HDR-meta drain needs.
+                if !flag.isStopped { Thread.sleep(forTimeInterval: 0.008) }
             }
             drainDone.signal()
         }

clients/apple/Sources/PunktfunkKit/InputCapture.swift

@@ -160,7 +160,13 @@ public final class InputCapture {
             previous.onPreempted?()
         }
         Self.activeCapture = self
-        if let mouse = GCMouse.current { attach(mouse: mouse) }
+        // Attach EVERY connected mouse, not just GCMouse.current. With two pointing devices (e.g.
+        // the iPad's own Magic Keyboard trackpad AND a Universal Control "V-UC Automouse"), only one
+        // is `current` at a time; attaching just that one left the OTHER device's motion handler
+        // uninstalled, so moving it did nothing. Each GCMouse delivers its own deltas through its own
+        // handler, so handling all of them lets either device drive. New arrivals are caught by the
+        // GCMouseDidConnect observer below.
+        for mouse in GCMouse.mice() { attach(mouse: mouse) }
         if let keyboard = GCKeyboard.coalesced { attach(keyboard: keyboard) }
         observers.append(NotificationCenter.default.addObserver(
             forName: .GCMouseDidConnect, object: nil, queue: .main

clients/apple/Sources/PunktfunkKit/Licenses.swift

@@ -27,10 +27,35 @@ public enum Licenses {
             + apache
     }
 
+    /// The bundled brand typeface (Geist Sans + Geist Mono) — SIL Open Font License 1.1. The
+    /// license file ships alongside the OTFs in `Resources/Fonts/`, satisfying the OFL's
+    /// distribution requirement; this surfaces it in the Acknowledgements screen too.
+    public static var fontLicense: String {
+        guard let url = Bundle.module.url(
+            forResource: "Geist-OFL", withExtension: "txt", subdirectory: "Fonts"),
+            let text = try? String(contentsOf: url, encoding: .utf8)
+        else { return "" }
+        return text
+    }
+
     /// Third-party software notices for the linked Rust crates (generated by
     /// `scripts/gen-third-party-notices.sh`).
     public static var thirdPartyNotices: String {
         let text = resource("THIRD-PARTY-NOTICES")
         return text.isEmpty ? "Third-party notices unavailable." : text
     }
+
+    /// `thirdPartyNotices` pre-split into render-sized line chunks. The full notices are ~885 KB /
+    /// 16k lines; a single SwiftUI `Text` that large overshoots CoreText/CoreAnimation's max
+    /// renderable height — it lays out for ages and draws blank past the limit — so the
+    /// Acknowledgements screen renders these chunks in a `LazyVStack` (only on-screen chunks lay
+    /// out, and no chunk is tall enough to clip). Split at line boundaries and joined with "\n";
+    /// the inter-chunk break is the `LazyVStack` row boundary, so no text is lost. Computed once.
+    public static let thirdPartyNoticesChunks: [String] = {
+        let lines = thirdPartyNotices.split(separator: "\n", omittingEmptySubsequences: false)
+        let chunkSize = 200
+        return stride(from: 0, to: lines.count, by: chunkSize).map { start in
+            lines[start..<min(start + chunkSize, lines.count)].joined(separator: "\n")
+        }
+    }()
 }

clients/apple/Sources/PunktfunkKit/MetalVideoPresenter.swift

@@ -1,21 +1,35 @@
-// Stage-2 presenter, present half: draw a decoded NV12 CVPixelBuffer into a CAMetalLayer
+// Stage-2 presenter, present half: draw a decoded NV12 / P010 / 4:4:4 CVPixelBuffer into a CAMetalLayer
-// drawable with a BT.709 YUV→RGB shader. The display link (owned by the hosting view) drives
+// drawable with a Y′CbCr→RGB shader. The hosting view's CADisplayLink drives `render` once per vsync
-// `render` once per vsync with the target present time, so a present can finally be stamped and
+// (via Stage2Pipeline.renderTick) with the target present time, so a present can be stamped and the
-// the present tail hand-paced. See docs apple-stage2-presenter.md.
+// present tail hand-paced. See docs apple-stage2-presenter.md.
 //
-// Main-thread only: created during view setup, `render` called from the view's CADisplayLink
+// Main-thread only: created during view setup, `render`/`configure` called from the view's CADisplayLink
-// (which fires on the main runloop). The Metal objects + texture cache are touched only here.
+// (which fires on the main runloop). The Metal objects + texture cache are touched only here. The one
+// exception is `setHdrMeta`, called from the pump thread — it hops the layer write to main so every
+// CALayer mutation stays on one thread.
 
 #if canImport(Metal) && canImport(QuartzCore)
 import CoreGraphics
 import CoreVideo
 import Metal
 import QuartzCore
+import os
 
-/// Runtime-compiled (no metallib build step needed in SwiftPM): a fullscreen triangle and a
+private let presenterLog = Logger(subsystem: "io.unom.punktfunk", category: "presenter")
-/// BT.709 limited-range NV12→RGB fragment shader. uv.y is flipped (1 - p.y) so the top-left-
+
-/// origin texture presents upright (NDC y is up), not upside down. (Colorspace is BT.709 SDR
+/// HDR reference white (BT.2408 "HDR Reference White"): the absolute luminance, in nits, that the
-/// for now — matches the host; 10-bit/HDR + other matrices are a later tie-in.)
+/// PQ signal's diffuse white sits at. Passed to `CAEDRMetadata.hdr10(opticalOutputScale:)`, it anchors
+/// 203-nit diffuse white at EDR 1.0 (the display's SDR-white level) and lets the system tone-map the
+/// brighter highlights into the panel's headroom. This is the missing anchor that made the old HDR path
+/// render "way too bright" (no `edrMetadata` → no reference-white anchoring); a LARGER value renders
+/// dimmer. Matches the host's standard PQ reference white.
+private let hdrReferenceWhiteNits: Float = 203.0
+
+/// Runtime-compiled (no metallib build step needed in SwiftPM): a fullscreen triangle and BT.709 SDR
+/// and BT.2020-PQ HDR Y′CbCr→RGB fragment shaders. uv.y is flipped (1 - p.y) so the top-left-origin
+/// texture presents upright (NDC y is up). The HDR shader outputs PQ-encoded R′G′B′ as-is — the
+/// CAMetalLayer's `itur_2100_PQ` colour space + `edrMetadata` tell the system compositor the samples
+/// are PQ and how to tone-map them (no EOTF here, matching the host's BT.2020 PQ emission).
 private let shaderSource = """
 #include <metal_stdlib>
 using namespace metal;
@@ -30,11 +44,46 @@ vertex VOut pf_vtx(uint vid [[vertex_id]]) {
     return o;
 }
 
+// Bicubic (Catmull-Rom) sampling of the single-channel luma plane. When the drawable is larger
+// than the decoded frame (a window/view bigger than the host's fixed mode), a bilinear upscale
+// looks soft; Catmull-Rom keeps edges crisp — matching AVSampleBufferDisplayLayer's (stage-1)
+// scaler — and reduces to the exact texel at 1:1, so a native-resolution present stays pixel-exact.
+// Nine bilinear taps (TheRealMJP's optimisation of the 16-tap kernel); `s` MUST be a linear
+// sampler. Luma carries the perceived detail, so only it gets bicubic; chroma stays bilinear.
+float catmullRomLuma(texture2d<float> tex, sampler s, float2 uv) {
+    float2 texSize = float2(tex.get_width(), tex.get_height());
+    float2 samplePos = uv * texSize;
+    float2 tc1 = floor(samplePos - 0.5) + 0.5;
+    float2 f = samplePos - tc1;
+    float2 w0 = f * (-0.5 + f * (1.0 - 0.5 * f));
+    float2 w1 = 1.0 + f * f * (-2.5 + 1.5 * f);
+    float2 w2 = f * (0.5 + f * (2.0 - 1.5 * f));
+    float2 w3 = f * f * (-0.5 + 0.5 * f);
+    float2 w12 = w1 + w2;
+    float2 off12 = w2 / w12;
+    float2 tc0 = (tc1 - 1.0) / texSize;
+    float2 tc3 = (tc1 + 2.0) / texSize;
+    float2 tc12 = (tc1 + off12) / texSize;
+    float r = 0.0;
+    r += tex.sample(s, float2(tc0.x,  tc0.y)).r  * (w0.x  * w0.y);
+    r += tex.sample(s, float2(tc12.x, tc0.y)).r  * (w12.x * w0.y);
+    r += tex.sample(s, float2(tc3.x,  tc0.y)).r  * (w3.x  * w0.y);
+    r += tex.sample(s, float2(tc0.x,  tc12.y)).r * (w0.x  * w12.y);
+    r += tex.sample(s, float2(tc12.x, tc12.y)).r * (w12.x * w12.y);
+    r += tex.sample(s, float2(tc3.x,  tc12.y)).r * (w3.x  * w12.y);
+    r += tex.sample(s, float2(tc0.x,  tc3.y)).r  * (w0.x  * w3.y);
+    r += tex.sample(s, float2(tc12.x, tc3.y)).r  * (w12.x * w3.y);
+    r += tex.sample(s, float2(tc3.x,  tc3.y)).r  * (w3.x  * w3.y);
+    return r;
+}
+
+// SDR: 8-bit NV12 / 4:4:4 (BT.709, limited/video range) → full-range RGB. Chroma is sampled at the
+// same UV as luma, so a full-size 4:4:4 chroma plane needs no shader change vs 4:2:0.
 fragment float4 pf_frag(VOut in [[stage_in]],
                         texture2d<float> lumaTex [[texture(0)]],
                         texture2d<float> chromaTex [[texture(1)]]) {
     constexpr sampler s(filter::linear, address::clamp_to_edge);
-    float y = lumaTex.sample(s, in.uv).r;
+    float y = catmullRomLuma(lumaTex, s, in.uv);
     float2 c = chromaTex.sample(s, in.uv).rg;
     // BT.709, 8-bit limited (video) range → full-range RGB.
     y = (y - 16.0/255.0) * (255.0/219.0);
@@ -46,18 +95,18 @@ fragment float4 pf_frag(VOut in [[stage_in]],
     return float4(saturate(float3(r, g, b)), 1.0);
 }
 
-// HDR: 10-bit P010 (BT.2020, limited range), Y'CbCr that is PQ-encoded. We apply the BT.2020
+// HDR: 10-bit P010 / 4:4:4 (BT.2020, limited range), Y′CbCr that is PQ-encoded. We apply the BT.2020
-// matrix to get PQ-encoded R'G'B' and output it as-is — the CAMetalLayer's itur_2100_PQ colour
+// matrix to get PQ-encoded R′G′B′ and output it as-is — the CAMetalLayer's itur_2100_PQ colour space
-// space + EDR tells the compositor the samples are PQ, so it does the PQ→display mapping. No EOTF
+// + edrMetadata tell the compositor the samples are PQ, so it does the PQ→display tone-map. No EOTF
-// here (matching the host, which emitted BT.2020 PQ). P010 stores the 10-bit code in the high bits
+// here. P010/x444 store the 10-bit code in the high bits of each 16-bit sample, so an .r16Unorm sample
-// of each 16-bit sample, so an .r16Unorm sample reads ~code/1023 (the /1024 vs /1023 error is < 0.1%).
+// reads ~code/1023 (the /1024 vs /1023 error is < 0.1%).
 fragment float4 pf_frag_hdr(VOut in [[stage_in]],
                             texture2d<float> lumaTex [[texture(0)]],
                             texture2d<float> chromaTex [[texture(1)]]) {
     constexpr sampler s(filter::linear, address::clamp_to_edge);
-    float y = lumaTex.sample(s, in.uv).r;
+    float y = catmullRomLuma(lumaTex, s, in.uv);
     float2 c = chromaTex.sample(s, in.uv).rg;
-    // BT.2020 10-bit limited (video) range → full-range PQ R'G'B'.
+    // BT.2020 10-bit limited (video) range → full-range PQ R′G′B′.
     y = (y - 64.0/1023.0) * (1023.0/876.0);
     float u = (c.x - 512.0/1023.0) * (1023.0/896.0);
     float v = (c.y - 512.0/1023.0) * (1023.0/896.0);
@@ -74,21 +123,34 @@ public final class MetalVideoPresenter {
 
     private let device: MTLDevice
     private let queue: MTLCommandQueue
-    /// SDR (BT.709 8-bit NV12 → bgra8) and HDR (BT.2020 PQ 10-bit P010 → rgba16Float) pipelines.
+    /// SDR (BT.709 8-bit → bgra8) and HDR (BT.2020 PQ 10-bit → rgba16Float) pipelines. Selected per
-    /// Selected per frame by `render`; the layer is reconfigured when the mode flips (HDR toggle).
+    /// frame in `render`; the layer is reconfigured to match when the session flips (HDR toggle).
     private let pipelineSDR: MTLRenderPipelineState
     private let pipelineHDR: MTLRenderPipelineState
     private var textureCache: CVMetalTextureCache?
-    /// Current layer configuration — switched lazily in `configure(hdr:)` when a frame's mode differs.
-    private var hdrActive = false
 
-    /// nil if Metal is unavailable (no GPU / a headless CI) — the caller falls back to stage-1.
+    /// Current layer configuration — switched in `configure(hdr:)` when a frame's HDR-ness differs.
-    public init?() {
+    /// Main-thread only (read + written from `render`/`configure`, all on the display-link runloop).
+    private var hdrActive = false
+    /// Last HDR mastering grade received via `setHdrMeta` (the host's 0xCE). Cached so a mid-session
+    /// SDR→HDR flip's `configureColor` re-applies the real grade instead of clobbering it back to the
+    /// bare reference-white anchor (an out-of-order race otherwise: `setHdrMeta` and the flip both write
+    /// `edrMetadata`). Main-thread only.
+    private var lastHdrMeta: PunktfunkConnection.HdrMeta?
+
+    #if DEBUG
+    /// Last logged "decoded→drawable" signature, so the diagnostic logs only on a size/HDR change.
+    private var lastSizeSig = ""
+    #endif
+
+    /// nil if Metal is unavailable (no GPU / a headless CI) or a shader fails to compile — the caller
+    /// falls back to stage-1.
+    public static func make() -> MetalVideoPresenter? {
         guard let device = MTLCreateSystemDefaultDevice(),
               let queue = device.makeCommandQueue()
         else { return nil }
-        self.device = device
+        let pipelineSDR: MTLRenderPipelineState
-        self.queue = queue
+        let pipelineHDR: MTLRenderPipelineState
         do {
             let library = try device.makeLibrary(source: shaderSource, options: nil)
             let vtx = library.makeFunction(name: "pf_vtx")
@@ -105,76 +167,148 @@ public final class MetalVideoPresenter {
         } catch {
             return nil
         }
-        CVMetalTextureCacheCreate(kCFAllocatorDefault, nil, device, nil, &textureCache)
+        var cache: CVMetalTextureCache?
-        guard textureCache != nil else { return nil }
+        CVMetalTextureCacheCreate(kCFAllocatorDefault, nil, device, nil, &cache)
+        guard let textureCache = cache else { return nil }
 
         let layer = CAMetalLayer()
         layer.device = device
         layer.pixelFormat = .bgra8Unorm
         layer.framebufferOnly = true
         layer.isOpaque = true
-        // Triple-buffer: more in-flight drawables before `nextDrawable()` (called on the
-        // display-link / MAIN thread) has to block waiting for one to free.
-        layer.maximumDrawableCount = 3
         #if os(macOS)
-        // The display link already paces exactly one present per vsync. Leaving the layer's
+        // The display link already paces exactly one present per vsync. Leaving the layer's own vsync
-        // own vsync wait on means `commandBuffer.present` ALSO blocks for the hardware vsync,
+        // wait on means `commandBuffer.present` ALSO blocks for the hardware vsync, so `nextDrawable()`
-        // so `nextDrawable()` stalls the MAIN thread until a drawable frees — windowed, the
+        // stalls the MAIN thread until a drawable frees — windowed, the WindowServer's looser
-        // WindowServer's looser compositing hides it; FULLSCREEN's tighter, more-direct path
+        // compositing hides it; FULLSCREEN's tighter path serializes the main thread to the display and
-        // serializes the main thread to the display and the stall surfaces as bad judder.
+        // the stall surfaces as bad judder. Disabling the layer-level sync lets present return promptly
-        // Disabling the layer-level sync lets present return promptly (the display link is the
+        // (the display link is the pacing source) — the fix for the fullscreen stutter. macOS-only.
-        // pacing source), which is what fixes the fullscreen stutter. macOS-only property.
         layer.displaySyncEnabled = false
         #endif
+        // Render the drawable at the DECODED frame's resolution (set per-frame in `render`) and let the
+        // system compositor scale it to the layer's bounds — the same `.resizeAspect` path stage-1's
+        // AVSampleBufferDisplayLayer uses. A native-resolution present is then pixel-exact (1:1, no
+        // shader scaling); a resized window rescales via the system's scaler.
+        layer.contentsGravity = .resizeAspect
+        // Triple-buffer: more in-flight drawables before `nextDrawable()` (called on the display-link /
+        // MAIN thread) has to block waiting for one to free.
+        layer.maximumDrawableCount = 3
+
+        return MetalVideoPresenter(
+            device: device, queue: queue, pipelineSDR: pipelineSDR, pipelineHDR: pipelineHDR,
+            textureCache: textureCache, layer: layer)
+    }
+
+    private init(
+        device: MTLDevice, queue: MTLCommandQueue, pipelineSDR: MTLRenderPipelineState,
+        pipelineHDR: MTLRenderPipelineState, textureCache: CVMetalTextureCache, layer: CAMetalLayer
+    ) {
+        self.device = device
+        self.queue = queue
+        self.pipelineSDR = pipelineSDR
+        self.pipelineHDR = pipelineHDR
+        self.textureCache = textureCache
         self.layer = layer
     }
 
-    /// Track the stream mode (the host can Reconfigure mid-stream). Size is in pixels.
+    /// Configure the layer + active pipeline for an SDR or HDR session. MAIN THREAD ONLY. Called once at
-    public func setDrawableSize(_ size: CGSize) {
+    /// session start and again per-frame from `render` (idempotent — the guard makes a same-state call a
-        guard size.width > 0, size.height > 0 else { return }
+    /// no-op), so a mid-session HDR toggle (the host re-inits its encoder; the decoded `frame.isHDR`
-        if layer.drawableSize != size { layer.drawableSize = size }
+    /// flips) reconfigures here automatically. HDR uses an rgba16Float drawable + BT.2020 PQ colour space
-    }
+    /// + EDR with a 203-nit reference-white anchor; SDR uses the plain 8-bit sRGB path.
-
+    public func configure(hdr: Bool) {
-    /// Reconfigure the layer for SDR or HDR when the stream mode flips (HDR toggle). HDR uses an
-    /// rgba16Float drawable + a BT.2020 PQ colour space + EDR, so the compositor PQ-maps to the
-    /// display; SDR uses the plain 8-bit sRGB path. Main-thread only (called from `render`).
-    private func configure(hdr: Bool) {
         guard hdr != hdrActive else { return }
         hdrActive = hdr
+        configureColor(hdr: hdr)
+    }
+
+    /// Set the layer's pixel format + colour config for SDR or HDR. MAIN THREAD ONLY. EDR is requested
+    /// on macOS + iOS (the old `#if os(macOS)` guard left iOS EDR half-engaged). tvOS has NO EDR API
+    /// (`wantsExtendedDynamicRangeContent`/`edrMetadata`/`CAEDRMetadata` are all unavailable there), so
+    /// it gets the PQ pixel format + colour space only — the tvOS compositor tone-maps from those.
+    private func configureColor(hdr: Bool) {
         if hdr {
             layer.pixelFormat = .rgba16Float
             layer.colorspace = CGColorSpace(name: CGColorSpace.itur_2100_PQ)
-            #if os(macOS)
+            #if !os(tvOS)
             layer.wantsExtendedDynamicRangeContent = true
+            // Anchor reference white. Re-apply the real grade if one already arrived (0xCE before the
+            // flip); otherwise the bare 203-nit anchor. Without this anchor the PQ signal is too bright.
+            layer.edrMetadata = makeEDR(lastHdrMeta)
             #endif
         } else {
+            // SDR: gamma-encoded BT.709 [0,1] in an 8-bit drawable; a nil colorspace tags it device/sRGB
+            // (the proven SDR path — never showed the "too bright" issue, which was HDR-only).
             layer.pixelFormat = .bgra8Unorm
             layer.colorspace = nil
-            #if os(macOS)
+            #if !os(tvOS)
             layer.wantsExtendedDynamicRangeContent = false
+            layer.edrMetadata = nil
             #endif
         }
     }
 
-    /// Draw one decoded frame to the next drawable and present it. `isHDR` selects the 10-bit
+    #if !os(tvOS)
-    /// BT.2020 PQ path (P010 input) vs the 8-bit BT.709 path (NV12 input). Returns true on success;
+    private func makeEDR(_ meta: PunktfunkConnection.HdrMeta?) -> CAEDRMetadata {
-    /// false when there's no drawable yet, a texture couldn't be made, or Metal errored — the
+        CAEDRMetadata.hdr10(
-    /// caller then doesn't stamp a present for this frame.
+            displayInfo: meta?.masteringDisplayColorVolume(),
+            contentInfo: meta?.contentLightLevelInfo(),
+            opticalOutputScale: hdrReferenceWhiteNits)
+    }
+    #endif
+
+    /// Update the HDR mastering metadata (drained from the host's 0xCE datagram) to refine the system
+    /// tone-map from the real grade. Called from the PUMP thread, so the layer write is hopped to MAIN
+    /// (every CALayer mutation stays on one thread). The grade is cached so a later SDR→HDR
+    /// `configureColor` re-applies it; the `edrMetadata` write is gated on `hdrActive` (setting it on an
+    /// SDR layer is harmless but pointless, and the flip will apply it anyway).
+    public func setHdrMeta(_ meta: PunktfunkConnection.HdrMeta) {
+        DispatchQueue.main.async { [weak self] in
+            guard let self else { return }
+            self.lastHdrMeta = meta
+            // tvOS has no edrMetadata — the cached grade is still kept above (harmless), it just can't
+            // be applied to the layer there. macOS/iOS refine the system tone-map from the real grade.
+            #if !os(tvOS)
+            if self.hdrActive { self.layer.edrMetadata = self.makeEDR(meta) }
+            #endif
+        }
+    }
+
+    /// Draw one decoded frame to the next drawable and present it. MAIN THREAD (the display link).
+    /// `isHDR` selects the 10-bit BT.2020 PQ path vs the 8-bit BT.709 path and is reconciled with the
+    /// layer config via `configure`. Returns true on success; false when there's no drawable yet, a
+    /// texture couldn't be made, or Metal errored — the caller then doesn't stamp a present.
     @discardableResult
     public func render(_ pixelBuffer: CVPixelBuffer, isHDR: Bool = false) -> Bool {
+        // Reconcile the layer with the decoded frame's HDR-ness (handles a mid-session SDR↔HDR flip).
         configure(hdr: isHDR)
-        // P010 stores 10-bit luma/chroma in 16-bit samples → R16/RG16; NV12 is 8-bit → R8/RG8.
+
-        let lumaFmt: MTLPixelFormat = isHDR ? .r16Unorm : .r8Unorm
+        // P010/x444 store 10-bit luma/chroma in 16-bit samples → R16/RG16; NV12/444v is 8-bit → R8/RG8.
-        let chromaFmt: MTLPixelFormat = isHDR ? .rg16Unorm : .rg8Unorm
+        // Derived from the actual decoded buffer so a 4:4:4 (full chroma plane) frame just works.
+        let pf = CVPixelBufferGetPixelFormatType(pixelBuffer)
+        let tenBit =
+            pf == kCVPixelFormatType_420YpCbCr10BiPlanarVideoRange
+            || pf == kCVPixelFormatType_420YpCbCr10BiPlanarFullRange
+            || pf == kCVPixelFormatType_444YpCbCr10BiPlanarVideoRange
+            || pf == kCVPixelFormatType_444YpCbCr10BiPlanarFullRange
         guard let textureCache,
-              let luma = makeTexture(pixelBuffer, plane: 0, format: lumaFmt, cache: textureCache),
+              let luma = makeTexture(
-              let chroma = makeTexture(pixelBuffer, plane: 1, format: chromaFmt, cache: textureCache)
+                pixelBuffer, plane: 0, format: tenBit ? .r16Unorm : .r8Unorm, cache: textureCache),
+              let chroma = makeTexture(
+                pixelBuffer, plane: 1, format: tenBit ? .rg16Unorm : .rg8Unorm, cache: textureCache)
         else { return false }
 
-        // The hosting view owns drawableSize (aspect-fit to its bounds); skip until it's laid
+        // Size the drawable to the decoded frame so the fullscreen triangle samples 1:1 (pixel-exact);
-        // out. The fullscreen triangle scales the decoded texture to fill the drawable.
+        // the layer's contentsGravity then scales it to the on-screen bounds via the system compositor
-        guard layer.drawableSize.width > 0, layer.drawableSize.height > 0,
+        // (matching stage-1). drawableSize does NOT track bounds (defaults to 0), so set it BEFORE
-              let drawable = layer.nextDrawable(),
+        // nextDrawable; re-set only on a change (first frame / Reconfigure / HDR flip).
+        let decodedSize = CGSize(
+            width: CVPixelBufferGetWidth(pixelBuffer), height: CVPixelBufferGetHeight(pixelBuffer))
+        if layer.drawableSize != decodedSize { layer.drawableSize = decodedSize }
+        #if DEBUG
+        logSizeIfChanged(decoded: decodedSize)
+        #endif
+        guard let drawable = layer.nextDrawable(),
               let commandBuffer = queue.makeCommandBuffer()
         else { return false }
 
@@ -186,24 +320,23 @@ public final class MetalVideoPresenter {
         guard let encoder = commandBuffer.makeRenderCommandEncoder(descriptor: pass) else {
             return false
         }
-        encoder.setRenderPipelineState(isHDR ? pipelineHDR : pipelineSDR)
+        encoder.setRenderPipelineState(hdrActive ? pipelineHDR : pipelineSDR)
         encoder.setFragmentTexture(CVMetalTextureGetTexture(luma), index: 0)
         encoder.setFragmentTexture(CVMetalTextureGetTexture(chroma), index: 1)
         encoder.drawPrimitives(type: .triangle, vertexStart: 0, vertexCount: 3)
         encoder.endEncoding()
         commandBuffer.present(drawable) // present at the next vsync — lowest latency
-        // Hold the CVMetalTextures + the source pixel buffer (its IOSurface) alive until the GPU
+        // Hold the CVMetalTextures + source pixel buffer (its IOSurface) alive until the GPU finishes
-        // finishes sampling — releasing them at scope exit could free the backing mid-read.
+        // sampling — releasing them at scope exit could free the backing mid-read.
         commandBuffer.addCompletedHandler { _ in _ = (luma, chroma, pixelBuffer) }
         commandBuffer.commit()
         return true
     }
 
-    /// Returns the CVMetalTexture (not just its MTLTexture) so the caller can keep it alive past
+    /// Returns the CVMetalTexture (not just its MTLTexture) so the caller can keep it alive past the
-    /// the draw — the MTLTexture is only valid while its CVMetalTexture is retained.
+    /// draw — the MTLTexture is only valid while its CVMetalTexture is retained.
     private func makeTexture(
-        _ pixelBuffer: CVPixelBuffer, plane: Int, format: MTLPixelFormat,
+        _ pixelBuffer: CVPixelBuffer, plane: Int, format: MTLPixelFormat, cache: CVMetalTextureCache
-        cache: CVMetalTextureCache
     ) -> CVMetalTexture? {
         let w = CVPixelBufferGetWidthOfPlane(pixelBuffer, plane)
         let h = CVPixelBufferGetHeightOfPlane(pixelBuffer, plane)
@@ -215,5 +348,16 @@ public final class MetalVideoPresenter {
         else { return nil }
         return cvTexture
     }
+
+    #if DEBUG
+    private func logSizeIfChanged(decoded: CGSize) {
+        let sig = "\(Int(decoded.width))x\(Int(decoded.height))|hdr\(hdrActive ? 1 : 0)"
+        if sig != lastSizeSig {
+            lastSizeSig = sig
+            let msg = "stage2: decoded \(Int(decoded.width))x\(Int(decoded.height)) hdr=\(hdrActive)"
+            presenterLog.info("\(msg, privacy: .public)")
+        }
+    }
+    #endif
 }
 #endif

clients/apple/Sources/PunktfunkKit/PointerLockChain.swift

@@ -0,0 +1,94 @@
+// Steers the system's iPad pointer-lock resolution down to a chosen "anchor" view controller.
+//
+// `UIViewController.prefersPointerLocked` is resolved the same way as the status bar: the system
+// walks DOWN from the window's root view controller through `childViewControllerForPointerLock`.
+// SwiftUI's hosting / container view controllers do NOT forward that query to their children, so a
+// `UIViewControllerRepresentable` controller buried in the SwiftUI tree (our StreamViewController)
+// is never consulted — its `prefersPointerLocked = true` is silently ignored and a Magic Keyboard
+// trackpad / mouse falls through to the absolute-pointer path instead of being captured.
+//
+// Swizzling the DEFAULT implementation isn't enough: the controllers that break the chain
+// (UIHostingController and SwiftUI's internal containers) provide their OWN implementation of the
+// property, so a base-class swizzle never runs for them. Instead we walk UP the LIVE `parent`
+// chain from the anchor to the window root and, on each real ancestor, force
+// `childViewControllerForPointerLock` to return the next controller toward the anchor. Each forced
+// value is a genuine direct child (we follow the actual containment chain), so the system's
+// downward walk reaches the anchor and reads its `prefersPointerLocked`.
+//
+// The forcing is per-INSTANCE — an associated object — gated behind a one-time per-CLASS IMP
+// swizzle. Only the specific controllers in the anchor's chain are affected; every other instance
+// of those classes keeps its original behavior (associated object nil → original IMP). The forced
+// values are cleared on disengage so the long-lived SwiftUI parents don't retain a stale controller
+// across sessions. Only the PUBLIC `childViewControllerForPointerLock` selector is touched
+// (App-Store-safe; no private API).
+
+#if os(iOS)
+import ObjectiveC
+import UIKit
+
+enum PointerLockChain {
+    private static var forcedChildKey: UInt8 = 0
+    /// Classes whose `childViewControllerForPointerLock` we've already IMP-swizzled (keyed by the
+    /// class object). Main-thread only — pointer-lock resolution and capture toggles are all main.
+    private static var swizzledClasses = Set<ObjectIdentifier>()
+    /// Ancestors we've stamped with a forced child this engagement, held weakly so a deallocated
+    /// SwiftUI controller drops out on its own (no dangling). disengage() clears every one — even
+    /// if the live `parent` chain has since broken — so a stamped parent can never retain a stale
+    /// controller subtree across sessions. One anchor is ever engaged at a time.
+    private static let stampedParents = NSHashTable<UIViewController>.weakObjects()
+
+    private static func forcedChild(of vc: UIViewController) -> UIViewController? {
+        objc_getAssociatedObject(vc, &forcedChildKey) as? UIViewController
+    }
+
+    private static func setForcedChild(_ child: UIViewController?, on vc: UIViewController) {
+        // RETAIN: while steering, the parent must keep the toward-anchor child alive. It's also
+        // already a strong child of `vc` via UIKit containment, so this adds no cycle (the reverse
+        // `.parent` link is weak), and disengage() always clears it — so it can't outlive a session.
+        objc_setAssociatedObject(vc, &forcedChildKey, child, .OBJC_ASSOCIATION_RETAIN_NONATOMIC)
+    }
+
+    /// Ensure `cls`'s `childViewControllerForPointerLock` getter consults the per-instance forced
+    /// child first, falling back to the class's original implementation. Idempotent per class.
+    private static func ensureSwizzled(_ cls: AnyClass) {
+        let id = ObjectIdentifier(cls)
+        guard !swizzledClasses.contains(id) else { return }
+        swizzledClasses.insert(id)
+        let selector = #selector(getter: UIViewController.childViewControllerForPointerLock)
+        guard let method = class_getInstanceMethod(cls, selector) else { return }
+        let originalIMP = method_getImplementation(method)
+        typealias OriginalFn = @convention(c) (AnyObject, Selector) -> UIViewController?
+        let original = unsafeBitCast(originalIMP, to: OriginalFn.self)
+        let forwarding: @convention(block) (UIViewController) -> UIViewController? = { vc in
+            if let forced = forcedChild(of: vc) { return forced }
+            return original(vc, selector)
+        }
+        method_setImplementation(method, imp_implementationWithBlock(forwarding))
+    }
+
+    /// Force every ancestor of `anchor` to forward pointer-lock resolution toward it, then ask the
+    /// system to re-resolve. No-op when `anchor` isn't in a view-controller hierarchy yet (it
+    /// re-runs from the anchor's appearance/parent callbacks once it is).
+    static func engage(_ anchor: UIViewController) {
+        disengage(anchor) // clear any prior engagement first (reparent / re-anchor)
+        var child = anchor
+        while let parent = child.parent {
+            ensureSwizzled(object_getClass(parent)!)
+            setForcedChild(child, on: parent)
+            stampedParents.add(parent)
+            child = parent
+        }
+        anchor.setNeedsUpdateOfPrefersPointerLocked()
+    }
+
+    /// Clear the forced forwarding on every stamped ancestor (so the SwiftUI parents stop retaining
+    /// the anchor's subtree) and re-resolve to drop the lock.
+    static func disengage(_ anchor: UIViewController) {
+        for parent in stampedParents.allObjects {
+            setForcedChild(nil, on: parent)
+        }
+        stampedParents.removeAllObjects()
+        anchor.setNeedsUpdateOfPrefersPointerLocked()
+    }
+}
+#endif

clients/apple/Sources/PunktfunkKit/Probe444Blobs.swift

@@ -0,0 +1,36 @@
+// Synthetic 4:4:4 HEVC keyframes used only by `Stage444Probe` to probe decode capability.
+//
+// Each is the first IDR access unit (VPS + SPS + PPS + IDR slice, Annex-B) of a 256×256 HEVC
+// Range-Extensions clip — `chroma_format_idc = 3` — generated offline with libx265:
+//   ffmpeg -f lavfi -i color=c=gray:s=256x256:r=30:d=0.1 -frames:v 3 \
+//          -pix_fmt yuv444p[10le] -c:v libx265 \
+//          -x265-params keyint=1:min-keyint=1:no-info=1:repeat-headers=1:aud=0 -f hevc out.hevc
+// 256×256 clears the hardware decoder's minimum-dimension floor (a 16×16 clip is rejected for every
+// chroma format). Validated to hardware-decode to `444v`/`x444` on Apple Silicon (M3).
+enum Probe444Blobs {
+    /// 256×256 HEVC Range-Extensions 4:4:4 keyframe (Annex-B): 134 bytes.
+    static let au444_8bit: [UInt8] = [
+        0x00, 0x00, 0x00, 0x01, 0x40, 0x01, 0x0c, 0x01, 0xff, 0xff, 0x04, 0x08, 0x00, 0x00, 0x03, 0x00,
+        0x9e, 0x28, 0x00, 0x00, 0x03, 0x00, 0x00, 0x3c, 0xba, 0x02, 0x40, 0x00, 0x00, 0x00, 0x01, 0x42,
+        0x01, 0x01, 0x04, 0x08, 0x00, 0x00, 0x03, 0x00, 0x9e, 0x28, 0x00, 0x00, 0x03, 0x00, 0x00, 0x3c,
+        0x90, 0x01, 0x01, 0x00, 0x80, 0xb2, 0xdd, 0x49, 0x26, 0x57, 0x80, 0xb4, 0x04, 0x00, 0x00, 0x03,
+        0x00, 0x04, 0x00, 0x00, 0x03, 0x00, 0x78, 0x20, 0x00, 0x00, 0x00, 0x01, 0x44, 0x01, 0xc1, 0x72,
+        0x86, 0x0c, 0x06, 0x24, 0x00, 0x00, 0x00, 0x01, 0x28, 0x01, 0xaf, 0x72, 0x15, 0xe8, 0x34, 0xeb,
+        0xae, 0xfb, 0xfe, 0x75, 0x57, 0xca, 0xc1, 0x71, 0x43, 0x16, 0xf5, 0xc2, 0x40, 0xbd, 0x80, 0xa6,
+        0x65, 0x35, 0x20, 0x28, 0x81, 0xa2, 0x5e, 0xc0, 0x93, 0x04, 0x10, 0x9b, 0x00, 0x34, 0xe0, 0x87,
+        0x00, 0x00, 0x03, 0x00, 0x5b, 0x40,
+    ]
+
+    /// 256×256 HEVC Range-Extensions 4:4:4 10-bit keyframe (Annex-B): 133 bytes.
+    static let au444_10bit: [UInt8] = [
+        0x00, 0x00, 0x00, 0x01, 0x40, 0x01, 0x0c, 0x01, 0xff, 0xff, 0x04, 0x08, 0x00, 0x00, 0x03, 0x00,
+        0x9c, 0x28, 0x00, 0x00, 0x03, 0x00, 0x00, 0x3c, 0xba, 0x02, 0x40, 0x00, 0x00, 0x00, 0x01, 0x42,
+        0x01, 0x01, 0x04, 0x08, 0x00, 0x00, 0x03, 0x00, 0x9c, 0x28, 0x00, 0x00, 0x03, 0x00, 0x00, 0x3c,
+        0x90, 0x01, 0x01, 0x00, 0x80, 0x9b, 0x2d, 0xd4, 0x92, 0x65, 0x78, 0x0b, 0x40, 0x40, 0x00, 0x00,
+        0x03, 0x00, 0x40, 0x00, 0x00, 0x07, 0x82, 0x00, 0x00, 0x00, 0x01, 0x44, 0x01, 0xc1, 0x72, 0x86,
+        0x0c, 0x06, 0x24, 0x00, 0x00, 0x00, 0x01, 0x28, 0x01, 0xaf, 0x72, 0x15, 0xe8, 0x34, 0xeb, 0xae,
+        0xfb, 0xfe, 0x75, 0x57, 0xca, 0xc1, 0x71, 0x43, 0x16, 0xf5, 0xc2, 0x40, 0xbd, 0x80, 0xa6, 0x65,
+        0x35, 0x20, 0x28, 0x81, 0xa2, 0x5e, 0xc0, 0x93, 0x04, 0x10, 0x9b, 0x00, 0x34, 0xe0, 0x87, 0x00,
+        0x00, 0x03, 0x00, 0x5b, 0x40,
+    ]
+}

clients/apple/Sources/PunktfunkKit/PunktfunkConnection.swift

@@ -182,6 +182,11 @@ public final class PunktfunkConnection {
         case dualSense = 2
         case xboxOne = 3
         case dualShock4 = 4
+        // Valve Steam Controller / Steam Deck (Linux UHID hid-steam hosts). Parity only on Apple —
+        // GameController never surfaces a 0x28DE HID device, so the client can't capture one; these
+        // exist so the resolved type round-trips and name parsing matches the host.
+        case steamController = 5
+        case steamDeck = 6
 
         /// Loose name parsing for env/dev hooks, mirroring the host's
         /// `GamepadPref::from_name`.
@@ -192,6 +197,8 @@ public final class PunktfunkConnection {
             case "dualsense", "ds", "ds5", "ps5": self = .dualSense
             case "xboxone", "xbox-one", "xboxseries", "series": self = .xboxOne
             case "dualshock4", "dualshock", "ds4", "ps4": self = .dualShock4
+            case "steamdeck", "steam-deck", "deck": self = .steamDeck
+            case "steamcontroller", "steam-controller", "steamcon": self = .steamController
             default: return nil
             }
         }
@@ -231,6 +238,13 @@ public final class PunktfunkConnection {
     public private(set) var colorFullRange: Bool = false
     /// Encoded bit depth (8 or 10).
     public private(set) var bitDepth: UInt8 = 8
+    /// The chroma subsampling the host resolved for this session, as the HEVC `chroma_format_idc`:
+    /// `1` = 4:2:0 (every pre-4:4:4 host, and the back-compat default) or `3` = full-chroma 4:4:4
+    /// (only when this client advertised `videoCap444` *and* the host could open a real 4:4:4
+    /// encoder). Drive the decoder's requested pixel format from this. See `isChroma444`.
+    public private(set) var chromaFormat: UInt8 = 1
+    /// Convenience: the resolved stream is full-chroma 4:4:4 (`chroma_format_idc == 3`).
+    public var isChroma444: Bool { chromaFormat == 3 }
     /// True when the negotiated stream is HDR (PQ or HLG transfer) — drive an HDR present path and
     /// drain `nextHdrMeta`.
     public var isHDR: Bool { colorTransfer == 16 || colorTransfer == 18 }
@@ -327,6 +341,9 @@ public final class PunktfunkConnection {
         colorMatrix = mtx
         colorFullRange = fullRange != 0
         bitDepth = depth
+        var cf: UInt8 = 1
+        _ = punktfunk_connection_chroma_format(handle, &cf)
+        chromaFormat = cf
         var ac: UInt8 = 2
         _ = punktfunk_connection_audio_channels(handle, &ac)
         resolvedAudioChannels = ac
@@ -598,6 +615,10 @@ public final class PunktfunkConnection {
     public static let videoCap10Bit: UInt8 = UInt8(PUNKTFUNK_VIDEO_CAP_10BIT)
     /// Video-capability bit: the client can present BT.2020 PQ HDR10 (implies 10-bit).
     public static let videoCapHDR: UInt8 = UInt8(PUNKTFUNK_VIDEO_CAP_HDR)
+    /// Video-capability bit: the client can decode a full-chroma 4:4:4 HEVC stream (Range
+    /// Extensions). Advertise only when the device can *hardware*-decode it (`Stage444Probe`);
+    /// the host then emits 4:4:4 only if it too opted in. `chromaFormat` reflects the real value.
+    public static let videoCap444: UInt8 = UInt8(PUNKTFUNK_VIDEO_CAP_444)
 
     /// Static HDR mastering metadata (SMPTE ST.2086 + content light level) the host sent for an HDR
     /// session. Mirrors the wire/ABI `PunktfunkHdrMeta`; primaries are in ST.2086 **G, B, R** order,

clients/apple/Sources/PunktfunkKit/Resources/Fonts/Geist-OFL.txt

@@ -0,0 +1,93 @@
+Copyright 2024 The Geist Project Authors (https://github.com/vercel/geist-font)
+
+This Font Software is licensed under the SIL Open Font License, Version 1.1.
+This license is copied below, and is also available with a FAQ at:
+https://openfontlicense.org
+
+
+-----------------------------------------------------------
+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
+-----------------------------------------------------------
+
+PREAMBLE
+The goals of the Open Font License (OFL) are to stimulate worldwide
+development of collaborative font projects, to support the font creation
+efforts of academic and linguistic communities, and to provide a free and
+open framework in which fonts may be shared and improved in partnership
+with others.
+
+The OFL allows the licensed fonts to be used, studied, modified and
+redistributed freely as long as they are not sold by themselves. The
+fonts, including any derivative works, can be bundled, embedded, 
+redistributed and/or sold with any software provided that any reserved
+names are not used by derivative works. The fonts and derivatives,
+however, cannot be released under any other type of license. The
+requirement for fonts to remain under this license does not apply
+to any document created using the fonts or their derivatives.
+
+DEFINITIONS
+"Font Software" refers to the set of files released by the Copyright
+Holder(s) under this license and clearly marked as such. This may
+include source files, build scripts and documentation.
+
+"Reserved Font Name" refers to any names specified as such after the
+copyright statement(s).
+
+"Original Version" refers to the collection of Font Software components as
+distributed by the Copyright Holder(s).
+
+"Modified Version" refers to any derivative made by adding to, deleting,
+or substituting -- in part or in whole -- any of the components of the
+Original Version, by changing formats or by porting the Font Software to a
+new environment.
+
+"Author" refers to any designer, engineer, programmer, technical
+writer or other person who contributed to the Font Software.
+
+PERMISSION & CONDITIONS
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of the Font Software, to use, study, copy, merge, embed, modify,
+redistribute, and sell modified and unmodified copies of the Font
+Software, subject to the following conditions:
+
+1) Neither the Font Software nor any of its individual components,
+in Original or Modified Versions, may be sold by itself.
+
+2) Original or Modified Versions of the Font Software may be bundled,
+redistributed and/or sold with any software, provided that each copy
+contains the above copyright notice and this license. These can be
+included either as stand-alone text files, human-readable headers or
+in the appropriate machine-readable metadata fields within text or
+binary files as long as those fields can be easily viewed by the user.
+
+3) No Modified Version of the Font Software may use the Reserved Font
+Name(s) unless explicit written permission is granted by the corresponding
+Copyright Holder. This restriction only applies to the primary font name as
+presented to the users.
+
+4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
+Software shall not be used to promote, endorse or advertise any
+Modified Version, except to acknowledge the contribution(s) of the
+Copyright Holder(s) and the Author(s) or with their explicit written
+permission.
+
+5) The Font Software, modified or unmodified, in part or in whole,
+must be distributed entirely under this license, and must not be
+distributed under any other license. The requirement for fonts to
+remain under this license does not apply to any document created
+using the Font Software.
+
+TERMINATION
+This license becomes null and void if any of the above conditions are
+not met.
+
+DISCLAIMER
+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
+OTHER DEALINGS IN THE FONT SOFTWARE.

clients/apple/Sources/PunktfunkKit/SessionAudio.swift

@@ -177,6 +177,16 @@ public final class SessionAudio {
     private var playbackEngine: AVAudioEngine?
     private var captureEngine: AVAudioEngine?
     private var drainStarted = false
+    #if !os(macOS)
+    /// AVAudioSession `setCategory`/`setActive` are synchronous and block on the audio server, so
+    /// they must not run on the main thread (UI stall — AVFoundation warns about it). PROCESS-WIDE
+    /// (static) so every SessionAudio shares one serial queue: the AVAudioSession is a process
+    /// singleton, and across a reconnect the old session's deactivate must be ordered before the
+    /// new session's activate (a per-instance queue would let them race and leave the new session's
+    /// audio deactivated). stop() enqueues its deactivate promptly so it lands before the next
+    /// session's activate.
+    private static let sessionQueue = DispatchQueue(label: "io.unom.punktfunk.audio.session")
+    #endif
 
     public init(connection: PunktfunkConnection) {
         self.connection = connection
@@ -189,37 +199,60 @@ public final class SessionAudio {
         flag.stop()
     }
 
-    /// Start playback (and, if enabled+authorized, the mic uplink). Empty UIDs = system
+    /// Start playback (and, if enabled+authorized, the mic uplink). Empty UIDs = system default
-    /// default device; on iOS the UIDs are ignored entirely (routes are
+    /// device; on iOS the UIDs are ignored entirely (routes are AVAudioSession-managed). On macOS
-    /// AVAudioSession-managed). Main thread (engine setup); returns after the engines
+    /// the engines start synchronously on the caller's (main) thread. On iOS/tvOS start() is
-    /// start — the mic may start slightly later if the permission prompt is pending.
+    /// ASYNCHRONOUS: it activates the AVAudioSession off the main thread, then starts the engines on
+    /// a later main-queue hop (gated by `!flag.isStopped`) — so playback is live shortly after, not
+    /// on return. The mic may start later still if the permission prompt is pending.
     public func start(speakerUID: String, micUID: String, micEnabled: Bool) {
-        #if os(iOS)
+        #if os(macOS)
-        // Route + policy live in the session, not per-engine: stereo playback, mic
+        // No AVAudioSession on macOS — start the engines directly (caller's thread, as before).
-        // capture when enabled, Bluetooth allowed. Failure is non-fatal (defaults).
+        startEngines(speakerUID: speakerUID, micUID: micUID, micEnabled: micEnabled)
+        #else
+        // Configure + activate the session OFF the main thread (it blocks on the audio server),
+        // then start the engines back on the main thread once it's active — engine routing/format
+        // depend on the active session. A stop() racing in between is caught by the flag guard.
+        Self.sessionQueue.async { [weak self] in
+            guard let self else { return }
+            self.activateAudioSession(micEnabled: micEnabled)
+            DispatchQueue.main.async { [weak self] in
+                guard let self, !self.flag.isStopped else { return }
+                self.startEngines(speakerUID: speakerUID, micUID: micUID, micEnabled: micEnabled)
+            }
+        }
+        #endif
+    }
+
+    #if !os(macOS)
+    /// Route + policy live in the session, not per-engine: stereo playback, mic capture when
+    /// enabled, Bluetooth allowed. Failure is non-fatal (defaults). Runs on `sessionQueue`.
+    private func activateAudioSession(micEnabled: Bool) {
         let session = AVAudioSession.sharedInstance()
         do {
+            #if os(iOS)
             if micEnabled {
-                // .defaultToSpeaker: .playAndRecord otherwise routes to the iPhone
+                // .defaultToSpeaker: .playAndRecord otherwise routes to the iPhone EARPIECE; only
-                // EARPIECE; only affects the built-in route (headphones/BT still win).
+                // affects the built-in route (headphones/BT still win).
                 try session.setCategory(
                     .playAndRecord, mode: .default,
                     options: [.allowBluetoothA2DP, .defaultToSpeaker])
             } else {
                 try session.setCategory(.playback, mode: .default)
             }
+            #else // tvOS — no app-accessible mic
+            try session.setCategory(.playback, mode: .default)
+            #endif
             try session.setActive(true)
         } catch {
             log.warning("AVAudioSession setup failed: \(error.localizedDescription)")
         }
-        #elseif os(tvOS)
-        do {
-            try AVAudioSession.sharedInstance().setCategory(.playback, mode: .default)
-            try AVAudioSession.sharedInstance().setActive(true)
-        } catch {
-            log.warning("AVAudioSession setup failed: \(error.localizedDescription)")
     }
     #endif
+
+    /// Build + start the playback engine (and the mic uplink when enabled + authorized). Main
+    /// thread (engine setup); on iOS/tvOS the session is already active by the time this runs.
+    private func startEngines(speakerUID: String, micUID: String, micEnabled: Bool) {
         startPlayback(speakerUID: speakerUID)
         #if os(tvOS)
         // No app-accessible microphone input on tvOS — playback only.
@@ -258,19 +291,24 @@ public final class SessionAudio {
             capture.stop()
         }
         playback?.stop()
-        if wasDraining {
-            _ = drainDone.wait(timeout: .now() + .milliseconds(400))
-        }
         #if !os(macOS)
-        // Release the session so audio we interrupted (Music, podcasts) gets its
+        // Release the session so audio we interrupted (Music, podcasts) gets its resume cue. Like
-        // resume cue.
+        // activation, setActive is synchronous/blocking — run it on the shared serial session queue
+        // (off the main thread). Enqueued HERE — engines already stopped, and BEFORE the drain wait
+        // below — so across a reconnect it lands ahead of the next session's activate on the shared
+        // queue (otherwise a deferred deactivate could deactivate the new session). Fire-and-forget.
+        Self.sessionQueue.async {
             do {
                 try AVAudioSession.sharedInstance().setActive(
                     false, options: .notifyOthersOnDeactivation)
             } catch {
                 log.warning("AVAudioSession deactivation failed: \(error.localizedDescription)")
             }
+        }
         #endif
+        if wasDraining {
+            _ = drainDone.wait(timeout: .now() + .milliseconds(400))
+        }
     }
 
     // MARK: - Playback (host → speaker)

clients/apple/Sources/PunktfunkKit/Stage2Pipeline.swift

@@ -1,21 +1,21 @@
-// Stage-2 presenter orchestrator: a pump thread pulls AUs → VideoDecoder; the decoder's async
+// Stage-2 presenter orchestrator: a pump thread pulls AUs → VideoDecoder; the decoder's async output
-// output drops the newest decoded frame into a 1-slot ring; the hosting view's display link
+// drops the newest decoded frame into a 1-slot ring; the hosting view's display link calls `renderTick`
-// calls `renderTick` once per vsync to draw + present the newest ready frame and stamp
+// once per vsync to draw + present the newest ready frame and stamp capture→present. Mirrors
-// capture→present. Mirrors StreamPump's lifecycle (one per start; cancel is permanent).
+// StreamPump's lifecycle (one per start; cancel is permanent).
 //
-// Threading: the pump runs on its own thread; the decoder callback on a VT thread; `renderTick`
+// Threading: the pump runs on its own thread; the decoder callback on a VT thread; `renderTick` +
-// + `setDrawableSize` + `start`/`stop` on the MAIN thread (the view's CADisplayLink fires there).
+// `start`/`stop` on the MAIN thread (the view's CADisplayLink fires there). Only the ring (lock-guarded)
-// Only the ring + decoder cross threads and both are internally locked.
+// and the decoder/presenter (internally locked / main-hopped) cross threads.
 
 #if canImport(Metal) && canImport(QuartzCore)
 import AVFoundation
 import Foundation
 import QuartzCore
 
-/// Weak-target wrapper for CADisplayLink. The link retains its target, so targeting a view
+/// Weak-target wrapper for CADisplayLink. The link retains its target, so targeting a view directly
-/// directly makes a `view → link → view` cycle that only `invalidate()` breaks — if a teardown
+/// makes a `view → link → view` cycle that only `invalidate()` breaks — if a teardown is ever missed
-/// is ever missed the view leaks and keeps ticking. This proxy holds the handler weakly, so the
+/// the view leaks and keeps ticking. This proxy holds the handler weakly, so the view can deallocate
-/// view can deallocate and its `deinit` invalidate the link.
+/// and its `deinit` invalidate the link.
 public final class DisplayLinkProxy: NSObject {
     private let onTick: (CADisplayLink) -> Void
     public init(_ onTick: @escaping (CADisplayLink) -> Void) { self.onTick = onTick }
@@ -44,10 +44,10 @@ private final class PumpToken: @unchecked Sendable {
     func cancel() { lock.lock(); live = false; lock.unlock() }
 }
 
-/// Throttled host keyframe requests for decode recovery. The decoder's async error callback
+/// Throttled host keyframe requests for decode recovery. The decoder's async error callback (a VT
-/// (a VT thread) and the pump thread (a submit failure) both signal a wedge; this coalesces
+/// thread) and the pump thread (a submit failure) both signal a wedge; this coalesces them so the
-/// them so the control stream isn't flooded while the decode stays stalled for several frames
+/// control stream isn't flooded while the decode stays stalled for several frames until the requested
-/// until the requested IDR lands. Bound to the live connection in `start`, unbound in `stop`.
+/// IDR lands. Bound to the live connection in `start`, unbound in `stop`.
 private final class KeyframeRecovery: @unchecked Sendable {
     private let lock = NSLock()
     private var connection: PunktfunkConnection?
@@ -60,7 +60,7 @@ private final class KeyframeRecovery: @unchecked Sendable {
     func request() {
         lock.lock()
         let now = DispatchTime.now().uptimeNanoseconds
-        let due = lastNs == 0 || now &- lastNs > 250_000_000 // ≥ 250 ms since the last request
+        let due = lastNs == 0 || now &- lastNs > 100_000_000 // ≥ 100 ms since the last request
         if due { lastNs = now }
         let conn = due ? connection : nil
         lock.unlock()
@@ -76,30 +76,36 @@ public final class Stage2Pipeline {
     private let recovery = KeyframeRecovery()
     private var token = PumpToken()
     private var offsetNs: Int64 = 0
+    /// Signalled when the pump thread exits, so `stop()` can join it (bounded) before `decoder.reset()`
+    /// — otherwise a pump iteration already past its `token.isLive` check can rebuild a decode session
+    /// right after the reset (a brief orphan session). `pumpJoinable` is armed by `start`, consumed by
+    /// the first `stop` (so the idempotent second `stop`/deinit doesn't block on an already-drained
+    /// semaphore). start/stop are sequential lifecycle calls, so the plain flag is safe.
+    private let pumpStopped = DispatchSemaphore(value: 0)
+    private var pumpJoinable = false
 
-    /// The Metal layer the hosting view installs + sizes. nil-init fails when Metal is
+    /// The Metal layer the hosting view installs + sizes.
-    /// unavailable so the caller can fall back to stage-1.
     public var layer: CAMetalLayer { presenter.layer }
 
-    /// `presentMeter` records capture→present (the glass-to-glass term). Returns nil if Metal
+    /// `presentMeter` records capture→present (the glass-to-glass term). Returns nil if Metal can't be
-    /// can't be set up (headless / no GPU) — caller falls back to the stage-1 presenter.
+    /// set up (headless / no GPU) — caller falls back to the stage-1 presenter.
     public init?(presentMeter: LatencyMeter) {
-        guard let presenter = MetalVideoPresenter() else { return nil }
+        guard let presenter = MetalVideoPresenter.make() else { return nil }
         self.presenter = presenter
         self.presentMeter = presentMeter
         let ring = ring
         let recovery = recovery
         self.decoder = VideoDecoder(
             onDecoded: { ring.submit($0) },
-            // Async decode failure (a bad P-frame referencing a lost/corrupt IDR): the pump
+            // Async decode failure (a bad P-frame referencing a lost/corrupt IDR): the pump resets to
-            // resets to re-gate on the next IDR, and we ask the host to send one now (infinite
+            // re-gate on the next IDR, and we ask the host to send one now (infinite GOP — it wouldn't
-            // GOP — it wouldn't otherwise come soon). Throttled in KeyframeRecovery.
+            // otherwise come soon). Throttled in KeyframeRecovery.
             onDecodeError: { _ in recovery.request() })
     }
 
-    /// Start pulling AUs into the decoder. `onFrame` fires per AU at receipt (capture→client
+    /// Start pulling AUs into the decoder. MAIN THREAD. `onFrame` fires per AU at receipt (capture→client
-    /// meter, exactly as stage-1); `onSessionEnd` on close. `clockOffsetNs` (host minus client)
+    /// meter, exactly as stage-1); `onSessionEnd` on close. `clockOffsetNs` (host minus client) makes the
-    /// makes the present stamp cross-machine valid.
+    /// present stamp cross-machine valid.
     public func start(
         connection: PunktfunkConnection,
         onFrame: (@Sendable (AccessUnit) -> Void)?,
@@ -108,43 +114,70 @@ public final class Stage2Pipeline {
         offsetNs = connection.clockOffsetNs
         recovery.bind(connection) // arm host-keyframe recovery for this session
         token = PumpToken() // fresh token per start — cancel is permanent (like StreamPump)
+
+        // Configure the decoder's chroma + the layer's initial colorimetry before the first frame. The
+        // chroma subsampling drives only the decode pixel format (orthogonal to HDR/depth); the HDR
+        // config is the Welcome's latched value, which a mid-session flip then overrides per-frame.
+        decoder.setChroma444(connection.isChroma444)
+        presenter.configure(hdr: connection.isHDR)
+
         let token = token
         let decoder = decoder
         let recovery = recovery
+        let presenter = presenter
+        let pumpStopped = pumpStopped
         let thread = Thread {
+            defer { pumpStopped.signal() } // let stop() join the pump (bounded) before decoder.reset()
             var format: CMVideoFormatDescription?
             var lastFramesDropped = connection.framesDropped()
+            // Persistent recovery WANT, not a one-shot edge (see StreamPump for the full rationale):
+            // keep asking until an IDR lands so a request swallowed by the throttle is re-sent.
+            var awaitingIDR = false
+            // 4:4:4 backstop: a run of decode/create failures in a 4:4:4 session means this device can't
+            // decode 4:4:4 at the negotiated resolution (the HW probe clears the common case but not a
+            // resolution-ceiling miss). End cleanly instead of looping on a black screen.
+            var decodeFailRun = 0
             while token.isLive {
                 do {
-                    // Loss recovery (the primary recovery path). The reassembler drops unrecoverable
+                    // Loss recovery (the primary path). The reassembler drops unrecoverable AUs and the
-                    // AUs (framesDropped) and the decoder then conceals the reference-missing delta
+                    // decoder conceals the reference-missing deltas — often WITHOUT an error callback —
-                    // frames that follow — often rendering them WITHOUT an error callback — so the
+                    // so key off the drop count climbing, then keep asking (awaitingIDR) until a fresh
-                    // onDecodeError trigger rarely fires after a real network blip. Ask the host for
+                    // IDR re-anchors decode.
-                    // a fresh IDR whenever the drop count climbs (throttled in KeyframeRecovery).
-                    // Polled every iteration so a total-loss drought recovers the moment packets
-                    // resume and the reassembler counts the gap.
                     let dropped = connection.framesDropped()
                     if dropped > lastFramesDropped {
                         lastFramesDropped = dropped
-                        recovery.request()
+                        awaitingIDR = true
                     }
-                    // Drain any HDR mastering-metadata update (0xCE) and hand it to the decoder, which
+                    if awaitingIDR { recovery.request() }
-                    // attaches it to subsequent HDR frames. Non-blocking; only HDR sessions emit these.
+                    // Drain HDR mastering metadata (0xCE) and hand it to the PRESENTER (→ CAEDRMetadata).
-                    if connection.isHDR, let meta = try? connection.nextHdrMeta(timeoutMs: 0) {
+                    // Polled UNCONDITIONALLY (not gated on connection.isHDR, the fixed Welcome flag): the
-                        decoder.setHdrMeta(meta)
+                    // host sends 0xCE only for HDR, INCLUDING a mid-session SDR→HDR transition (a game
+                    // entering HDR — the host re-inits its encoder) the Welcome flag would never reflect.
+                    // Non-blocking; nil for an SDR stream.
+                    if let meta = try? connection.nextHdrMeta(timeoutMs: 0) {
+                        presenter.setHdrMeta(meta)
                     }
                     guard let au = try connection.nextAU(timeoutMs: 100) else { continue }
                     onFrame?(au)
                     if let f = AnnexB.formatDescription(fromIDR: au.data) {
                         format = f          // refreshed on every IDR (mode changes included)
+                        awaitingIDR = false // a fresh IDR re-anchored decode — recovery complete
                     }
                     guard let f = format, token.isLive else { continue }
-                    if !decoder.decode(au: au, format: f) {
+                    if decoder.decode(au: au, format: f) {
-                        // Submit/decoder error: drop the session and re-gate on the next IDR's
+                        decodeFailRun = 0
-                        // in-band parameter sets (a delta frame can't recover) — stage-1's policy
+                    } else {
-                        // — and ask the host for that IDR now (infinite GOP; throttled).
+                        // Submit/decoder error: drop the session and re-gate on the next IDR's in-band
+                        // parameter sets (a delta frame can't recover) and keep asking for that IDR.
                         decoder.reset()
-                        recovery.request()
+                        awaitingIDR = true
+                        decodeFailRun += 1
+                        // ~3 s of solid failure in a 4:4:4 session (and only there — a 4:2:0 loss
+                        // recovers within a GOP) ⇒ 4:4:4 isn't decodable here; end the session.
+                        if connection.isChroma444, decodeFailRun >= 180 {
+                            if token.isLive { onSessionEnd?() }
+                            break
+                        }
                     }
                 } catch {
                     if token.isLive { onSessionEnd?() }
@@ -154,27 +187,30 @@ public final class Stage2Pipeline {
         }
         thread.name = "punktfunk-stage2-pump"
         thread.qualityOfService = .userInteractive
+        pumpJoinable = true
         thread.start()
     }
 
-    /// MAIN thread, once per vsync. Present the newest ready frame (if any) and stamp
+    /// MAIN thread, once per vsync. Present the newest ready frame (if any) and stamp capture→present at
-    /// capture→present at `targetPresentNs` — the display link's target present instant, already
+    /// `targetPresentNs` — the display link's target present instant, already converted to
-    /// converted to `CLOCK_REALTIME` (see `realtimeNs(forDisplayLinkTimestamp:)`).
+    /// `CLOCK_REALTIME` (see `realtimeNs(forDisplayLinkTimestamp:)`).
     public func renderTick(targetPresentNs: Int64) {
         guard let frame = ring.take() else { return }
         guard presenter.render(frame.pixelBuffer, isHDR: frame.isHDR) else { return }
         presentMeter.record(ptsNs: frame.ptsNs, atNs: targetPresentNs, offsetNs: offsetNs)
     }
 
-    /// MAIN thread. Keep the drawable matched to the negotiated mode (host can Reconfigure).
+    /// Stop the pump (≤ one poll timeout) and drop the decode session. MAIN THREAD; idempotent. Does not
-    public func setDrawableSize(_ size: CGSize) {
+    /// close the connection. A restart needs a fresh Stage2Pipeline (cancel is permanent).
-        presenter.setDrawableSize(size)
-    }
-
-    /// Stop the pump (≤ one poll timeout) and drop the decode session. Does not close the
-    /// connection. A restart needs a fresh Stage2Pipeline (cancel is permanent).
     public func stop() {
         token.cancel()
+        // Join the pump (bounded: ≤ one nextAU poll + an in-flight decode) before resetting the decoder,
+        // so the pump can't rebuild a session right after the reset. Only the first stop joins; a
+        // repeat/deinit stop skips the already-drained semaphore.
+        if pumpJoinable {
+            pumpJoinable = false
+            _ = pumpStopped.wait(timeout: .now() + 0.5)
+        }
         decoder.reset()
         recovery.bind(nil) // stop requesting keyframes once the session is torn down
     }
@@ -182,8 +218,8 @@ public final class Stage2Pipeline {
     deinit { token.cancel() }
 
     /// Convert a `CADisplayLink.targetTimestamp` (CACurrentMediaTime basis) to a `CLOCK_REALTIME`
-    /// nanosecond instant — the present clock the AU pts + skew offset live in. Projects to the
+    /// nanosecond instant — the present clock the AU pts + skew offset live in. Projects to the target
-    /// target present time (when the frame is actually on glass), not the moment we drew.
+    /// present time (when the frame is actually on glass), not the moment we drew.
     public static func realtimeNs(forDisplayLinkTimestamp t: CFTimeInterval) -> Int64 {
         let caNow = CACurrentMediaTime()
         var ts = timespec()

clients/apple/Sources/PunktfunkKit/Stage444Probe.swift

@@ -0,0 +1,83 @@
+// Runtime 4:4:4 HEVC decode-capability probe.
+//
+// We advertise `VIDEO_CAP_444` (so the host upgrades to a full-chroma 4:4:4 stream) ONLY when this
+// device can decode 4:4:4 HEVC *in hardware* — software 4:4:4 decode works but is far too slow for a
+// real-time stream at the negotiated resolution, so a software-only device must keep 4:2:0.
+//
+// `VTIsHardwareDecodeSupported(HEVC)` and the HEVC-decoder-capabilities dictionary report HEVC HW
+// decode but expose nothing about `chroma_format_idc`, so the only reliable signal is to actually
+// create a *hardware-required* `VTDecompressionSession` for a tiny synthetic 4:4:4 keyframe and
+// confirm it both creates and decodes to the expected biplanar 4:4:4 pixel format. Validated on an
+// Apple M3 (HW 4:4:4 8- and 10-bit decode to `444v`/`x444`); a software-only decoder fails the
+// hardware-required create and we fall back to 4:2:0.
+//
+// The probe blobs are 256×256 (above the hardware decoder's minimum-dimension floor — a 16×16 clip
+// is rejected for ALL chroma formats, including 4:2:0) HEVC Range-Extensions keyframes generated
+// offline with libx265; see scripts notes. Results are cached (device-static) in lazy statics.
+
+import CoreMedia
+import CoreVideo
+import Foundation
+import VideoToolbox
+
+public enum Stage444Probe {
+    /// True iff this device hardware-decodes 8-bit 4:4:4 HEVC (the host's current 4:4:4 path —
+    /// BT.709 limited `yuv444p`). Cached after first evaluation.
+    public static let hwDecode444_8bit: Bool = probeHardware444(
+        au: Probe444Blobs.au444_8bit,
+        want: kCVPixelFormatType_444YpCbCr8BiPlanarVideoRange,
+        fullRangeSibling: kCVPixelFormatType_444YpCbCr8BiPlanarFullRange)
+
+    /// True iff this device hardware-decodes 10-bit 4:4:4 HEVC (the 4:4:4 ∩ HDR/10-bit intersection).
+    /// Cached after first evaluation.
+    public static let hwDecode444_10bit: Bool = probeHardware444(
+        au: Probe444Blobs.au444_10bit,
+        want: kCVPixelFormatType_444YpCbCr10BiPlanarVideoRange,
+        fullRangeSibling: kCVPixelFormatType_444YpCbCr10BiPlanarFullRange)
+
+    /// Create a hardware-REQUIRED `VTDecompressionSession` for the synthetic 4:4:4 keyframe and
+    /// decode it, returning true only when the decoder produces the expected (video- or full-range)
+    /// biplanar 4:4:4 pixel format. Any failure (no hardware path, wrong output format, decode error)
+    /// → false → we keep 4:2:0.
+    private static func probeHardware444(
+        au auBytes: [UInt8], want: OSType, fullRangeSibling: OSType
+    ) -> Bool {
+        let data = Data(auBytes)
+        guard let format = AnnexB.formatDescription(fromIDR: data) else { return false }
+        // Require a hardware decoder — a software false-positive would make us advertise 4:4:4 and
+        // then decode every real frame on the CPU, blowing the latency budget.
+        let spec: [CFString: Any] = [
+            kVTVideoDecoderSpecification_RequireHardwareAcceleratedVideoDecoder: true,
+        ]
+        let attrs: [CFString: Any] = [
+            kCVPixelBufferPixelFormatTypeKey: want,
+            kCVPixelBufferMetalCompatibilityKey: true,
+        ]
+        var session: VTDecompressionSession?
+        let created = VTDecompressionSessionCreate(
+            allocator: kCFAllocatorDefault, formatDescription: format,
+            decoderSpecification: spec as CFDictionary, imageBufferAttributes: attrs as CFDictionary,
+            outputCallback: nil, decompressionSessionOut: &session)
+        guard created == noErr, let session else { return false }
+        defer { VTDecompressionSessionInvalidate(session) }
+
+        let au = AccessUnit(data: data, ptsNs: 0, frameIndex: 0, flags: 0)
+        guard let sample = AnnexB.sampleBuffer(au: au, format: format) else { return false }
+
+        var produced: OSType = 0
+        let done = DispatchSemaphore(value: 0)
+        let status = VTDecompressionSessionDecodeFrame(
+            session, sampleBuffer: sample,
+            flags: [._EnableAsynchronousDecompression], infoFlagsOut: nil
+        ) { status, _, imageBuffer, _, _ in
+            if status == noErr, let imageBuffer {
+                produced = CVPixelBufferGetPixelFormatType(imageBuffer)
+            }
+            done.signal()
+        }
+        guard status == noErr else { return false }
+        VTDecompressionSessionWaitForAsynchronousFrames(session)
+        _ = done.wait(timeout: .now() + 1.0)
+        return produced == want || produced == fullRangeSibling
+    }
+}

clients/apple/Sources/PunktfunkKit/StreamPump.swift

@@ -6,6 +6,9 @@
 
 import AVFoundation
 import Foundation
+import os
+
+private let pumpLog = Logger(subsystem: "io.unom.punktfunk", category: "video")
 
 /// Cancellation handle owned by exactly one pump thread — a restart hands the old pump
 /// its own token, so it can never be revived by a newer start().
@@ -47,44 +50,74 @@ final class StreamPump {
             var format: CMVideoFormatDescription?
             var lastKeyframeRequest = Date.distantPast
             var lastFramesDropped = connection.framesDropped()
-            // Coalesced host keyframe request: the decode stays wedged for several frames until
+            // Recovery is a persistent WANT, not a one-shot edge: set it on detected loss (or a
-            // the IDR lands, so requesting on every frame would flood the control stream.
+            // decoder reset), retry the throttled request EVERY iteration, and clear it only when a
+            // fresh IDR actually re-anchors decode. The old code advanced `lastFramesDropped` on the
+            // same edge it fired the throttled request — so a request swallowed by the throttle (a
+            // second drop within the window, e.g. the lost recovery IDR itself being pruned) was
+            // never re-sent: the counter went flat, the climb never re-fired, and the picture stayed
+            // frozen for good while audio kept playing. The iPhone's lossy Wi-Fi hits this where the
+            // Mac's Ethernet never does.
+            var awaitingIDR = false
+            var awaitingSince = Date.distantPast // when the current recovery began (for the resume log)
+            var wasFailed = false
+            // Coalesced host keyframe request. 100 ms throttle (matches the working Android path):
+            // fast enough that a lost recovery IDR is re-requested promptly, bounded so a sustained
+            // freeze can't flood the control stream.
             func requestKeyframeThrottled() {
                 let now = Date()
-                if now.timeIntervalSince(lastKeyframeRequest) > 0.25 {
+                if now.timeIntervalSince(lastKeyframeRequest) > 0.1 {
                     connection.requestKeyframe()
                     lastKeyframeRequest = now
                 }
             }
             while token.isLive {
                 do {
-                    // Loss recovery (the primary recovery path). Under the host's infinite GOP the
+                    // Loss recovery (the primary path). Under the host's infinite GOP the only
-                    // only recovery keyframe is one we request. The reassembler drops unrecoverable
+                    // recovery keyframe is one we request. The reassembler drops unrecoverable AUs
-                    // AUs (framesDropped); the decoder then *conceals* the reference-missing delta
+                    // (framesDropped); the decoder then *conceals* the reference-missing deltas — a
-                    // frames that follow — a frozen / garbage picture, WITHOUT flipping the layer to
+                    // frozen / garbage picture that never flips the layer to .failed — so key off the
-                    // .failed — so the .failed check below rarely fires after a real network blip.
+                    // drop count climbing, then keep asking (awaitingIDR) until an IDR lands. Polled
-                    // Ask the host for a fresh IDR whenever the drop count climbs. Polled every
+                    // every iteration so a total-loss drought still recovers when packets resume.
-                    // iteration (not just per AU) so a total-loss drought still recovers the moment
-                    // packets resume and the reassembler counts the gap.
                     let dropped = connection.framesDropped()
                     if dropped > lastFramesDropped {
-                        lastFramesDropped = dropped
+                        // Log only on the false→true transition (once per recovery cycle), not per
-                        requestKeyframeThrottled()
+                        // dropped AU, so heavy loss doesn't spam the log.
+                        if !awaitingIDR {
+                            awaitingSince = Date()
+                            pumpLog.notice(
+                                "video: unrecoverable drop (framesDropped=\(dropped, privacy: .public)) — requesting recovery IDR")
                         }
+                        lastFramesDropped = dropped
+                        awaitingIDR = true
+                    }
+                    if awaitingIDR { requestKeyframeThrottled() }
+
                     guard let au = try connection.nextAU(timeoutMs: 100) else { continue }
                     onFrame?(au)
-                    if let f = AnnexB.formatDescription(fromIDR: au.data) {
+                    let idrFormat = AnnexB.formatDescription(fromIDR: au.data)
+                    if let f = idrFormat {
                         format = f          // refreshed on every IDR (mode changes included)
+                        if awaitingIDR {
+                            let ms = Int(Date().timeIntervalSince(awaitingSince) * 1000)
+                            pumpLog.notice("video: recovery IDR received — resumed after \(ms, privacy: .public) ms")
                         }
-                    if layer.status == .failed {
+                        awaitingIDR = false // a fresh IDR re-anchored decode — recovery complete
+                    }
+                    let failed = layer.status == .failed
+                    if failed {
                         // Decode wedged hard (the cold-first-connect case — a lost/corrupt opening
-                        // IDR): flush and re-gate on the next in-band parameter sets (resuming with
+                        // IDR): flush and, unless THIS AU is the recovering IDR (re-anchored above),
-                        // a delta frame can't recover), AND ask the host for a fresh IDR. Throttled:
+                        // re-gate on the next in-band parameter sets and keep asking — enqueuing a
-                        // the layer stays .failed across several polls until the IDR lands.
+                        // delta into a failed layer can't recover it.
+                        if !wasFailed { pumpLog.warning("video: display layer .failed — flushing + re-anchoring") }
                         layer.flush()
-                        format = AnnexB.formatDescription(fromIDR: au.data)
+                        if idrFormat == nil {
-                        requestKeyframeThrottled()
+                            format = nil
+                            awaitingIDR = true
                         }
+                    }
+                    wasFailed = failed
                     guard let f = format,
                           let sample = AnnexB.sampleBuffer(au: au, format: f),
                           token.isLive // don't enqueue a stale frame after a restart

clients/apple/Sources/PunktfunkKit/StreamView.swift

@@ -137,8 +137,8 @@ public struct StreamView: NSViewRepresentable {
 public final class StreamLayerView: NSView {
     private let displayLayer = AVSampleBufferDisplayLayer()
     private var pump: StreamPump?
-    /// Stage-2 presenter (opt-in via `punktfunk.presenter`): a CAMetalLayer sublayer driven by a
+    /// Stage-2 presenter (default): a CAMetalLayer sublayer driven by a display link instead of the
-    /// display link instead of the StreamPump → displayLayer path. nil = stage-1 (default).
+    /// StreamPump → displayLayer path. nil = stage-1 (Metal-unavailable fallback / DEBUG toggle).
     var presentMeter: LatencyMeter?
     private var stage2: Stage2Pipeline?
     private var stage2Link: CADisplayLink?
@@ -245,6 +245,15 @@ public final class StreamLayerView: NSView {
         layoutMetalLayer() // keep the stage-2 sublayer aspect-fit to the view
     }
 
+    public override func setFrameSize(_ newSize: NSSize) {
+        super.setFrameSize(newSize)
+        // `layout()` isn't guaranteed on a manual-frame (no-Auto-Layout) live resize, so the
+        // stage-2 metal sublayer's drawableSize could stay at the old size while the view grows —
+        // the compositor then upscales a too-small drawable and the video turns blocky. Resize the
+        // drawable here too so it always tracks the window's pixel size (no stale upscale).
+        layoutMetalLayer()
+    }
+
     // MARK: - Capture state machine
 
     /// Clicking into the video engages capture; that click is local (engagement), so
@@ -549,10 +558,17 @@ public final class StreamLayerView: NSView {
         cursorVisible = false
         _ = connection.resolvedCompositor // (was: Auto → gamescope; kept to document intent)
 
-        // Presenter choice — default stage-1 (the known-good AVSampleBufferDisplayLayer). Stage-2
+        // Presenter choice — stage-2 is the DEFAULT (explicit VTDecompressionSession decode + a
-        // (`punktfunk.presenter == "stage2"`) takes explicit VTDecompressionSession decode + a
+        // CAMetalLayer/display-link present): it can detect + recover a wedged decoder where
-        // CAMetalLayer/display-link present; it falls back here if Metal can't be set up.
+        // stage-1's AVSampleBufferDisplayLayer freezes hard on a lost HEVC reference. Stage-1 is
-        if UserDefaults.standard.string(forKey: DefaultsKey.presenter) == "stage2",
+        // reachable only via the DEBUG presenter toggle; release always takes stage-2 (the stage-1
+        // pump below stays the automatic fallback if Metal is missing).
+        #if DEBUG
+        let forceStage1 = UserDefaults.standard.string(forKey: DefaultsKey.presenter) == "stage1"
+        #else
+        let forceStage1 = false
+        #endif
+        if !forceStage1,
            let meter = presentMeter,
            let pipeline = Stage2Pipeline(presentMeter: meter) {
             startStage2(pipeline, connection: connection, onFrame: onFrame, onSessionEnd: onSessionEnd)
@@ -593,9 +609,11 @@ public final class StreamLayerView: NSView {
             targetPresentNs: Stage2Pipeline.realtimeNs(forDisplayLinkTimestamp: link.targetTimestamp))
     }
 
-    /// Aspect-fit the metal sublayer in the view (the host streams at the client's native mode,
+    /// Position the metal sublayer aspect-fit in the view (the host streams at the client's native
-    /// so this is usually the full bounds; it letterboxes a resized window). drawableSize is the
+    /// mode, so this is usually the full bounds; it letterboxes a resized window). Only the layer
-    /// layer's pixel size — the fullscreen-triangle shader scales the decoded texture to fill it.
+    /// FRAME is set here — the presenter sizes the drawable to the decoded frame and the layer's
+    /// contentsGravity (.resizeAspect) scales it to this frame via the system compositor, so a
+    /// resized window rescales through the system's filter (matching stage-1) instead of the shader.
     private func layoutMetalLayer() {
         guard let metalLayer, let connection else { return }
         let mode = connection.currentMode()
@@ -604,14 +622,12 @@ public final class StreamLayerView: NSView {
                 aspectRatio: CGSize(width: Int(mode.width), height: Int(mode.height)),
                 insideRect: bounds)
             : bounds
-        let scale = window?.backingScaleFactor ?? 1
         // No implicit resize animation; refresh contentsScale on a retina↔non-retina move.
         CATransaction.begin()
         CATransaction.setDisableActions(true)
-        metalLayer.contentsScale = scale
+        metalLayer.contentsScale = window?.backingScaleFactor ?? 1
         metalLayer.frame = fit
         CATransaction.commit()
-        stage2?.setDrawableSize(CGSize(width: fit.width * scale, height: fit.height * scale))
     }
 
     public override func viewDidChangeBackingProperties() {
@@ -622,7 +638,7 @@ public final class StreamLayerView: NSView {
     private func teardownStage2() {
         stage2Link?.invalidate()
         stage2Link = nil
-        stage2?.stop()
+        stage2?.stop() // stops the pump (synchronous join) + drops the decode session
         stage2 = nil
         metalLayer?.removeFromSuperlayer()
         metalLayer = nil

clients/apple/Sources/PunktfunkKit/StreamViewIOS.swift

@@ -11,13 +11,18 @@
 // host mode, so the host's rescale is the identity).
 //
 // A hardware mouse/trackpad is a pointer, not a finger. When the scene is pointer-LOCKED
-// (full-screen + frontmost iPad) GCMouse delivers raw relative deltas and the system hides
+// (full-screen + frontmost iPad, and the user hasn't disabled pointer capture in Settings —
-// the cursor — the gaming-grade path. When it CAN'T lock (Stage Manager, not frontmost,
+// see PointerLockChain, which steers the lock request through SwiftUI's hosting controllers)
-// iPhone) the system shows its own cursor and routes the mouse through UIKit's pointer path:
+// GCMouse delivers raw relative deltas and the system hides the cursor — the gaming-grade path.
-// hover + indirect-pointer touches, which we forward as ABSOLUTE cursor position (+ buttons)
+// InputCapture handles EVERY connected mouse (GCMouse.mice), not just the current one, so a
-// so the host cursor tracks the visible local one. We never forward an indirect pointer as a
+// trackpad + a second pointer (e.g. a Universal Control mouse) both drive. When the scene CAN'T
-// touch — doing so hid the cursor and made the host see taps instead of a moving mouse.
+// lock (Stage Manager, not frontmost, iPhone, capture disabled) the system shows its own cursor
-// GCMouse is gated off whenever the lock isn't held so the two paths can't double-send.
+// and routes the mouse through UIKit's pointer path: hover + indirect-pointer touches, which we
+// forward as ABSOLUTE cursor position (+ buttons) so the host cursor tracks the visible local one.
+// We never forward an indirect pointer as a touch — doing so hid the cursor and made the host see
+// taps instead of a moving mouse. The two paths are mutually exclusive on `gcMouseForwarding`
+// (== locked): GCMouse forwards only WHILE locked, the UIKit indirect path (motion, buttons AND
+// scroll) only while NOT locked — so a pointer that emits both channels under lock can't double-send.
 // Hardware keyboard forwarding shares InputCapture with macOS — auto-engaged when streaming
 // starts, ⌘⎋ toggles (detected from the HID stream; there is no NSEvent monitor here).
 //
@@ -92,8 +97,8 @@ public final class StreamViewController: UIViewController {
     public private(set) var connection: PunktfunkConnection?
     private var pump: StreamPump?
     private var observers: [NSObjectProtocol] = []
-    /// Stage-2 presenter (opt-in via `punktfunk.presenter`): a CAMetalLayer sublayer driven by a
+    /// Stage-2 presenter (default): a CAMetalLayer sublayer driven by a CADisplayLink instead of the
-    /// CADisplayLink instead of the StreamPump → displayLayer path. nil = stage-1 (default).
+    /// StreamPump → displayLayer path. nil = stage-1 (Metal-unavailable fallback / DEBUG toggle).
     var presentMeter: LatencyMeter?
     private var stage2: Stage2Pipeline?
     private var stage2Link: CADisplayLink?
@@ -136,6 +141,13 @@ public final class StreamViewController: UIViewController {
 
     public override func loadView() {
         view = StreamLayerUIView()
+        // Re-size the stage-2 drawable if the display scale changes without a bounds change (e.g.
+        // moving to an external display at a different scale) — the iOS analogue of macOS's
+        // viewDidChangeBackingProperties relayout. The handler takes the VC as its argument, so it
+        // doesn't capture self (no retain cycle with the registration).
+        registerForTraitChanges([UITraitDisplayScale.self]) { (vc: StreamViewController, _) in
+            vc.layoutMetalLayer()
+        }
         #if os(iOS)
         // Hide the iPadOS cursor while it hovers the video: the host renders its own
         // cursor from our deltas, so the local one only diverges from it. This hides the
@@ -148,19 +160,58 @@ public final class StreamViewController: UIViewController {
     }
 
     #if os(iOS)
-    // Pointer lock is only meaningful on iPad (iPhone has no hardware-pointer lock) and
+    /// Whether the user wants the mouse/trackpad pointer CAPTURED (pointer lock → relative
-    // only when capture is engaged. The system additionally requires full-screen + frontmost
+    /// movement, the gaming default) rather than forwarded as an absolute position (desktop
-    // and may drop it (Slide Over/Stage Manager/backgrounding) — verified in setCaptured().
+    /// use). Read live from UserDefaults so it tracks the Settings toggle; defaults to on when
-    public override var prefersPointerLocked: Bool {
+    /// unset. iPad-only — gated again in `prefersPointerLocked`.
-        captured && UIDevice.current.userInterfaceIdiom == .pad
+    private var pointerCaptureEnabled: Bool {
+        UserDefaults.standard.object(forKey: DefaultsKey.pointerCapture) as? Bool ?? true
     }
+
+    /// Whether the pointer should be CAPTURED right now: iPad, capture engaged, and the user
+    /// hasn't opted into the absolute (desktop) pointer. The system additionally requires
+    /// full-screen + frontmost and may drop the lock (Slide Over/Stage Manager/backgrounding) —
+    /// syncPointerLock() handles the actual grant/drop and falls back to absolute when unlocked.
+    private var wantsPointerLock: Bool {
+        captured && pointerCaptureEnabled && UIDevice.current.userInterfaceIdiom == .pad
+    }
+
+    public override var prefersPointerLocked: Bool { wantsPointerLock }
     public override var prefersHomeIndicatorAutoHidden: Bool { true }
 
-    // If SwiftUI's UIHostingController reparents us, a plain container parent that forwards
+    // NOTE: we deliberately do NOT override `childViewControllerForPointerLock`. The default
-    // its pointer-lock decision to its children will then reach this VC. (UIHostingController
+    // returns nil, which tells the system to use THIS controller's own `prefersPointerLocked` —
-    // itself does not consult children, which is why GCMouse deltas can never arrive there —
+    // exactly what we want, since `PointerLockChain` forces our SwiftUI ancestors to forward the
-    // the touch path, always forwarded, is the unconditional fallback.)
+    // downward walk to us and we are the terminal anchor. Returning `self` here would make the
-    public override var childViewControllerForPointerLock: UIViewController? { self }
+    // system ask the same controller forever (it keeps delegating to the returned child) →
+    // unbounded recursion → stack overflow once the chain actually reaches us.
+
+    /// (Re)build or tear down the forced pointer-lock forwarding chain from this controller to the
+    /// window root so the system actually resolves our `prefersPointerLocked`. Safe to call
+    /// repeatedly — it no-ops until the view is in a window with a parent chain, and re-runs from
+    /// the appearance/parent callbacks once SwiftUI has placed us.
+    private func updatePointerLockChain() {
+        // Engaging needs a live parent chain to the window root; disengaging is always safe and
+        // must run even after the view has left the window (session teardown) so the stamped
+        // SwiftUI ancestors are cleared.
+        if wantsPointerLock, view.window != nil {
+            PointerLockChain.engage(self)
+        } else {
+            PointerLockChain.disengage(self)
+        }
+    }
+
+    public override func viewDidAppear(_ animated: Bool) {
+        super.viewDidAppear(animated)
+        // SwiftUI places us in the hierarchy AFTER start()'s setCaptured(true), and may reparent us
+        // later — re-anchor the chain here so a lock requested before we had a parent still lands.
+        updatePointerLockChain()
+    }
+
+    public override func didMove(toParent parent: UIViewController?) {
+        super.didMove(toParent: parent)
+        updatePointerLockChain() // chain shape changed — re-anchor (or no-op if not yet in a window)
+    }
     #endif
 
     func start(
@@ -190,18 +241,22 @@ public final class StreamViewController: UIViewController {
             guard self?.captureEnabled == true else { return }
             connection?.send(event)
         }
-        // Indirect pointer (mouse/trackpad with no lock) → absolute cursor + buttons, routed
+        // Indirect pointer (mouse/trackpad) WITHOUT a lock → absolute cursor + buttons + scroll.
-        // through InputCapture so the forwarding gate and release-on-blur apply uniformly.
+        // While the scene is pointer-LOCKED the GCMouse path owns motion AND buttons AND scroll, so
+        // the whole UIKit indirect path is gated off here (`gcMouseForwarding`). The trackpad and a
+        // mouse BOTH report through GCMouse under lock and ALSO emit UIKit indirect-pointer events
+        // (pinned at the locked position) — without this gate a click double-sends (GCMouse + UIKit)
+        // and a second pointer (e.g. a Universal Control mouse) competes with the trackpad. The gate
+        // is the exact mirror of the GCMouse handlers, which fire only while locked.
         streamView.onPointerMoveAbs = { [weak self] p in
-            self?.inputCapture?.sendMouseAbs(
+            guard let self, self.inputCapture?.gcMouseForwarding == false else { return }
+            self.inputCapture?.sendMouseAbs(
                 x: p.x, y: p.y, surfaceWidth: p.w, surfaceHeight: p.h)
         }
         streamView.onPointerButton = { [weak self] button, down in
-            self?.inputCapture?.sendMouseButton(button, pressed: down)
+            guard let self, self.inputCapture?.gcMouseForwarding == false else { return }
+            self.inputCapture?.sendMouseButton(button, pressed: down)
         }
-        // Trackpad two-finger / wheel scroll → host scroll. The pan recognizer is the
-        // UNLOCKED regime; while locked, GCMouse's scroll handler owns it — mirror the
-        // sendMouseAbs !gcMouseForwarding gate so the two can't double-send.
         streamView.onScroll = { [weak self] dx, dy in
             guard let self, self.inputCapture?.gcMouseForwarding == false else { return }
             self.inputCapture?.sendScroll(dx: dx, dy: dy)
@@ -219,10 +274,17 @@ public final class StreamViewController: UIViewController {
         inputCapture = capture
         #endif
 
-        // Presenter choice — default stage-1 (the known-good AVSampleBufferDisplayLayer). Stage-2
+        // Presenter choice — stage-2 is the DEFAULT (VTDecompressionSession decode + a
-        // (`punktfunk.presenter == "stage2"`) takes VTDecompressionSession decode + a
+        // CAMetalLayer/display-link present): it can detect + recover a wedged decoder, where
-        // CAMetalLayer/display-link present; falls back here if Metal can't be set up.
+        // stage-1's AVSampleBufferDisplayLayer freezes hard on a lost HEVC reference frame with no
-        if UserDefaults.standard.string(forKey: DefaultsKey.presenter) == "stage2",
+        // way to recover. Stage-1 is reachable only via the DEBUG presenter toggle; release always
+        // takes stage-2 (the stage-1 pump below stays the automatic fallback if Metal is missing).
+        #if DEBUG
+        let forceStage1 = UserDefaults.standard.string(forKey: DefaultsKey.presenter) == "stage1"
+        #else
+        let forceStage1 = false
+        #endif
+        if !forceStage1,
            let meter = presentMeter,
            let pipeline = Stage2Pipeline(presentMeter: meter) {
             startStage2(pipeline, connection: connection, onFrame: onFrame, onSessionEnd: onSessionEnd)
@@ -300,8 +362,8 @@ public final class StreamViewController: UIViewController {
         onFrame: (@Sendable (AccessUnit) -> Void)?, onSessionEnd: (@Sendable () -> Void)?
     ) {
         let metal = pipeline.layer
-        metal.contentsScale = streamView.contentScaleFactor
         // Composites OVER the idle (un-enqueued in stage-2) AVSampleBufferDisplayLayer base.
+        // (contentsScale + frame are set by layoutMetalLayer() just below.)
         streamView.layer.addSublayer(metal)
         metalLayer = metal
         stage2 = pipeline
@@ -325,9 +387,20 @@ public final class StreamViewController: UIViewController {
         layoutMetalLayer()
     }
 
-    /// Aspect-fit the metal sublayer in the view (the host streams at the client's native mode,
+    /// The display scale to render the metal drawable at. `traitCollection.displayScale` is the
-    /// so this is usually the full bounds). drawableSize is the layer's pixel size; the shader's
+    /// canonical render scale and is reliable once the controller is in the hierarchy;
-    /// fullscreen triangle scales the decoded texture to fill it.
+    /// `view.contentScaleFactor` can read 1.0 before the view attaches to a window/screen, which
+    /// would size the drawable at point resolution → a pixelated, upscaled mess. Falls back to the
+    /// main screen scale if the trait is still unspecified.
+    private var renderScale: CGFloat {
+        let s = traitCollection.displayScale
+        return s > 0 ? s : UIScreen.main.scale
+    }
+
+    /// Position the metal sublayer aspect-fit in the view (the host streams at the client's native
+    /// mode, so this is usually the full bounds). Only the layer FRAME is set here — the presenter
+    /// sizes the drawable to the decoded frame and the layer's contentsGravity (.resizeAspect)
+    /// scales it to this frame via the system compositor (matching stage-1's videoGravity).
     private func layoutMetalLayer() {
         guard let metalLayer, let connection else { return }
         let mode = connection.currentMode()
@@ -337,19 +410,17 @@ public final class StreamViewController: UIViewController {
                 aspectRatio: CGSize(width: Int(mode.width), height: Int(mode.height)),
                 insideRect: bounds)
             : bounds
-        let scale = streamView.contentScaleFactor
         CATransaction.begin()
         CATransaction.setDisableActions(true) // don't animate the resize
-        metalLayer.contentsScale = scale
+        metalLayer.contentsScale = renderScale
         metalLayer.frame = fit
         CATransaction.commit()
-        stage2?.setDrawableSize(CGSize(width: fit.width * scale, height: fit.height * scale))
     }
 
     private func teardownStage2() {
         stage2Link?.invalidate()
         stage2Link = nil
-        stage2?.stop()
+        stage2?.stop() // stops the pump (synchronous join) + drops the decode session
         stage2 = nil
         metalLayer?.removeFromSuperlayer()
         metalLayer = nil
@@ -369,6 +440,7 @@ public final class StreamViewController: UIViewController {
             captured = false
         }
         setNeedsUpdateOfPrefersPointerLocked()
+        updatePointerLockChain() // (re)anchor the SwiftUI ancestors so the lock actually resolves
         syncPointerLock() // resolve cursor + GCMouse/absolute routing for the current state
         let onCaptureChange = onCaptureChange
         let captured = captured

clients/apple/Sources/PunktfunkKit/VideoDecoder.swift

@@ -49,11 +49,10 @@ public final class VideoDecoder: @unchecked Sendable {
     /// pump can re-gate on the next IDR.
     private let onDecodeError: @Sendable (OSStatus) -> Void
 
-    /// Latest source HDR mastering metadata (from `PunktfunkConnection.nextHdrMeta`), attached to
+    /// Whether the negotiated stream is full-chroma 4:4:4 (`connection.isChroma444`), set once at
-    /// each decoded HDR pixel buffer so the compositor tone-maps from the real grade. Guarded by its
+    /// session start before any decode. Selects the 4:4:4 decode pixel format (orthogonal to bit
-    /// own lock — written by the pump thread, read on the VT decode callback.
+    /// depth / HDR). Read inside `createSessionLocked` under `lock`.
-    private let metaLock = NSLock()
+    private var chroma444 = false
-    private var hdrMeta: PunktfunkConnection.HdrMeta?
 
     public init(
         onDecoded: @escaping @Sendable (ReadyFrame) -> Void,
@@ -65,12 +64,13 @@ public final class VideoDecoder: @unchecked Sendable {
 
     deinit { teardown() }
 
-    /// Set the source HDR mastering metadata (drained from `PunktfunkConnection.nextHdrMeta`). It's
+    /// Select the chroma subsampling of the decode output (4:2:0 vs full-chroma 4:4:4). Call once at
-    /// attached to subsequent decoded HDR pixel buffers. Thread-safe; cheap to call on each update.
+    /// session start, before decoding, from `connection.isChroma444`. Takes effect on the next
-    public func setHdrMeta(_ meta: PunktfunkConnection.HdrMeta) {
+    /// session (re)build. Thread-safe.
-        metaLock.lock()
+    public func setChroma444(_ on: Bool) {
-        hdrMeta = meta
+        lock.lock()
-        metaLock.unlock()
+        chroma444 = on
+        lock.unlock()
     }
 
     /// Submit one AU for asynchronous decode, (re)creating the session if `format` changed. The
@@ -135,8 +135,10 @@ public final class VideoDecoder: @unchecked Sendable {
 
     /// True when `newFormat` carries a PQ (SMPTE ST 2084) or HLG transfer function — i.e. the host
     /// is sending HDR (BT.2020). VideoToolbox populates the transfer-function extension from the
-    /// HEVC VUI, so this tracks the *stream*, switching dynamically when the user toggles HDR
+    /// HEVC VUI, so this picks the decode bit depth (10-bit P010/x444 vs 8-bit NV12/444v) from the
-    /// (the host re-emits parameter sets with the new VUI → a new format desc → session rebuild).
+    /// stream. The present-side HDR config (colorspace/EDR/shader) is latched once per session from
+    /// the Welcome (`connection.isHDR`), which the host does NOT flip mid-session — so this predicate
+    /// and that config agree for the session (a `#if DEBUG` assert in the presenter guards it).
     static func isHDRFormat(_ format: CMVideoFormatDescription) -> Bool {
         guard
             let tf = CMFormatDescriptionGetExtension(
@@ -157,11 +159,18 @@ public final class VideoDecoder: @unchecked Sendable {
         session = nil
         format = nil
 
+        // Decode pixel format is a 2×2 of (chroma, depth/HDR), both biplanar so the presenter binds
+        // plane 0 = luma, plane 1 = interleaved chroma uniformly — 4:4:4 just delivers a full-size
+        // chroma plane. 10-bit (P010 / `x444`) for HDR (PQ/HLG), 8-bit (NV12 / `444v`) otherwise.
         let hdr = Self.isHDRFormat(newFormat)
-        let pixelFormat =
+        let pixelFormat: OSType = {
-            hdr
+            switch (chroma444, hdr) {
-            ? kCVPixelFormatType_420YpCbCr10BiPlanarVideoRange // P010 (10-bit)
+            case (false, false): return kCVPixelFormatType_420YpCbCr8BiPlanarVideoRange // NV12
-            : kCVPixelFormatType_420YpCbCr8BiPlanarVideoRange // NV12 (8-bit)
+            case (false, true): return kCVPixelFormatType_420YpCbCr10BiPlanarVideoRange // P010
+            case (true, false): return kCVPixelFormatType_444YpCbCr8BiPlanarVideoRange // 444v
+            case (true, true): return kCVPixelFormatType_444YpCbCr10BiPlanarVideoRange // x444
+            }
+        }()
         let imageAttrs: [CFString: Any] = [
             kCVPixelBufferMetalCompatibilityKey: true,
             kCVPixelBufferPixelFormatTypeKey: pixelFormat,
@@ -169,11 +178,20 @@ public final class VideoDecoder: @unchecked Sendable {
         var callback = VTDecompressionOutputCallbackRecord(
             decompressionOutputCallback: decoderOutputCallback,
             decompressionOutputRefCon: Unmanaged.passUnretained(self).toOpaque())
+        // 4:4:4 sessions REQUIRE a hardware decoder: we only advertise 4:4:4 when the hardware probe
+        // passed, so a hardware-incapable mode (e.g. a resolution past the HW 4:4:4 ceiling) must fail
+        // HERE, synchronously, letting the pump's backstop end the session — rather than silently
+        // falling back to a software 4:4:4 decoder far too slow for a real-time stream. 4:2:0 keeps the
+        // software fallback (nil spec) as a robustness net.
+        let spec: CFDictionary? =
+            chroma444
+            ? [kVTVideoDecoderSpecification_RequireHardwareAcceleratedVideoDecoder: true] as CFDictionary
+            : nil
         var newSession: VTDecompressionSession?
         let status = VTDecompressionSessionCreate(
             allocator: kCFAllocatorDefault,
             formatDescription: newFormat,
-            decoderSpecification: nil, // hardware by default
+            decoderSpecification: spec,
             imageBufferAttributes: imageAttrs as CFDictionary,
             outputCallback: &callback,
             decompressionSessionOut: &newSession)
@@ -195,26 +213,17 @@ public final class VideoDecoder: @unchecked Sendable {
         // pts was stamped at timescale 1e9 (AnnexB.sampleBuffer); normalize defensively.
         let p = CMTimeConvertScale(pts, timescale: 1_000_000_000, method: .default)
         let ptsNs = p.value > 0 ? UInt64(p.value) : 0
-        // HDR iff the decoder produced a 10-bit P010 buffer (we only request P010 for PQ streams).
+        // HDR iff the decoder produced a 10-bit buffer (we only request a 10-bit format for PQ/HLG
+        // streams). Covers 4:2:0 (P010) and 4:4:4 (`x444`), video- and full-range, so a 10-bit 4:4:4
+        // HDR frame isn't misclassified as SDR. (The mastering metadata is applied to the presenter's
+        // CAMetalLayer via CAEDRMetadata, not to this source buffer — a separate-drawable presenter
+        // never composites the source buffer's attachments, so attaching them here would be dead.)
+        let fmt = CVPixelBufferGetPixelFormatType(imageBuffer)
         let isHDR =
-            CVPixelBufferGetPixelFormatType(imageBuffer)
+            fmt == kCVPixelFormatType_420YpCbCr10BiPlanarVideoRange
-            == kCVPixelFormatType_420YpCbCr10BiPlanarVideoRange
+            || fmt == kCVPixelFormatType_420YpCbCr10BiPlanarFullRange
-        // Attach the source's mastering display + content light level (ST.2086 / CEA-861.3) so the
+            || fmt == kCVPixelFormatType_444YpCbCr10BiPlanarVideoRange
-        // compositor tone-maps from the real grade rather than inferring from the PQ colourspace
+            || fmt == kCVPixelFormatType_444YpCbCr10BiPlanarFullRange
-        // alone. The SEI byte payloads map 1:1 to these CVImageBuffer attachment keys.
-        if isHDR {
-            metaLock.lock()
-            let meta = hdrMeta
-            metaLock.unlock()
-            if let meta {
-                CVBufferSetAttachment(
-                    imageBuffer, kCVImageBufferMasteringDisplayColorVolumeKey,
-                    meta.masteringDisplayColorVolume() as CFData, .shouldPropagate)
-                CVBufferSetAttachment(
-                    imageBuffer, kCVImageBufferContentLightLevelInfoKey,
-                    meta.contentLightLevelInfo() as CFData, .shouldPropagate)
-            }
-        }
         onDecoded(
             ReadyFrame(ptsNs: ptsNs, decodedNs: decodedNs, pixelBuffer: imageBuffer, isHDR: isHDR))
     }

clients/apple/Tests/PunktfunkKitTests/MetalPresenterTests.swift

@@ -0,0 +1,69 @@
+import XCTest
+
+#if canImport(Metal)
+import CoreVideo
+import Metal
+import QuartzCore
+@testable import PunktfunkKit
+
+final class MetalPresenterTests: XCTestCase {
+    /// `MetalVideoPresenter.make()` compiles the runtime Metal shaders (the BT.709/BT.2020 YUV→RGB
+    /// fragment shaders plus the Catmull-Rom luma sampler). A `nil` result on a GPU-equipped host
+    /// means a shader failed to compile — this catches a malformed shader before it silently
+    /// degrades stage-2 to a stage-1 fallback on device.
+    func testPresenterInitCompilesShaders() throws {
+        guard MTLCreateSystemDefaultDevice() != nil else {
+            throw XCTSkip("no Metal device available in this environment")
+        }
+        XCTAssertNotNil(
+            MetalVideoPresenter.make(),
+            "stage-2 Metal shaders failed to compile (presenter init returned nil)")
+    }
+
+    /// The HDR fix: `configure(hdr:)` must put the layer into the BT.2020-PQ EDR configuration with a
+    /// reference-white anchor (`edrMetadata`) — the missing anchor was what made HDR render "too
+    /// bright". SDR must use the plain 8-bit path with EDR off and no metadata. A mid-session flip is a
+    /// per-mode reconfigure, so the round trip back to SDR must fully restore the SDR config.
+    func testConfigureHDRSetsEDRAnchor() throws {
+        guard let presenter = MetalVideoPresenter.make() else {
+            throw XCTSkip("no Metal device available in this environment")
+        }
+        presenter.configure(hdr: true)
+        XCTAssertEqual(presenter.layer.pixelFormat, .rgba16Float, "HDR uses an EDR-capable drawable")
+        XCTAssertNotNil(presenter.layer.colorspace, "HDR layer must be tagged (itur_2100_PQ)")
+        XCTAssertTrue(
+            presenter.layer.wantsExtendedDynamicRangeContent, "EDR must be requested on all platforms")
+        XCTAssertNotNil(
+            presenter.layer.edrMetadata,
+            "HDR must anchor reference white via edrMetadata (the fix for 'too bright')")
+
+        // Mid-session HDR→SDR flip: the 8-bit path, EDR off, no metadata.
+        presenter.configure(hdr: false)
+        XCTAssertEqual(presenter.layer.pixelFormat, .bgra8Unorm, "SDR uses the plain 8-bit drawable")
+        XCTAssertFalse(presenter.layer.wantsExtendedDynamicRangeContent)
+        XCTAssertNil(presenter.layer.edrMetadata)
+    }
+
+    /// `render` with a freshly-allocated NV12 buffer must present without crashing or hanging — the
+    /// main-thread present path is the highest-risk part of the stage-2 rewrite. (A headless CI with no
+    /// display can still allocate a drawable from a CAMetalLayer; if it can't, render returns false,
+    /// which is also a valid non-crashing outcome.)
+    func testRenderDoesNotCrashOnNV12Frame() throws {
+        guard let presenter = MetalVideoPresenter.make() else {
+            throw XCTSkip("no Metal device available in this environment")
+        }
+        presenter.configure(hdr: false)
+        var pb: CVPixelBuffer?
+        let attrs: [CFString: Any] = [kCVPixelBufferMetalCompatibilityKey: true]
+        let status = CVPixelBufferCreate(
+            kCFAllocatorDefault, 256, 256, kCVPixelFormatType_420YpCbCr8BiPlanarVideoRange,
+            attrs as CFDictionary, &pb)
+        guard status == kCVReturnSuccess, let pixelBuffer = pb else {
+            throw XCTSkip("could not allocate a test pixel buffer")
+        }
+        // Just asserting it returns (true or false) without trapping — the layer may have no drawable
+        // source headless, so a false return is acceptable.
+        _ = presenter.render(pixelBuffer, isHDR: false)
+    }
+}
+#endif

clients/apple/Tests/PunktfunkKitTests/Stage444Tests.swift

@@ -0,0 +1,68 @@
+// 4:4:4 decode-path coverage: the hardware-capability probe is stable/cached, and a real 4:4:4 HEVC
+// keyframe decodes through VideoDecoder to a biplanar 4:4:4 pixel buffer. Reuses the same synthetic
+// 4:4:4 blobs the runtime probe ships with.
+
+import CoreVideo
+import VideoToolbox
+import XCTest
+@testable import PunktfunkKit
+
+private final class FrameBox: @unchecked Sendable {
+    let lock = NSLock()
+    var frame: ReadyFrame?
+    var error: OSStatus?
+}
+
+final class Stage444Tests: XCTestCase {
+    /// The capability probe is device-static and cached — reading it twice must return the same value
+    /// (and must never crash, including where 4:4:4 is unsupported → false).
+    func testProbeIsStableAndCached() {
+        XCTAssertEqual(Stage444Probe.hwDecode444_8bit, Stage444Probe.hwDecode444_8bit)
+        XCTAssertEqual(Stage444Probe.hwDecode444_10bit, Stage444Probe.hwDecode444_10bit)
+    }
+
+    /// A real 8-bit 4:4:4 HEVC keyframe (the embedded probe blob) decodes through `VideoDecoder` with
+    /// `setChroma444(true)` to a 256×256 biplanar 4:4:4 (`444v`/`444f`) buffer classified SDR.
+    /// (4:4:4 sessions require a hardware decoder — skip where there isn't one, which is exactly where
+    /// the client wouldn't advertise 4:4:4 anyway.)
+    func testVideoDecoderDecodes444() throws {
+        try XCTSkipUnless(
+            Stage444Probe.hwDecode444_8bit, "no hardware 4:4:4 decode on this device")
+        let data = Data(Probe444Blobs.au444_8bit)
+        let format = try XCTUnwrap(
+            AnnexB.formatDescription(fromIDR: data), "the 4:4:4 blob must yield a format description")
+        let au = AccessUnit(data: data, ptsNs: 7_000_000, frameIndex: 0, flags: 0)
+
+        let box = FrameBox()
+        let done = DispatchSemaphore(value: 0)
+        let decoder = VideoDecoder(
+            onDecoded: { f in box.lock.lock(); box.frame = f; box.lock.unlock(); done.signal() },
+            onDecodeError: { s in box.lock.lock(); box.error = s; box.lock.unlock(); done.signal() })
+        decoder.setChroma444(true)
+
+        XCTAssertTrue(decoder.decode(au: au, format: format), "4:4:4 frame submit should succeed")
+        XCTAssertEqual(done.wait(timeout: .now() + 10), .success, "the decode callback must fire")
+        decoder.reset()
+
+        box.lock.lock(); let frame = box.frame; let error = box.error; box.lock.unlock()
+        XCTAssertNil(error.map { "decode error \($0)" })
+        let ready = try XCTUnwrap(frame, "a 4:4:4 ReadyFrame must be delivered")
+        XCTAssertEqual(CVPixelBufferGetWidth(ready.pixelBuffer), 256)
+        XCTAssertEqual(CVPixelBufferGetHeight(ready.pixelBuffer), 256)
+        let pf = CVPixelBufferGetPixelFormatType(ready.pixelBuffer)
+        XCTAssertTrue(
+            pf == kCVPixelFormatType_444YpCbCr8BiPlanarVideoRange
+                || pf == kCVPixelFormatType_444YpCbCr8BiPlanarFullRange,
+            "expected a biplanar 4:4:4 8-bit buffer, got \(fourCCString(pf))")
+        XCTAssertFalse(ready.isHDR, "an 8-bit BT.709 4:4:4 stream is SDR")
+        // The chroma plane (plane 1) must be FULL resolution for 4:4:4 (vs half for 4:2:0) — this is
+        // what lets the unchanged shader sample chroma at the luma UV.
+        XCTAssertEqual(CVPixelBufferGetWidthOfPlane(ready.pixelBuffer, 1), 256)
+        XCTAssertEqual(CVPixelBufferGetHeightOfPlane(ready.pixelBuffer, 1), 256)
+    }
+
+    private func fourCCString(_ t: OSType) -> String {
+        let b = [UInt8(t >> 24 & 0xff), UInt8(t >> 16 & 0xff), UInt8(t >> 8 & 0xff), UInt8(t & 0xff)]
+        return String(bytes: b, encoding: .ascii) ?? "\(t)"
+    }
+}

clients/decky/src/index.tsx

@@ -294,7 +294,13 @@ const RESOLUTIONS: [number, number, string][] = [
   [2560, 1440, "2560 × 1440"],
 ];
 const REFRESH = [0, 30, 60, 90, 120];
-const GAMEPADS = ["auto", "xbox360", "dualsense"];
+const GAMEPADS = ["auto", "xbox360", "dualsense", "steamdeck"];
+const GAMEPAD_LABELS: Record<string, string> = {
+  auto: "Automatic",
+  xbox360: "Xbox 360",
+  dualsense: "DualSense",
+  steamdeck: "Steam Deck",
+};
 
 const SettingsSection: FC = () => {
   const [s, setS] = useState<StreamSettings | null>(null);
@@ -355,14 +361,17 @@ const SettingsSection: FC = () => {
       />
       <Field label="Gamepad type" childrenContainerWidth="max">
         <Dropdown
-          rgOptions={GAMEPADS.map((g) => ({
+          rgOptions={GAMEPADS.map((g) => ({ data: g, label: GAMEPAD_LABELS[g] ?? g }))}
-            data: g,
-            label: g === "auto" ? "Automatic" : g === "xbox360" ? "Xbox 360" : "DualSense",
-          }))}
           selectedOption={s.gamepad}
           onChange={(o) => patch({ gamepad: o.data as string })}
         />
       </Field>
+      {s.gamepad === "steamdeck" && (
+        <Field
+          label="⚠ Disable Steam Input"
+          description="Steam Deck mode forwards the paddles, both trackpads, and gyro to the host. For that, Steam Input must be OFF for punktfunk: on the game page tap ⚙ → Controller Settings → set Steam Input to Off. Otherwise Steam keeps the Deck's controls and only the sticks + buttons reach the host."
+        />
+      )}
       <ToggleField
         label="Stream microphone"
         checked={s.mic_enabled}

clients/decky/src/steam.ts

@@ -113,12 +113,35 @@ async function ensureShortcut(): Promise<number> {
   return appId;
 }
 
+/**
+ * Best-effort: turn Steam Input OFF for our shortcut so SDL's HIDAPI Steam Deck driver can open the
+ * Deck's controls (paddles · trackpads · gyro) directly. There is no confirmed-stable SteamClient
+ * API for this, so it is feature-detected and MUST never block or throw into the launch — the manual
+ * toggle (game page → ⚙ → Controller Settings → Steam Input Off, surfaced in the plugin Settings) is
+ * the documented source of truth. No-op when the optional API is absent.
+ */
+function disableSteamInputForShortcut(appId: number): void {
+  try {
+    const input = (
+      SteamClient as unknown as {
+        Input?: { SetSteamInputEnabledForApp?: (appId: number, enabled: boolean) => void };
+      }
+    ).Input;
+    input?.SetSteamInputEnabledForApp?.(appId, false);
+  } catch {
+    /* a controller tweak must never break the launch */
+  }
+}
+
 /**
  * Launch a stream to `host:port` fullscreen in Gaming Mode. Encodes the target into the
  * shortcut's launch options (so one generic shortcut serves every host), then RunGame.
  */
 export async function launchStream(host: string, port: number): Promise<void> {
   const appId = await ensureShortcut();
+  // Best-effort so the Deck's rich controls reach the client; no-op if the API is absent (the user
+  // disables Steam Input manually — see the Settings instruction).
+  disableSteamInputForShortcut(appId);
   const target = port && port !== 9777 ? `${host}:${port}` : host;
   // KEY=value ... %command% — the wrapper reads PF_HOST from the environment.
   SteamClient.Apps.SetAppLaunchOptions(appId, `PF_HOST=${target} %command%`);

clients/linux/src/app.rs

@@ -767,6 +767,7 @@ fn start_session_with(app: Rc<App>, req: ConnectRequest, pin: Option<[u8; 32]>,
                         connector,
                         frames.take().expect("Connected delivered once"),
                         app.gamepad.escape_events(),
+                        app.gamepad.disconnect_events(),
                         handle.stop.clone(),
                         inhibit,
                         &title,

clients/linux/src/gamepad.rs

@@ -18,7 +18,7 @@ use punktfunk_core::quic::{HidOutput, RichInput};
 use std::collections::HashMap;
 use std::sync::mpsc::{Receiver, Sender};
 use std::sync::{Arc, Mutex};
-use std::time::Duration;
+use std::time::{Duration, Instant};
 
 /// Motion scale constants, shared convention with the Swift client (`GamepadWire`):
 /// derived from hid-playstation's math over the host's fixed calibration blob. SDL hands
@@ -33,8 +33,15 @@ const G: f32 = 9.80665;
 /// is the only way out. Four simultaneous buttons that no game uses as a deliberate
 /// combo, so it can't be triggered by normal play. Still forwarded to the host (the user
 /// is leaving anyway); we only also raise the escape signal.
+///
+/// **Escalation:** a quick press leaves fullscreen / releases capture; *holding* the same
+/// chord for [`DISCONNECT_HOLD`] ends the session. Deliberately NOT the Steam / QAM buttons —
+/// those are the marquee pass-through controls that now reach the host's game-mode UI.
 const ESCAPE_CHORD: [u32; 4] = [wire::BTN_LB, wire::BTN_RB, wire::BTN_START, wire::BTN_BACK];
 
+/// Hold the [`ESCAPE_CHORD`] at least this long to disconnect (escalates the leave-fullscreen press).
+const DISCONNECT_HOLD: Duration = Duration::from_millis(1500);
+
 #[derive(Clone, Debug)]
 pub struct PadInfo {
     pub id: u32,
@@ -58,6 +65,7 @@ impl PadInfo {
             GamepadPref::DualSense => "DualSense",
             GamepadPref::DualShock4 => "DualShock 4",
             GamepadPref::XboxOne => "Xbox One",
+            GamepadPref::SteamDeck => "Steam Deck",
             _ => "",
         }
     }
@@ -89,6 +97,9 @@ pub struct GamepadService {
     /// Fires once per press of the [`ESCAPE_CHORD`]; the stream page consumes it to leave
     /// fullscreen + release capture.
     escape_rx: async_channel::Receiver<()>,
+    /// Fires once when the [`ESCAPE_CHORD`] is held past [`DISCONNECT_HOLD`]; the stream page
+    /// consumes it to end the session (the controller equivalent of Ctrl+Alt+Shift+D).
+    disconnect_rx: async_channel::Receiver<()>,
 }
 
 impl GamepadService {
@@ -98,11 +109,12 @@ impl GamepadService {
         let pinned = Arc::new(Mutex::new(None));
         let (ctl, ctl_rx) = std::sync::mpsc::channel();
         let (escape_tx, escape_rx) = async_channel::unbounded();
+        let (disconnect_tx, disconnect_rx) = async_channel::unbounded();
         let (p, a, pin) = (pads.clone(), active.clone(), pinned.clone());
         if let Err(e) = std::thread::Builder::new()
             .name("punktfunk-gamepad".into())
             .spawn(move || {
-                if let Err(e) = run(&p, &a, &pin, &ctl_rx, &escape_tx) {
+                if let Err(e) = run(&p, &a, &pin, &ctl_rx, &escape_tx, &disconnect_tx) {
                     tracing::warn!(error = %e, "gamepad service ended — pads disabled");
                 }
             })
@@ -115,6 +127,7 @@ impl GamepadService {
             pinned,
             ctl,
             escape_rx,
+            disconnect_rx,
         }
     }
 
@@ -124,6 +137,12 @@ impl GamepadService {
         self.escape_rx.clone()
     }
 
+    /// A receiver that yields one `()` when the escape chord is held past [`DISCONNECT_HOLD`]
+    /// (controller disconnect). A fresh clone per call; the stream page spawns a future on it.
+    pub fn disconnect_events(&self) -> async_channel::Receiver<()> {
+        self.disconnect_rx.clone()
+    }
+
     pub fn pads(&self) -> Vec<PadInfo> {
         self.pads.lock().unwrap().clone()
     }
@@ -188,6 +207,13 @@ fn button_bit(b: sdl3::gamepad::Button) -> Option<u32> {
         Button::DPadLeft => wire::BTN_DPAD_LEFT,
         Button::DPadRight => wire::BTN_DPAD_RIGHT,
         Button::Touchpad => wire::BTN_TOUCHPAD,
+        // Back grips / paddles (Steam Deck L4/L5/R4/R5, Xbox Elite P1–P4) + the misc/Share button.
+        // PADDLE1/2/3/4 = R4/L4/R5/L5 (see the host `input::gamepad`).
+        Button::RightPaddle1 => wire::BTN_PADDLE1,
+        Button::LeftPaddle1 => wire::BTN_PADDLE2,
+        Button::RightPaddle2 => wire::BTN_PADDLE3,
+        Button::LeftPaddle2 => wire::BTN_PADDLE4,
+        Button::Misc1 => wire::BTN_MISC1,
         _ => return None,
     })
 }
@@ -259,11 +285,22 @@ struct Worker {
     /// Wire state of the active pad — zeroed on the wire at switch/detach.
     last_axis: [i32; 6],
     held_buttons: Vec<u32>,
+    /// Touchpad contacts the host believes are down, keyed by `(surface, finger)` — lifted on pad
+    /// switch / detach so a contact held at that moment doesn't stick. surface 0 = the legacy single
+    /// touchpad, 1/2 = a Steam left/right pad.
+    held_touches: std::collections::HashSet<(u8, u8)>,
     last_accel: [i16; 3],
     /// Raises the UI escape signal; the escape chord fires it once per press.
     escape_tx: async_channel::Sender<()>,
+    /// Raises the UI disconnect signal when the escape chord is held past [`DISCONNECT_HOLD`].
+    disconnect_tx: async_channel::Sender<()>,
     /// The escape chord is fully held — latched so it fires once, not every poll.
     chord_armed: bool,
+    /// When the escape chord became fully held (drives the hold-to-disconnect escalation); `None`
+    /// when the chord is broken.
+    chord_since: Option<Instant>,
+    /// The disconnect signal already fired for the current hold — latched so it fires once.
+    disconnect_fired: bool,
 }
 
 impl Worker {
@@ -275,13 +312,22 @@ impl Worker {
 
     fn pad_info(&self, id: u32) -> Option<PadInfo> {
         let pad = self.opened.get(&id)?;
+        let mut pref = pref_for_type(
+            self.subsystem
+                .type_for_id(sdl3::sys::joystick::SDL_JoystickID(id)),
+        );
+        // There is no SDL gamepad type for the Steam Deck / Steam Controller, so detect Valve by
+        // VID/PID (Deck 0x1205, SC wired 0x1102, SC dongle 0x1142) — the host then builds the virtual
+        // hid-steam pad with the back grips + dual trackpads and the right glyph identity.
+        if pad.vendor_id() == Some(0x28DE)
+            && matches!(pad.product_id(), Some(0x1205 | 0x1102 | 0x1142))
+        {
+            pref = GamepadPref::SteamDeck;
+        }
         Some(PadInfo {
             id,
             name: pad.name().unwrap_or_else(|| "Controller".into()),
-            pref: pref_for_type(
+            pref,
-                self.subsystem
-                    .type_for_id(sdl3::sys::joystick::SDL_JoystickID(id)),
-            ),
         })
     }
 
@@ -297,32 +343,90 @@ impl Worker {
                 }
                 *v = i32::MIN;
             }
+            // Lift any touchpad contact the host still believes is down (surface 0 = legacy pad).
+            for (surface, finger) in self.held_touches.drain() {
+                let rich = if surface == 0 {
+                    RichInput::Touchpad {
+                        pad: 0,
+                        finger,
+                        active: false,
+                        x: 0,
+                        y: 0,
+                    }
+                } else {
+                    RichInput::TouchpadEx {
+                        pad: 0,
+                        surface,
+                        finger,
+                        touch: false,
+                        click: false,
+                        x: 0,
+                        y: 0,
+                        pressure: 0,
+                    }
+                };
+                let _ = c.send_rich_input(rich);
+            }
         } else {
             self.held_buttons.clear();
             self.last_axis = [i32::MIN; 6];
+            self.held_touches.clear();
         }
+        // A held chord doesn't survive a flush (detach / pad-switch) — clear its latches too.
+        self.reset_chord();
     }
 
     /// Raise the UI escape signal when the [`ESCAPE_CHORD`] just completed (latched so it
-    /// fires once per press). Called after each button-down updates `held_buttons`.
+    /// fires once per press) and start the hold-to-disconnect timer. Called after each
+    /// button-down updates `held_buttons`.
     fn maybe_fire_escape(&mut self) {
         if self.chord_armed {
             return;
         }
         if ESCAPE_CHORD.iter().all(|b| self.held_buttons.contains(b)) {
             self.chord_armed = true;
+            self.chord_since = Some(Instant::now());
             let _ = self.escape_tx.try_send(());
-            tracing::info!("gamepad escape chord (L1+R1+Start+Select) — leaving fullscreen");
+            tracing::info!(
+                "gamepad escape chord (L1+R1+Start+Select) — leaving fullscreen (hold to disconnect)"
+            );
+        }
+    }
+
+    /// Fire the disconnect signal once the escape chord has been continuously held past
+    /// [`DISCONNECT_HOLD`]. Polled from the main loop so the hold completes without new events.
+    fn maybe_fire_disconnect(&mut self) {
+        if self.disconnect_fired {
+            return;
+        }
+        if let Some(since) = self.chord_since {
+            if since.elapsed() >= DISCONNECT_HOLD {
+                self.disconnect_fired = true;
+                let _ = self.disconnect_tx.try_send(());
+                tracing::info!("gamepad escape chord held — disconnecting");
+            }
         }
     }
 
     /// Re-arm once the chord is broken (any of its buttons released).
     fn rearm_escape(&mut self) {
         if self.chord_armed && !ESCAPE_CHORD.iter().all(|b| self.held_buttons.contains(b)) {
-            self.chord_armed = false;
+            self.reset_chord();
         }
     }
 
+    /// Clear the escape/disconnect chord latches. Called at every session boundary
+    /// ([`flush_held`](Self::flush_held) on detach/pad-switch + on attach): the hold-to-disconnect
+    /// path *always* ends the session while the chord is still physically held, so the matching
+    /// button-up events arrive after detach (dropped by the `attached` guard) and `rearm_escape`
+    /// never runs — without this the latched state would leak into the next session and either
+    /// swallow its first chord press or fire a stale disconnect on connect.
+    fn reset_chord(&mut self) {
+        self.chord_armed = false;
+        self.chord_since = None;
+        self.disconnect_fired = false;
+    }
+
     /// Sensors stream only while a session wants them (they cost USB/BT bandwidth).
     fn set_sensors(&mut self, enabled: bool) {
         let Some(id) = self.active_id() else { return };
@@ -335,6 +439,56 @@ impl Worker {
             }
         }
     }
+
+    /// Forward one touchpad contact on the rich-input plane. A multi-touchpad pad (Steam Deck / Steam
+    /// Controller) sends `TouchpadEx` with the surface (SDL touchpad 0 = left → 1, 1 = right → 2) and
+    /// signed coordinates; a single-touchpad pad (DualSense) keeps the legacy `Touchpad` (unsigned).
+    fn forward_touch(
+        &mut self,
+        which: u32,
+        touchpad: u32,
+        finger: u8,
+        x: f32,
+        y: f32,
+        active: bool,
+    ) {
+        let Some(c) = self.attached.as_ref() else {
+            return;
+        };
+        let multi = self
+            .opened
+            .get(&which)
+            .map(|p| p.touchpads_count() >= 2)
+            .unwrap_or(false);
+        let (cx, cy) = (x.clamp(0.0, 1.0), y.clamp(0.0, 1.0));
+        let surface = if multi { (touchpad as u8) + 1 } else { 0 };
+        let rich = if multi {
+            RichInput::TouchpadEx {
+                pad: 0,
+                surface,
+                finger,
+                touch: active,
+                click: false,
+                x: (cx * 65535.0 - 32768.0) as i16,
+                y: (cy * 65535.0 - 32768.0) as i16,
+                pressure: 0,
+            }
+        } else {
+            RichInput::Touchpad {
+                pad: 0,
+                finger,
+                active,
+                x: (cx * 65535.0) as u16,
+                y: (cy * 65535.0) as u16,
+            }
+        };
+        let _ = c.send_rich_input(rich);
+        if active {
+            self.held_touches.insert((surface, finger));
+        } else {
+            self.held_touches.remove(&(surface, finger));
+        }
+    }
 }
 
 #[allow(clippy::too_many_lines)]
@@ -344,11 +498,18 @@ fn run(
     pinned_out: &Mutex<Option<u32>>,
     ctl: &Receiver<Ctl>,
     escape_tx: &async_channel::Sender<()>,
+    disconnect_tx: &async_channel::Sender<()>,
 ) -> Result<(), String> {
     // Off-main-thread + no video subsystem: keep SDL away from signals, poll pads on its
     // own thread.
     sdl3::hint::set("SDL_NO_SIGNAL_HANDLERS", "1");
     sdl3::hint::set("SDL_JOYSTICK_THREAD", "1");
+    // Let SDL's HIDAPI drivers open Valve Steam Controller / Steam Deck devices directly, so the
+    // paddles, both trackpads, and gyro arrive as first-class SDL gamepad inputs. On a Deck in Game
+    // Mode, Steam Input still holds the device — the user must disable Steam Input for this app (see
+    // the Decky UX); on a desktop client (or a Deck with Steam Input off) the hints just work.
+    sdl3::hint::set("SDL_JOYSTICK_HIDAPI_STEAMDECK", "1");
+    sdl3::hint::set("SDL_JOYSTICK_HIDAPI_STEAM", "1");
     let sdl = sdl3::init().map_err(|e| e.to_string())?;
     let subsystem = sdl.gamepad().map_err(|e| e.to_string())?;
     let mut pump = sdl.event_pump().map_err(|e| e.to_string())?;
@@ -361,9 +522,13 @@ fn run(
         attached: None,
         last_axis: [i32::MIN; 6],
         held_buttons: Vec::new(),
+        held_touches: std::collections::HashSet::new(),
         last_accel: [0; 3],
         escape_tx: escape_tx.clone(),
+        disconnect_tx: disconnect_tx.clone(),
         chord_armed: false,
+        chord_since: None,
+        disconnect_fired: false,
     };
 
     let publish = |w: &Worker| {
@@ -381,6 +546,7 @@ fn run(
                 Ok(Ctl::Attach(c)) => {
                     w.attached = Some(c);
                     w.last_axis = [i32::MIN; 6];
+                    w.reset_chord(); // every session starts un-latched (Attach doesn't flush)
                     w.set_sensors(true);
                 }
                 Ok(Ctl::Detach) => {
@@ -474,9 +640,11 @@ fn run(
                         send(w.attached.as_ref().unwrap(), InputKind::GamepadAxis, id, v);
                     }
                 }
-                // DualSense touchpad → the rich-input plane, normalized 0..=65535.
+                // Touchpad contacts → the rich-input plane. One pad (DualSense) keeps the legacy
+                // `Touchpad`; two pads (Steam Deck / Steam Controller) send `TouchpadEx` per surface.
                 Event::ControllerTouchpadDown {
                     which,
+                    touchpad,
                     finger,
                     x,
                     y,
@@ -484,41 +652,23 @@ fn run(
                 }
                 | Event::ControllerTouchpadMotion {
                     which,
+                    touchpad,
                     finger,
                     x,
                     y,
                     ..
                 } if active == Some(which) && w.attached.is_some() => {
-                    let _ = w
+                    w.forward_touch(which, touchpad as u32, finger as u8, x, y, true);
-                        .attached
-                        .as_ref()
-                        .unwrap()
-                        .send_rich_input(RichInput::Touchpad {
-                            pad: 0,
-                            finger: finger as u8,
-                            active: true,
-                            x: (x.clamp(0.0, 1.0) * 65535.0) as u16,
-                            y: (y.clamp(0.0, 1.0) * 65535.0) as u16,
-                        });
                 }
                 Event::ControllerTouchpadUp {
                     which,
+                    touchpad,
                     finger,
                     x,
                     y,
                     ..
                 } if active == Some(which) && w.attached.is_some() => {
-                    let _ = w
+                    w.forward_touch(which, touchpad as u32, finger as u8, x, y, false);
-                        .attached
-                        .as_ref()
-                        .unwrap()
-                        .send_rich_input(RichInput::Touchpad {
-                            pad: 0,
-                            finger: finger as u8,
-                            active: false,
-                            x: (x.clamp(0.0, 1.0) * 65535.0) as u16,
-                            y: (y.clamp(0.0, 1.0) * 65535.0) as u16,
-                        });
                 }
                 // Motion: accel events update the cache; each gyro event ships a sample
                 // (the DualSense reports both at ~250 Hz). Scale convention shared with
@@ -559,6 +709,10 @@ fn run(
             }
         }
 
+        // Escalate a held escape chord to a disconnect (polled — the hold completes with no
+        // new button events; the chord itself is only detected while a session is attached).
+        w.maybe_fire_disconnect();
+
         // Feedback planes (this thread is their single consumer). The host re-sends
         // rumble state periodically, so a generous duration with refresh-on-update is
         // safe — a dropped stop heals within ~500 ms.

clients/linux/src/ui_stream.rs

@@ -124,12 +124,13 @@ impl Capture {
     }
 }
 
-#[allow(clippy::too_many_lines)]
+#[allow(clippy::too_many_lines, clippy::too_many_arguments)]
 pub fn new(
     window: &adw::ApplicationWindow,
     connector: Arc<NativeClient>,
     frames: async_channel::Receiver<DecodedFrame>,
     escape_rx: async_channel::Receiver<()>,
+    disconnect_rx: async_channel::Receiver<()>,
     stop: Arc<AtomicBool>,
     inhibit_shortcuts: bool,
     title: &str,
@@ -152,7 +153,7 @@ pub fn new(
     stats_label.set_margin_top(12);
 
     let hint = gtk::Label::new(Some(
-        "Click the stream to capture input · Ctrl+Alt+Shift+Q releases",
+        "Click the stream to capture input · Ctrl+Alt+Shift+Q releases · Ctrl+Alt+Shift+D disconnects",
     ));
     hint.add_css_class("osd");
     hint.set_halign(gtk::Align::Center);
@@ -163,7 +164,9 @@ pub fn new(
     // Flashed when entering fullscreen — the only exit affordances once the header bar is
     // hidden (F11 on a keyboard; the L1+R1+Start+Select chord on a controller, which is the
     // only way out on a Steam Deck).
-    let fs_hint = gtk::Label::new(Some("F11  ·  L1 + R1 + Start + Select  —  exit fullscreen"));
+    let fs_hint = gtk::Label::new(Some(
+        "F11  ·  L1 + R1 + Start + Select  —  exit fullscreen (hold to disconnect)",
+    ));
     fs_hint.add_css_class("osd");
     fs_hint.set_halign(gtk::Align::Center);
     fs_hint.set_valign(gtk::Align::Start);
@@ -297,6 +300,7 @@ pub fn new(
         key.set_propagation_phase(gtk::PropagationPhase::Capture);
         let cap = capture.clone();
         let window_k = window.clone();
+        let stop_kb = stop.clone();
         key.connect_key_pressed(move |_, keyval, keycode, state| {
             let chord = gdk::ModifierType::CONTROL_MASK
                 | gdk::ModifierType::ALT_MASK
@@ -309,6 +313,13 @@ pub fn new(
                 }
                 return glib::Propagation::Stop;
             }
+            // Ctrl+Alt+Shift+D — leave the session. Now that Steam / QAM pass through to the host,
+            // the capture toggle alone can't end a stream, so this is the keyboard's explicit exit.
+            if state.contains(chord) && keyval.to_lower() == gdk::Key::d {
+                cap.release();
+                stop_kb.store(true, Ordering::SeqCst);
+                return glib::Propagation::Stop;
+            }
             if keyval == gdk::Key::F11 {
                 if window_k.is_fullscreen() {
                     window_k.unfullscreen();
@@ -442,6 +453,24 @@ pub fn new(
         })
     };
 
+    // Controller disconnect (escape chord held past the hold threshold) → end the session, the
+    // controller equivalent of Ctrl+Alt+Shift+D. Setting `stop` ends the session pump, which pops
+    // this page (and fires `hidden` below). One-shot — the session is going away.
+    let disconnect_future = {
+        let window = window.clone();
+        let cap = capture.clone();
+        let stop_d = stop.clone();
+        glib::spawn_future_local(async move {
+            if disconnect_rx.recv().await.is_ok() {
+                cap.release();
+                if window.is_fullscreen() {
+                    window.unfullscreen();
+                }
+                stop_d.store(true, Ordering::SeqCst);
+            }
+        })
+    };
+
     // The page's `hidden` fires once navigation away completes (back button, pop on
     // session end) — NOT on the transient unmap/map cycle a NavigationView push performs.
     {
@@ -449,6 +478,7 @@ pub fn new(
         let stop_h = stop.clone();
         let handlers = RefCell::new(Some((fs_handler, active_handler)));
         let escape_future = RefCell::new(Some(escape_future));
+        let disconnect_future = RefCell::new(Some(disconnect_future));
         page.connect_hidden(move |_| {
             tracing::debug!("stream page hidden — ending session");
             if let Some((fs, active)) = handlers.borrow_mut().take() {
@@ -458,6 +488,9 @@ pub fn new(
             if let Some(f) = escape_future.borrow_mut().take() {
                 f.abort();
             }
+            if let Some(f) = disconnect_future.borrow_mut().take() {
+                f.abort();
+            }
             if window.is_fullscreen() {
                 window.unfullscreen();
             }

clients/windows/src/gamepad.rs

@@ -169,6 +169,13 @@ fn button_bit(b: sdl3::gamepad::Button) -> Option<u32> {
         Button::DPadLeft => wire::BTN_DPAD_LEFT,
         Button::DPadRight => wire::BTN_DPAD_RIGHT,
         Button::Touchpad => wire::BTN_TOUCHPAD,
+        // Back grips / paddles (Steam Deck L4/L5/R4/R5, Xbox Elite P1–P4) + the misc/Share button.
+        // PADDLE1/2/3/4 = R4/L4/R5/L5 (see the host `input::gamepad`).
+        Button::RightPaddle1 => wire::BTN_PADDLE1,
+        Button::LeftPaddle1 => wire::BTN_PADDLE2,
+        Button::RightPaddle2 => wire::BTN_PADDLE3,
+        Button::LeftPaddle2 => wire::BTN_PADDLE4,
+        Button::Misc1 => wire::BTN_MISC1,
         _ => return None,
     })
 }
@@ -240,6 +247,9 @@ struct Worker {
     /// Wire state of the active pad — zeroed on the wire at switch/detach.
     last_axis: [i32; 6],
     held_buttons: Vec<u32>,
+    /// Touchpad contacts the host believes are down, keyed by `(surface, finger)` — lifted on pad
+    /// switch / detach. surface 0 = the legacy single touchpad, 1/2 = a Steam left/right pad.
+    held_touches: std::collections::HashSet<(u8, u8)>,
     last_accel: [i16; 3],
 }
 
@@ -252,13 +262,21 @@ impl Worker {
 
     fn pad_info(&self, id: u32) -> Option<PadInfo> {
         let pad = self.opened.get(&id)?;
+        let mut pref = pref_for_type(
+            self.subsystem
+                .type_for_id(sdl3::sys::joystick::SDL_JoystickID(id)),
+        );
+        // No SDL type for the Steam Deck / Steam Controller — detect Valve by VID/PID (Deck 0x1205,
+        // SC wired 0x1102, SC dongle 0x1142) so the host builds the virtual hid-steam pad.
+        if pad.vendor_id() == Some(0x28DE)
+            && matches!(pad.product_id(), Some(0x1205 | 0x1102 | 0x1142))
+        {
+            pref = GamepadPref::SteamDeck;
+        }
         Some(PadInfo {
             id,
             name: pad.name().unwrap_or_else(|| "Controller".into()),
-            pref: pref_for_type(
+            pref,
-                self.subsystem
-                    .type_for_id(sdl3::sys::joystick::SDL_JoystickID(id)),
-            ),
         })
     }
 
@@ -274,9 +292,33 @@ impl Worker {
                 }
                 *v = i32::MIN;
             }
+            for (surface, finger) in self.held_touches.drain() {
+                let rich = if surface == 0 {
+                    RichInput::Touchpad {
+                        pad: 0,
+                        finger,
+                        active: false,
+                        x: 0,
+                        y: 0,
+                    }
+                } else {
+                    RichInput::TouchpadEx {
+                        pad: 0,
+                        surface,
+                        finger,
+                        touch: false,
+                        click: false,
+                        x: 0,
+                        y: 0,
+                        pressure: 0,
+                    }
+                };
+                let _ = c.send_rich_input(rich);
+            }
         } else {
             self.held_buttons.clear();
             self.last_axis = [i32::MIN; 6];
+            self.held_touches.clear();
         }
     }
 
@@ -292,6 +334,56 @@ impl Worker {
             }
         }
     }
+
+    /// Forward one touchpad contact on the rich-input plane. A multi-touchpad pad (Steam Deck / Steam
+    /// Controller) sends `TouchpadEx` with the surface (SDL touchpad 0 = left → 1, 1 = right → 2) and
+    /// signed coordinates; a single-touchpad pad (DualSense) keeps the legacy `Touchpad` (unsigned).
+    fn forward_touch(
+        &mut self,
+        which: u32,
+        touchpad: u32,
+        finger: u8,
+        x: f32,
+        y: f32,
+        active: bool,
+    ) {
+        let Some(c) = self.attached.as_ref() else {
+            return;
+        };
+        let multi = self
+            .opened
+            .get(&which)
+            .map(|p| p.touchpads_count() >= 2)
+            .unwrap_or(false);
+        let (cx, cy) = (x.clamp(0.0, 1.0), y.clamp(0.0, 1.0));
+        let surface = if multi { (touchpad as u8) + 1 } else { 0 };
+        let rich = if multi {
+            RichInput::TouchpadEx {
+                pad: 0,
+                surface,
+                finger,
+                touch: active,
+                click: false,
+                x: (cx * 65535.0 - 32768.0) as i16,
+                y: (cy * 65535.0 - 32768.0) as i16,
+                pressure: 0,
+            }
+        } else {
+            RichInput::Touchpad {
+                pad: 0,
+                finger,
+                active,
+                x: (cx * 65535.0) as u16,
+                y: (cy * 65535.0) as u16,
+            }
+        };
+        let _ = c.send_rich_input(rich);
+        if active {
+            self.held_touches.insert((surface, finger));
+        } else {
+            self.held_touches.remove(&(surface, finger));
+        }
+    }
 }
 
 #[allow(clippy::too_many_lines)]
@@ -305,6 +397,10 @@ fn run(
     // thread.
     sdl3::hint::set("SDL_NO_SIGNAL_HANDLERS", "1");
     sdl3::hint::set("SDL_JOYSTICK_THREAD", "1");
+    // Let SDL's HIDAPI drivers open Valve Steam Controller / Steam Deck devices directly, so the
+    // paddles, both trackpads, and gyro arrive as first-class SDL gamepad inputs.
+    sdl3::hint::set("SDL_JOYSTICK_HIDAPI_STEAMDECK", "1");
+    sdl3::hint::set("SDL_JOYSTICK_HIDAPI_STEAM", "1");
     let sdl = sdl3::init().map_err(|e| e.to_string())?;
     let subsystem = sdl.gamepad().map_err(|e| e.to_string())?;
     let mut pump = sdl.event_pump().map_err(|e| e.to_string())?;
@@ -317,6 +413,7 @@ fn run(
         attached: None,
         last_axis: [i32::MIN; 6],
         held_buttons: Vec::new(),
+        held_touches: std::collections::HashSet::new(),
         last_accel: [0; 3],
     };
 
@@ -426,9 +523,11 @@ fn run(
                         send(w.attached.as_ref().unwrap(), InputKind::GamepadAxis, id, v);
                     }
                 }
-                // DualSense touchpad → the rich-input plane, normalized 0..=65535.
+                // Touchpad contacts → the rich-input plane. One pad (DualSense) keeps the legacy
+                // `Touchpad`; two pads (Steam Deck / Steam Controller) send `TouchpadEx` per surface.
                 Event::ControllerTouchpadDown {
                     which,
+                    touchpad,
                     finger,
                     x,
                     y,
@@ -436,41 +535,23 @@ fn run(
                 }
                 | Event::ControllerTouchpadMotion {
                     which,
+                    touchpad,
                     finger,
                     x,
                     y,
                     ..
                 } if active == Some(which) && w.attached.is_some() => {
-                    let _ = w
+                    w.forward_touch(which, touchpad as u32, finger as u8, x, y, true);
-                        .attached
-                        .as_ref()
-                        .unwrap()
-                        .send_rich_input(RichInput::Touchpad {
-                            pad: 0,
-                            finger: finger as u8,
-                            active: true,
-                            x: (x.clamp(0.0, 1.0) * 65535.0) as u16,
-                            y: (y.clamp(0.0, 1.0) * 65535.0) as u16,
-                        });
                 }
                 Event::ControllerTouchpadUp {
                     which,
+                    touchpad,
                     finger,
                     x,
                     y,
                     ..
                 } if active == Some(which) && w.attached.is_some() => {
-                    let _ = w
+                    w.forward_touch(which, touchpad as u32, finger as u8, x, y, false);
-                        .attached
-                        .as_ref()
-                        .unwrap()
-                        .send_rich_input(RichInput::Touchpad {
-                            pad: 0,
-                            finger: finger as u8,
-                            active: false,
-                            x: (x.clamp(0.0, 1.0) * 65535.0) as u16,
-                            y: (y.clamp(0.0, 1.0) * 65535.0) as u16,
-                        });
                 }
                 // Motion: accel events update the cache; each gyro event ships a sample (the
                 // DualSense reports both at ~250 Hz). Scale convention shared with the other

crates/pf-driver-proto/src/lib.rs

@@ -80,7 +80,14 @@ pub mod control {
         pub width: u32,
         pub height: u32,
         pub refresh_hz: u32,
-        pub _reserved: u32,
+        /// Host-preferred per-client monitor id (`1..=15`) — the EDID serial / IddCx `ConnectorIndex` /
+        /// `ContainerId` the driver names this monitor by. A given client (keyed by its cert fingerprint)
+        /// gets a STABLE id across reconnects, so the OS device path + EDID stay identical and Windows
+        /// reapplies that client's saved per-monitor config (DPI scaling). `0` = AUTO: the driver
+        /// allocates the lowest-free id (the original slot-based behavior — used for anonymous/TOFU and
+        /// GameStream sessions). Byte-compatible with the old `_reserved` (offset 20): an un-upgraded
+        /// driver ignores it (→ auto), which the host detects via [`AddReply::resolved_monitor_id`].
+        pub preferred_monitor_id: u32,
     }
 
     /// `IOCTL_ADD` reply: the OS target id + the adapter LUID the IDD landed on (split low/high to
@@ -91,7 +98,11 @@ pub mod control {
         pub adapter_luid_low: u32,
         pub adapter_luid_high: i32,
         pub target_id: u32,
-        pub _reserved: u32,
+        /// The monitor id the driver ACTUALLY used — echoes [`AddRequest::preferred_monitor_id`] when the
+        /// preference was honored, or the auto-allocated id otherwise. Byte-compatible with the old
+        /// `_reserved` (offset 12): an un-upgraded driver leaves it `0`, so the host can tell its
+        /// preference was ignored (stale driver) and log it instead of silently losing per-client config.
+        pub resolved_monitor_id: u32,
     }
 
     /// `IOCTL_REMOVE` input.
@@ -129,11 +140,13 @@ pub mod control {
         assert!(offset_of!(AddRequest, width) == 8);
         assert!(offset_of!(AddRequest, height) == 12);
         assert!(offset_of!(AddRequest, refresh_hz) == 16);
+        assert!(offset_of!(AddRequest, preferred_monitor_id) == 20);
 
         assert!(size_of::<AddReply>() == 16);
         assert!(offset_of!(AddReply, adapter_luid_low) == 0);
         assert!(offset_of!(AddReply, adapter_luid_high) == 4);
         assert!(offset_of!(AddReply, target_id) == 8);
+        assert!(offset_of!(AddReply, resolved_monitor_id) == 12);
 
         assert!(size_of::<RemoveRequest>() == 8);
         assert!(offset_of!(RemoveRequest, session_id) == 0);
@@ -436,11 +449,25 @@ mod tests {
             width: 3840,
             height: 2160,
             refresh_hz: 120,
-            _reserved: 0,
+            preferred_monitor_id: 7,
         };
         let bytes = bytemuck::bytes_of(&req);
         assert_eq!(bytes.len(), 24);
         assert_eq!(*bytemuck::from_bytes::<control::AddRequest>(bytes), req);
+        // preferred_monitor_id occupies the old `_reserved` slot at offset 20 — byte-compatible.
+        assert_eq!(bytes[20..24], 7u32.to_le_bytes());
+
+        let reply = control::AddReply {
+            adapter_luid_low: 0x1234_5678,
+            adapter_luid_high: -2,
+            target_id: 262,
+            resolved_monitor_id: 7,
+        };
+        let rbytes = bytemuck::bytes_of(&reply);
+        assert_eq!(rbytes.len(), 16);
+        assert_eq!(*bytemuck::from_bytes::<control::AddReply>(rbytes), reply);
+        // resolved_monitor_id occupies the old `_reserved` slot at offset 12 — byte-compatible.
+        assert_eq!(rbytes[12..16], 7u32.to_le_bytes());
     }
 
     #[test]

crates/punktfunk-core/src/abi.rs

@@ -492,6 +492,10 @@ pub const PUNKTFUNK_HIDOUT_LED: u8 = 1;
 pub const PUNKTFUNK_HIDOUT_PLAYER_LEDS: u8 = 2;
 /// `PunktfunkHidOutput::kind` — one adaptive-trigger effect (`which` + `effect`/`effect_len` valid).
 pub const PUNKTFUNK_HIDOUT_TRIGGER: u8 = 3;
+/// `PunktfunkHidOutput::kind` — a trackpad haptic pulse (Steam Controller voice-coils). `which` =
+/// side (0 = right pad, 1 = left pad); `effect[0..6]` packs `amplitude` / `period` / `count` as
+/// little-endian `u16`s with `effect_len = 6`. Clients without trackpad coils drop it.
+pub const PUNKTFUNK_HIDOUT_TRACKPAD_HAPTIC: u8 = 4;
 /// Capacity of `PunktfunkHidOutput::effect` (the DualSense trigger parameter block).
 pub const PUNKTFUNK_HID_EFFECT_MAX: u8 = 11;
 
@@ -559,6 +563,23 @@ impl PunktfunkHidOutput {
                 out.effect[..n].copy_from_slice(&effect[..n]);
                 out.effect_len = n as u8;
             }
+            HidOutput::TrackpadHaptic {
+                pad,
+                side,
+                amplitude,
+                period,
+                count,
+            } => {
+                // No new struct (PunktfunkHidOutput has no size guard): pack into the existing
+                // `which` (side) + `effect[0..6]` (amplitude/period/count LE), `effect_len = 6`.
+                out.kind = PUNKTFUNK_HIDOUT_TRACKPAD_HAPTIC;
+                out.pad = *pad;
+                out.which = *side;
+                out.effect[0..2].copy_from_slice(&amplitude.to_le_bytes());
+                out.effect[2..4].copy_from_slice(&period.to_le_bytes());
+                out.effect[4..6].copy_from_slice(&count.to_le_bytes());
+                out.effect_len = 6;
+            }
         }
         out
     }
@@ -618,6 +639,11 @@ impl PunktfunkHdrMeta {
 pub const PUNKTFUNK_RICH_TOUCHPAD: u8 = 1;
 /// `PunktfunkRichInput::kind` — a motion sample (`gyro`/`accel` valid).
 pub const PUNKTFUNK_RICH_MOTION: u8 = 2;
+/// `RichInput::TouchpadEx` kind on the wire — an extended trackpad contact that identifies the
+/// surface (0 single / 1 Steam-left / 2 Steam-right) and carries click + pressure. The host decodes
+/// it today; *sending* it from a C client needs the size-prefixed `PunktfunkRichInputEx` +
+/// `punktfunk_connection_send_rich_input2` (added with client capture).
+pub const PUNKTFUNK_RICH_TOUCHPAD_EX: u8 = 3;
 
 /// One rich client→host input for the host's virtual DualSense
 /// ([`punktfunk_connection_send_rich_input`]): a touchpad contact or a motion sample. Set `kind`
@@ -666,6 +692,77 @@ impl PunktfunkRichInput {
     }
 }
 
+/// Forward-compatible superset of [`PunktfunkRichInput`] that can also express the rich Steam
+/// surfaces: a *second* trackpad (`surface`), a distinct `click` vs touch, signed coordinates, and
+/// pressure. Sent via [`punktfunk_connection_send_rich_input2`] — the only way a C client can emit a
+/// `TouchpadEx`. The caller MUST set `struct_size = sizeof(PunktfunkRichInputEx)` (the ABI-skew
+/// guard, like [`PunktfunkConfig`]); the legacy [`PunktfunkRichInput`] +
+/// [`punktfunk_connection_send_rich_input`] stay byte-for-byte for existing callers.
+#[cfg(feature = "quic")]
+#[repr(C)]
+#[derive(Clone, Copy)]
+pub struct PunktfunkRichInputEx {
+    /// MUST equal `sizeof(PunktfunkRichInputEx)`.
+    pub struct_size: u32,
+    /// One of `PUNKTFUNK_RICH_*` (`TOUCHPAD` / `MOTION` / `TOUCHPAD_EX`).
+    pub kind: u8,
+    /// Gamepad index.
+    pub pad: u8,
+    /// Touchpad/TouchpadEx: contact id.
+    pub finger: u8,
+    /// Touchpad/TouchpadEx: 1 = finger down / touching, 0 = lifted.
+    pub active: u8,
+    /// TouchpadEx: which surface — 0 = single/DualSense, 1 = Steam left pad, 2 = Steam right pad.
+    pub surface: u8,
+    /// TouchpadEx: 1 = the pad is physically clicked (depressed), distinct from a touch contact.
+    pub click: u8,
+    /// Reserved for alignment; set to 0.
+    pub _reserved: [u8; 2],
+    /// TouchpadEx: x coordinate — **signed**, centred at 0 (the real Steam report convention). For a
+    /// legacy `TOUCHPAD` kind sent through this struct, store the unsigned `0..=65535` value's bits.
+    pub x: i16,
+    /// TouchpadEx: y coordinate — signed, centred at 0.
+    pub y: i16,
+    /// TouchpadEx: contact pressure (`0` if the surface has no force sensor).
+    pub pressure: u16,
+    /// Motion: gyro (pitch, yaw, roll), raw signed-16.
+    pub gyro: [i16; 3],
+    /// Motion: accelerometer (x, y, z), raw signed-16.
+    pub accel: [i16; 3],
+}
+
+#[cfg(feature = "quic")]
+impl PunktfunkRichInputEx {
+    fn to_rich(self) -> Option<crate::quic::RichInput> {
+        use crate::quic::RichInput;
+        match self.kind {
+            PUNKTFUNK_RICH_TOUCHPAD_EX => Some(RichInput::TouchpadEx {
+                pad: self.pad,
+                surface: self.surface,
+                finger: self.finger,
+                touch: self.active != 0,
+                click: self.click != 0,
+                x: self.x,
+                y: self.y,
+                pressure: self.pressure,
+            }),
+            PUNKTFUNK_RICH_MOTION => Some(RichInput::Motion {
+                pad: self.pad,
+                gyro: self.gyro,
+                accel: self.accel,
+            }),
+            PUNKTFUNK_RICH_TOUCHPAD => Some(RichInput::Touchpad {
+                pad: self.pad,
+                finger: self.finger,
+                active: self.active != 0,
+                x: self.x as u16,
+                y: self.y as u16,
+            }),
+            _ => None,
+        }
+    }
+}
+
 /// Read an optional NUL-terminated UTF-8 string parameter; `Err` = invalid pointer/UTF-8.
 #[cfg(feature = "quic")]
 unsafe fn opt_cstr<'a>(p: *const std::os::raw::c_char) -> std::result::Result<Option<&'a str>, ()> {
@@ -714,6 +811,22 @@ pub const PUNKTFUNK_GAMEPAD_XBOXONE: u32 = 3;
 /// DualSense (minus adaptive triggers / player LEDs / mute). Honored only where available (Linux
 /// hosts); otherwise the host falls back to X-Box 360.
 pub const PUNKTFUNK_GAMEPAD_DUALSHOCK4: u32 = 4;
+/// UHID classic Steam Controller (Valve `28DE:1102`, kernel `hid-steam`): dual trackpads, gyro,
+/// two grip paddles. Reserved — currently folds to `XBOX360` until its backend lands.
+pub const PUNKTFUNK_GAMEPAD_STEAMCONTROLLER: u32 = 5;
+/// UHID Steam Deck controller (Valve `28DE:1205`, kernel `hid-steam`): full Deck gamepad incl. the
+/// four back grips, a right trackpad, and the IMU; re-grabbed by Steam Input with native glyphs when
+/// Steam runs on the host. Honored only where available (Linux hosts); else folds to X-Box 360.
+pub const PUNKTFUNK_GAMEPAD_STEAMDECK: u32 = 6;
+
+/// Extended `InputEvent` gamepad button bits for embedders building raw events: the four back grips
+/// (Steam L4/L5/R4/R5 ≙ Xbox-Elite P1–P4) + the misc/capture button, in Moonlight's
+/// `buttonFlags2 << 16` namespace. Mirror `input::gamepad::BTN_PADDLE1..4` / `BTN_MISC1`.
+pub const PUNKTFUNK_GAMEPAD_BTN_PADDLE1: u32 = 0x0001_0000;
+pub const PUNKTFUNK_GAMEPAD_BTN_PADDLE2: u32 = 0x0002_0000;
+pub const PUNKTFUNK_GAMEPAD_BTN_PADDLE3: u32 = 0x0004_0000;
+pub const PUNKTFUNK_GAMEPAD_BTN_PADDLE4: u32 = 0x0008_0000;
+pub const PUNKTFUNK_GAMEPAD_BTN_MISC1: u32 = 0x0020_0000;
 
 /// Connect to a `punktfunk/1` host and start a session at `width`x`height`@`refresh_hz`.
 /// Blocks up to `timeout_ms` for the handshake. Returns NULL on failure. Equivalent to
@@ -742,11 +855,28 @@ const _: () = {
 // Keep the ABI gamepad constants in lockstep with the wire enum (compile-time guard against drift).
 const _: () = {
     use crate::config::GamepadPref;
+    use crate::input::gamepad as g;
     assert!(PUNKTFUNK_GAMEPAD_AUTO == GamepadPref::Auto.to_u8() as u32);
     assert!(PUNKTFUNK_GAMEPAD_XBOX360 == GamepadPref::Xbox360.to_u8() as u32);
     assert!(PUNKTFUNK_GAMEPAD_DUALSENSE == GamepadPref::DualSense.to_u8() as u32);
     assert!(PUNKTFUNK_GAMEPAD_XBOXONE == GamepadPref::XboxOne.to_u8() as u32);
     assert!(PUNKTFUNK_GAMEPAD_DUALSHOCK4 == GamepadPref::DualShock4.to_u8() as u32);
+    assert!(PUNKTFUNK_GAMEPAD_STEAMCONTROLLER == GamepadPref::SteamController.to_u8() as u32);
+    assert!(PUNKTFUNK_GAMEPAD_STEAMDECK == GamepadPref::SteamDeck.to_u8() as u32);
+    // Extended button bits mirror the wire `input::gamepad` constants.
+    assert!(PUNKTFUNK_GAMEPAD_BTN_PADDLE1 == g::BTN_PADDLE1);
+    assert!(PUNKTFUNK_GAMEPAD_BTN_PADDLE2 == g::BTN_PADDLE2);
+    assert!(PUNKTFUNK_GAMEPAD_BTN_PADDLE3 == g::BTN_PADDLE3);
+    assert!(PUNKTFUNK_GAMEPAD_BTN_PADDLE4 == g::BTN_PADDLE4);
+    assert!(PUNKTFUNK_GAMEPAD_BTN_MISC1 == g::BTN_MISC1);
+};
+
+// The additive M3 kinds (TouchpadEx / TrackpadHaptic) must never grow the legacy ABI structs —
+// they have no `struct_size` guard, so a layout change would corrupt old-built callers' buffers.
+#[cfg(feature = "quic")]
+const _: () = {
+    assert!(core::mem::size_of::<PunktfunkRichInput>() == 20);
+    assert!(core::mem::size_of::<PunktfunkHidOutput>() == 19);
 };
 
 /// Trust: `pin_sha256` (NULL or 32 bytes) is the expected SHA-256 fingerprint of the host's
@@ -1727,6 +1857,43 @@ pub unsafe extern "C" fn punktfunk_connection_send_rich_input(
     })
 }
 
+/// Send a rich client→host input via the forward-compatible [`PunktfunkRichInputEx`] — the only way
+/// a C client can emit a `TouchpadEx` (a second trackpad / signed coords / pressure). Set
+/// `rich->struct_size = sizeof(PunktfunkRichInputEx)`; a smaller (older-layout) value is rejected.
+///
+/// # Safety
+/// `c` is a valid connection handle; `rich` is null or points to at least its declared
+/// `struct_size` bytes.
+#[cfg(feature = "quic")]
+#[no_mangle]
+pub unsafe extern "C" fn punktfunk_connection_send_rich_input2(
+    c: *mut PunktfunkConnection,
+    rich: *const PunktfunkRichInputEx,
+) -> PunktfunkStatus {
+    guard(|| {
+        let c = match unsafe { c.as_ref() } {
+            Some(c) => c,
+            None => return PunktfunkStatus::NullPointer,
+        };
+        if rich.is_null() {
+            return PunktfunkStatus::NullPointer;
+        }
+        // Read only the 4-byte size prefix first to bound the subsequent full read (the
+        // `PunktfunkConfig` ABI-skew precedent).
+        let declared = unsafe { std::ptr::addr_of!((*rich).struct_size).read_unaligned() } as usize;
+        if declared < std::mem::size_of::<PunktfunkRichInputEx>() {
+            return PunktfunkStatus::InvalidArg;
+        }
+        match unsafe { *rich }.to_rich() {
+            Some(r) => match c.inner.send_rich_input(r) {
+                Ok(()) => PunktfunkStatus::Ok,
+                Err(e) => e.status(),
+            },
+            None => PunktfunkStatus::InvalidArg,
+        }
+    })
+}
+
 /// The currently active session mode — the Welcome's, until an accepted
 /// [`punktfunk_connection_request_mode`] switches it. Safe any time after connect.
 ///

crates/punktfunk-core/src/config.rs

@@ -137,8 +137,9 @@ impl CompositorPref {
 /// host decide (its `PUNKTFUNK_GAMEPAD` env var, else X-Box 360). A concrete preference is
 /// honored only if that backend is available on the host (DualSense / DualShock 4 need Linux UHID);
 /// otherwise the host falls back and reports the real choice in `Welcome`. The wire form is a single
-/// byte (`0 = Auto`, `1 = Xbox360`, `2 = DualSense`, `3 = XboxOne`, `4 = DualShock4`), appended to
+/// byte (`0 = Auto`, `1 = Xbox360`, `2 = DualSense`, `3 = XboxOne`, `4 = DualShock4`,
-/// `Hello`/`Welcome` — older peers simply omit/ignore it (an unknown byte degrades to `Auto`).
+/// `5 = SteamController`, `6 = SteamDeck`), appended to `Hello`/`Welcome` — older peers simply
+/// omit/ignore it (an unknown byte degrades to `Auto`).
 #[derive(Clone, Copy, Debug, PartialEq, Eq, Default)]
 pub enum GamepadPref {
     /// Let the host pick (its `PUNKTFUNK_GAMEPAD` env var, else X-Box 360).
@@ -155,10 +156,19 @@ pub enum GamepadPref {
     /// UHID DualShock 4 (kernel `hid-playstation`, ≥ 6.2) — lightbar, touchpad, motion, rumble. Like
     /// `DualSense` minus adaptive triggers / player LEDs / mute. Needs Linux UHID on the host.
     DualShock4,
+    /// UHID classic Steam Controller (Valve `28DE:1102`, kernel `hid-steam`) — dual trackpads, gyro,
+    /// two grip paddles, trackpad-only haptics. Needs Linux UHID. *(Reserved; its backend is not yet
+    /// built — currently folds to `Xbox360`; the Deck identity below is the implemented one.)*
+    SteamController,
+    /// UHID Steam Deck controller (Valve `28DE:1205`, kernel `hid-steam`) — full Deck gamepad incl.
+    /// the four back grips (L4/L5/R4/R5), a right trackpad, and the IMU; re-grabbed by Steam Input
+    /// with native glyphs when Steam runs on the host. Needs Linux UHID.
+    SteamDeck,
 }
 
 impl GamepadPref {
-    /// Wire byte. `0 = Auto`, `1 = Xbox360`, `2 = DualSense`, `3 = XboxOne`, `4 = DualShock4`.
+    /// Wire byte. `0 = Auto`, `1 = Xbox360`, `2 = DualSense`, `3 = XboxOne`, `4 = DualShock4`,
+    /// `5 = SteamController`, `6 = SteamDeck`.
     pub const fn to_u8(self) -> u8 {
         match self {
             GamepadPref::Auto => 0,
@@ -166,6 +176,8 @@ impl GamepadPref {
             GamepadPref::DualSense => 2,
             GamepadPref::XboxOne => 3,
             GamepadPref::DualShock4 => 4,
+            GamepadPref::SteamController => 5,
+            GamepadPref::SteamDeck => 6,
         }
     }
 
@@ -177,6 +189,8 @@ impl GamepadPref {
             2 => GamepadPref::DualSense,
             3 => GamepadPref::XboxOne,
             4 => GamepadPref::DualShock4,
+            5 => GamepadPref::SteamController,
+            6 => GamepadPref::SteamDeck,
             _ => GamepadPref::Auto,
         }
     }
@@ -192,12 +206,14 @@ impl GamepadPref {
                 GamepadPref::XboxOne
             }
             "dualshock4" | "dualshock" | "ds4" | "ps4" => GamepadPref::DualShock4,
+            "steamdeck" | "steam-deck" | "deck" => GamepadPref::SteamDeck,
+            "steamcontroller" | "steam-controller" | "steamcon" => GamepadPref::SteamController,
             _ => return None,
         })
     }
 
     /// Canonical lowercase identifier (`"auto"`, `"xbox360"`, `"dualsense"`, `"xboxone"`,
-    /// `"dualshock4"`).
+    /// `"dualshock4"`, `"steamcontroller"`, `"steamdeck"`).
     pub fn as_str(self) -> &'static str {
         match self {
             GamepadPref::Auto => "auto",
@@ -205,6 +221,8 @@ impl GamepadPref {
             GamepadPref::DualSense => "dualsense",
             GamepadPref::XboxOne => "xboxone",
             GamepadPref::DualShock4 => "dualshock4",
+            GamepadPref::SteamController => "steamcontroller",
+            GamepadPref::SteamDeck => "steamdeck",
         }
     }
 }
@@ -381,4 +399,27 @@ mod tests {
         c.fec.fec_percent = 15; // 250 + ceil(250*15/100)=288 > 255
         assert!(c.validate().is_err());
     }
+
+    #[test]
+    fn gamepad_pref_steam_roundtrip() {
+        use GamepadPref::*;
+        // Wire-byte round-trip for the Steam additions; an unknown byte still degrades to Auto.
+        for (p, b) in [(SteamController, 5u8), (SteamDeck, 6)] {
+            assert_eq!(p.to_u8(), b);
+            assert_eq!(GamepadPref::from_u8(b), p);
+        }
+        assert_eq!(GamepadPref::from_u8(99), Auto);
+        // Name parsing + canonical-name round-trip.
+        assert_eq!(GamepadPref::from_name("steamdeck"), Some(SteamDeck));
+        assert_eq!(GamepadPref::from_name("deck"), Some(SteamDeck));
+        assert_eq!(
+            GamepadPref::from_name("steamcontroller"),
+            Some(SteamController)
+        );
+        assert_eq!(SteamDeck.as_str(), "steamdeck");
+        assert_eq!(
+            GamepadPref::from_name(SteamController.as_str()),
+            Some(SteamController)
+        );
+    }
 }

crates/punktfunk-core/src/input.rs

@@ -66,10 +66,24 @@ pub mod gamepad {
     pub const BTN_B: u32 = 0x2000;
     pub const BTN_X: u32 = 0x4000;
     pub const BTN_Y: u32 = 0x8000;
+    // Extended buttons in Moonlight's `buttonFlags2 << 16` namespace (see `gamestream/gamepad.rs`),
+    // so the GameStream paddle path and the native path share one host injector map. The four Steam
+    // Deck back grips (L4/L5/R4/R5) reuse the four GameStream/Xbox-Elite paddle slots — a semantic
+    // 1:1 for binding (the device identity carries the glyph distinction).
+    /// Back grip R4 — SDL `RightPaddle1` / GameStream `PADDLE1`.
+    pub const BTN_PADDLE1: u32 = 0x0001_0000;
+    /// Back grip L4 — SDL `LeftPaddle1` / GameStream `PADDLE2`.
+    pub const BTN_PADDLE2: u32 = 0x0002_0000;
+    /// Back grip R5 — SDL `RightPaddle2` / GameStream `PADDLE3`.
+    pub const BTN_PADDLE3: u32 = 0x0004_0000;
+    /// Back grip L5 — SDL `LeftPaddle2` / GameStream `PADDLE4`.
+    pub const BTN_PADDLE4: u32 = 0x0008_0000;
     /// DualSense touchpad click. Moonlight's extended-button position (`buttonFlags2`
     /// merges in at `<< 16`, see `gamestream/gamepad.rs`), so GameStream clients land on
     /// the same bit. Only the DualSense backend renders it; the xpad has no such button.
     pub const BTN_TOUCHPAD: u32 = 0x10_0000;
+    /// Misc / capture button — the Deck `…`/quick-access, Share/Capture / GameStream `MISC`.
+    pub const BTN_MISC1: u32 = 0x0020_0000;
 
     /// Axis ids for `InputKind::GamepadAxis`.
     pub const AXIS_LS_X: u32 = 0;

crates/punktfunk-core/src/quic.rs

@@ -1218,6 +1218,7 @@ pub fn decode_mic_datagram(b: &[u8]) -> Option<(u32, u64, &[u8])> {
 
 const RICH_TOUCHPAD: u8 = 0x01;
 const RICH_MOTION: u8 = 0x02;
+const RICH_TOUCHPAD_EX: u8 = 0x03;
 
 /// A rich client→host controller input beyond the fixed [`InputEvent`](crate::input::InputEvent):
 /// the DualSense touchpad and motion sensors. `pad` is the gamepad index. Wire form is
@@ -1241,6 +1242,22 @@ pub enum RichInput {
         gyro: [i16; 3],
         accel: [i16; 3],
     },
+    /// A richer trackpad contact that also identifies *which* physical pad (Steam Controller / Deck
+    /// have two), carries a separate click vs touch state, and a pressure reading. `surface`:
+    /// `0` = the single / DualSense touchpad, `1` = the Steam left pad, `2` = the Steam right pad.
+    /// Coordinates are **signed** (centred at 0), matching the real Steam report; `pressure` is `0`
+    /// for a surface with no force sensor. New clients send this for every touch surface; the host
+    /// decodes both `Touchpad` (`0x01`) and `TouchpadEx` (`0x03`) indefinitely.
+    TouchpadEx {
+        pad: u8,
+        surface: u8,
+        finger: u8,
+        touch: bool,
+        click: bool,
+        x: i16,
+        y: i16,
+        pressure: u16,
+    },
 }
 
 impl RichInput {
@@ -1264,6 +1281,22 @@ impl RichInput {
                     out.extend_from_slice(&v.to_le_bytes());
                 }
             }
+            RichInput::TouchpadEx {
+                pad,
+                surface,
+                finger,
+                touch,
+                click,
+                x,
+                y,
+                pressure,
+            } => {
+                let state = (touch as u8) | ((click as u8) << 1);
+                out.extend_from_slice(&[RICH_TOUCHPAD_EX, pad, surface, finger, state]);
+                out.extend_from_slice(&x.to_le_bytes());
+                out.extend_from_slice(&y.to_le_bytes());
+                out.extend_from_slice(&pressure.to_le_bytes());
+            }
         }
         out
     }
@@ -1288,6 +1321,16 @@ impl RichInput {
                     accel: [i16at(9), i16at(11), i16at(13)],
                 })
             }
+            RICH_TOUCHPAD_EX if b.len() >= 12 => Some(RichInput::TouchpadEx {
+                pad: b[2],
+                surface: b[3],
+                finger: b[4],
+                touch: b[5] & 0x01 != 0,
+                click: b[5] & 0x02 != 0,
+                x: i16::from_le_bytes([b[6], b[7]]),
+                y: i16::from_le_bytes([b[8], b[9]]),
+                pressure: u16::from_le_bytes([b[10], b[11]]),
+            }),
             _ => None,
         }
     }
@@ -1296,6 +1339,7 @@ impl RichInput {
 const HIDOUT_LED: u8 = 0x01;
 const HIDOUT_PLAYER_LEDS: u8 = 0x02;
 const HIDOUT_TRIGGER: u8 = 0x03;
+const HIDOUT_TRACKPAD_HAPTIC: u8 = 0x04;
 
 /// DualSense feedback flowing host → client (what a game wrote to the host's virtual pad).
 /// Wire form `[0xCD][kind][pad][fields…]`. The rich analog of the fixed rumble datagram;
@@ -1309,6 +1353,16 @@ pub enum HidOutput {
     /// One adaptive-trigger effect: `which` 0 = L2, 1 = R2; `effect` is the raw DualSense
     /// trigger parameter block (mode + params) for the client to replay on a real controller.
     Trigger { pad: u8, which: u8, effect: Vec<u8> },
+    /// A trackpad haptic pulse for a Steam Controller's voice-coil actuators (its only "rumble").
+    /// `side` 0 = right pad, 1 = left pad; `amplitude` + `period` (µs off-time) + `count` (pulses)
+    /// synthesize a buzz. A client without trackpad coils drops it (or maps it to ordinary rumble).
+    TrackpadHaptic {
+        pad: u8,
+        side: u8,
+        amplitude: u16,
+        period: u16,
+        count: u16,
+    },
 }
 
 impl HidOutput {
@@ -1325,6 +1379,18 @@ impl HidOutput {
                 out.extend_from_slice(&[HIDOUT_TRIGGER, *pad, *which]);
                 out.extend_from_slice(effect);
             }
+            HidOutput::TrackpadHaptic {
+                pad,
+                side,
+                amplitude,
+                period,
+                count,
+            } => {
+                out.extend_from_slice(&[HIDOUT_TRACKPAD_HAPTIC, *pad, *side]);
+                out.extend_from_slice(&amplitude.to_le_bytes());
+                out.extend_from_slice(&period.to_le_bytes());
+                out.extend_from_slice(&count.to_le_bytes());
+            }
         }
         out
     }
@@ -1349,6 +1415,13 @@ impl HidOutput {
                 which: b[3],
                 effect: b[4..].to_vec(),
             }),
+            HIDOUT_TRACKPAD_HAPTIC if b.len() >= 10 => Some(HidOutput::TrackpadHaptic {
+                pad: b[2],
+                side: b[3],
+                amplitude: u16::from_le_bytes([b[4], b[5]]),
+                period: u16::from_le_bytes([b[6], b[7]]),
+                count: u16::from_le_bytes([b[8], b[9]]),
+            }),
             _ => None,
         }
     }
@@ -2486,6 +2559,16 @@ mod tests {
                 gyro: [-100, 200, -300],
                 accel: [16384, -8192, 1],
             },
+            RichInput::TouchpadEx {
+                pad: 2,
+                surface: 1,
+                finger: 1,
+                touch: true,
+                click: false,
+                x: -12345,
+                y: 30000,
+                pressure: 4000,
+            },
         ] {
             let d = ev.encode();
             assert_eq!(d[0], RICH_INPUT_MAGIC);
@@ -2494,7 +2577,8 @@ mod tests {
         // Disjoint from the fixed input datagram (0xC8); unknown kind + truncation → None.
         assert!(RichInput::decode(&[crate::input::INPUT_MAGIC; 18]).is_none());
         assert!(RichInput::decode(&[RICH_INPUT_MAGIC, 0x7F]).is_none()); // unknown kind
-        assert!(RichInput::decode(&[RICH_INPUT_MAGIC, RICH_TOUCHPAD, 0]).is_none());
+        assert!(RichInput::decode(&[RICH_INPUT_MAGIC, RICH_TOUCHPAD, 0]).is_none()); // short
+        assert!(RichInput::decode(&[RICH_INPUT_MAGIC, RICH_TOUCHPAD_EX, 0, 0, 0, 0]).is_none());
         // short
     }
 
@@ -2516,6 +2600,13 @@ mod tests {
                 which: 1,
                 effect: vec![0x26, 0x90, 0xA0, 0xFF, 0x00, 0x00],
             },
+            HidOutput::TrackpadHaptic {
+                pad: 0,
+                side: 1,
+                amplitude: 0x1234,
+                period: 0x5678,
+                count: 9,
+            },
         ];
         for ev in &cases {
             let d = ev.encode();

crates/punktfunk-host/Cargo.toml

@@ -89,6 +89,9 @@ tokio = { version = "1", features = ["rt", "rt-multi-thread", "net", "time"] }
 wayland-client = "0.31"
 wayland-protocols-wlr = { version = "0.3", features = ["client"] }
 wayland-protocols-misc = { version = "0.3", features = ["client"] }
+# `xdg-output` (zxdg_output_v1): the per-output *logical* geometry (post-scale size + global
+# position), used by the KWin fake_input backend to map absolute coordinates under display scaling.
+wayland-protocols = { version = "0.32", features = ["client"] }
 # Codegen for KDE's `zkde_screencast_unstable_v1` (vendored in `protocols/`): create a KWin
 # virtual output sized to the client's resolution and get its PipeWire node (KRdp's path).
 # `wayland-backend` is referenced by the generated interface tables.
@@ -119,6 +122,10 @@ ash = "0.38"
 # `libcuda.so.1` is dlopen'd at runtime (NOT link-time) so one Linux binary runs on NVIDIA
 # (zero-copy via CUDA) AND on AMD/Intel (VAAPI, no NVIDIA driver present) — see `zerocopy::cuda`.
 libloading = "0.8"
+# Vendored + trimmed `usbip` server core (no libusb) — presents a virtual Steam Deck over USB/IP
+# so the local `vhci_hcd` attaches it: the shippable, Secure-Boot-clean, Steam-Input-promotable
+# virtual-Deck transport on non-SteamOS hosts (`inject/linux/steam_usbip.rs`). See the crate's NOTICE.
+usbip-sim = { path = "vendor/usbip-sim" }
 
 [target.'cfg(target_os = "windows")'.dependencies]
 # Windows host backends. `windows` covers the Win32/CCD APIs the SudoVDA virtual-display backend
@@ -175,6 +182,9 @@ windows = { version = "0.62", features = [
     # Windows service supervisor (src/service.rs): a kill-on-close job object so a service crash never
     # orphans the SYSTEM host it launched into the interactive session.
     "Win32_System_JobObjects",
+    # CoCreateInstance(PolicyConfigClient) — set the default audio playback/recording endpoints via the
+    # undocumented IPolicyConfig (audio/windows/audio_control.rs) so mic + desktop audio auto-wire.
+    "Win32_System_Com",
 ] }
 # The SCM plumbing for the `service` subcommand (define_windows_service! / dispatcher / control
 # handler / ServiceManager install). Wraps the Win32 service API; the supervision loop itself uses

crates/punktfunk-host/src/audio.rs

@@ -42,6 +42,7 @@ pub fn open_audio_capture(channels: u32) -> Result<Box<dyn AudioCapturer>> {
 
 #[cfg(target_os = "windows")]
 pub fn open_audio_capture(channels: u32) -> Result<Box<dyn AudioCapturer>> {
+    audio_control::ensure_wired_once();
     wasapi_cap::WasapiLoopbackCapturer::open(channels)
         .map(|c| Box::new(c) as Box<dyn AudioCapturer>)
 }
@@ -77,6 +78,7 @@ pub fn open_virtual_mic(channels: u32) -> Result<Box<dyn VirtualMic>> {
 
 #[cfg(target_os = "windows")]
 pub fn open_virtual_mic(channels: u32) -> Result<Box<dyn VirtualMic>> {
+    audio_control::ensure_wired_once();
     wasapi_mic::WasapiVirtualMic::open(channels).map(|m| Box::new(m) as Box<dyn VirtualMic>)
 }
 
@@ -85,6 +87,9 @@ pub fn open_virtual_mic(_channels: u32) -> Result<Box<dyn VirtualMic>> {
     anyhow::bail!("virtual mic requires Linux + PipeWire or Windows + a virtual audio device")
 }
 
+#[cfg(target_os = "windows")]
+#[path = "audio/windows/audio_control.rs"]
+mod audio_control;
 #[cfg(target_os = "linux")]
 mod linux;
 #[cfg(target_os = "windows")]

crates/punktfunk-host/src/audio/windows/audio_control.rs

@@ -0,0 +1,227 @@
+//! Windows audio device auto-wiring — production mic + desktop-audio passthrough with zero manual
+//! setup.
+//!
+//! A headless host has no real audio output, so BOTH the desktop-audio loopback ([`super::wasapi_cap`])
+//! and the virtual mic ([`super::wasapi_mic`]) must run on VIRTUAL audio cables — and on DIFFERENT
+//! ones, or the loopback re-captures the injected mic (an infinite echo). The installer bundles
+//! VB-Audio Virtual Cable (the mic target: its "CABLE Input" render endpoint → "CABLE Output" capture)
+//! and the host auto-installs the Steam Streaming pair (a loopback-capable render). This module wires
+//! them up at startup so no manual Sound-settings fiddling is ever needed:
+//!
+//! * default **PLAYBACK**  → a loopback-capable render that is NOT the mic cable (a real output device
+//!   if one exists, else the Steam Streaming Microphone; **never** the Steam Streaming Speakers, whose
+//!   loopback is silent — validated live). This is the endpoint [`super::wasapi_cap`] loopback-captures
+//!   for desktop audio.
+//! * default **RECORDING** → the virtual mic's capture endpoint (VB-Cable "CABLE Output") so host apps
+//!   record the client's mic by default.
+//!
+//! [`super::wasapi_mic::find_device`] then resolves the mic INJECT target to "CABLE Input" — a render
+//! candidate that is NOT the default playback — guaranteeing loopback ≠ mic, so there is no echo.
+//!
+//! Setting a default endpoint uses the undocumented `IPolicyConfig` COM interface (the only way to set
+//! a default device programmatically — neither the `windows` nor `wasapi` crate exposes it; it is the
+//! same call `mmsys.cpl` makes). Opt out with `PUNKTFUNK_KEEP_DEFAULT` to leave the user's chosen
+//! defaults untouched.
+
+// Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it.
+#![deny(clippy::undocumented_unsafe_blocks)]
+
+use anyhow::{anyhow, bail, Result};
+use std::ffi::c_void;
+use std::sync::Once;
+use wasapi::Direction;
+
+/// Run the audio device auto-wiring exactly once per process, before the first capturer/mic opens.
+/// Blocks until done so the default playback is set before the loopback captures it. Best-effort:
+/// every failure is logged, never fatal (the host then falls back to whatever the current defaults
+/// are — exactly the pre-wiring behaviour).
+pub(crate) fn ensure_wired_once() {
+    static WIRED: Once = Once::new();
+    WIRED.call_once(|| {
+        if std::env::var_os("PUNKTFUNK_KEEP_DEFAULT").is_some() {
+            tracing::info!("PUNKTFUNK_KEEP_DEFAULT set — leaving the audio default devices untouched");
+            return;
+        }
+        // Run on a dedicated COM-MTA thread so we never collide with the caller's apartment mode
+        // (the capture/mic threads each initialize their own COM separately).
+        let handle = std::thread::Builder::new()
+            .name("pf-audio-wiring".into())
+            .spawn(|| {
+                if wasapi::initialize_mta().ok().is_err() {
+                    tracing::warn!("audio wiring: COM init (MTA) failed — skipping");
+                    return;
+                }
+                if let Err(e) = ensure_audio_wiring() {
+                    tracing::warn!(error = %format!("{e:#}"),
+                        "audio auto-wiring failed — mic/desktop audio may need manual device defaults");
+                }
+            });
+        if let Ok(h) = handle {
+            let _ = h.join();
+        }
+    });
+}
+
+/// `(friendly_name, endpoint_id)` for every ACTIVE endpoint in direction `dir`.
+fn list_endpoints(dir: Direction) -> Vec<(String, String)> {
+    let mut out = Vec::new();
+    let Ok(en) = wasapi::DeviceEnumerator::new() else {
+        return out;
+    };
+    let Ok(coll) = en.get_device_collection(&dir) else {
+        return out;
+    };
+    let Ok(n) = coll.get_nbr_devices() else {
+        return out;
+    };
+    for i in 0..n {
+        if let Ok(dev) = coll.get_device_at_index(i) {
+            let id = dev.get_id().unwrap_or_default();
+            if id.is_empty() {
+                continue;
+            }
+            out.push((dev.get_friendlyname().unwrap_or_default(), id));
+        }
+    }
+    out
+}
+
+/// Pick the loopback + mic-capture devices and set them as the default playback/recording.
+fn ensure_audio_wiring() -> Result<()> {
+    let renders = list_endpoints(Direction::Render);
+    let captures = list_endpoints(Direction::Capture);
+    if renders.is_empty() {
+        bail!("no active render endpoints to wire");
+    }
+
+    // A render is unusable as the desktop-audio loopback if it is a VB-Cable endpoint (reserved for
+    // the mic inject) or the Steam Streaming Speakers (its loopback is silent — validated live).
+    let excluded_loopback =
+        |ln: &str| ln.contains("cable") || ln.contains("steam streaming speakers");
+    // "virtual-ish" = a known virtual cable; a render WITHOUT these markers is a real output device,
+    // the best loopback source (apps render there and the operator can also hear it).
+    let virtualish = |ln: &str| {
+        ln.contains("virtual")
+            || ln.contains("cable")
+            || ln.contains("steam streaming")
+            || ln.contains("voicemeeter")
+    };
+    let loopback = renders
+        .iter()
+        .find(|(n, _)| {
+            let ln = n.to_lowercase();
+            !excluded_loopback(&ln) && !virtualish(&ln)
+        })
+        .or_else(|| {
+            renders
+                .iter()
+                .find(|(n, _)| n.to_lowercase().contains("steam streaming microphone"))
+        })
+        .or_else(|| {
+            renders
+                .iter()
+                .find(|(n, _)| !excluded_loopback(&n.to_lowercase()))
+        });
+
+    // The virtual mic's CAPTURE endpoint host apps record from — VB-Cable "CABLE Output" preferred.
+    let mic_capture = captures
+        .iter()
+        .find(|(n, _)| n.to_lowercase().contains("cable output"))
+        .or_else(|| {
+            captures
+                .iter()
+                .find(|(n, _)| n.to_lowercase().contains("steam streaming microphone"))
+        })
+        .or_else(|| {
+            captures.iter().find(|(n, _)| {
+                let ln = n.to_lowercase();
+                ln.contains("voicemeeter") || ln.contains("virtual")
+            })
+        });
+
+    match loopback {
+        Some((name, id)) => match set_default_endpoint(id) {
+            Ok(()) => tracing::info!(device = %name,
+                "audio wiring: default playback = desktop-audio loopback source"),
+            Err(e) => tracing::warn!(device = %name, error = %format!("{e:#}"),
+                "audio wiring: failed to set the default playback device"),
+        },
+        None => {
+            tracing::warn!("audio wiring: no usable desktop-audio loopback render endpoint found")
+        }
+    }
+    if let Some((name, id)) = mic_capture {
+        match set_default_endpoint(id) {
+            Ok(()) => tracing::info!(device = %name,
+                "audio wiring: default recording = virtual mic (apps record the client's mic)"),
+            Err(e) => tracing::warn!(device = %name, error = %format!("{e:#}"),
+                "audio wiring: failed to set the default recording device"),
+        }
+    }
+    Ok(())
+}
+
+// --- IPolicyConfig (undocumented): set a default audio endpoint by id, for all three roles. ---
+
+/// The `IPolicyConfig` vtable. Only `SetDefaultEndpoint` is called; the 10 methods between `Release`
+/// and it (`GetMixFormat` … `SetPropertyValue`) are placeholders so the slot offset is correct.
+#[repr(C)]
+struct IPolicyConfigVtbl {
+    query_interface: unsafe extern "system" fn(
+        *mut c_void,
+        *const windows::core::GUID,
+        *mut *mut c_void,
+    ) -> windows::core::HRESULT,
+    add_ref: unsafe extern "system" fn(*mut c_void) -> u32,
+    release: unsafe extern "system" fn(*mut c_void) -> u32,
+    _reserved: [*const c_void; 10],
+    set_default_endpoint: unsafe extern "system" fn(
+        *mut c_void,
+        windows::core::PCWSTR,
+        u32,
+    ) -> windows::core::HRESULT,
+    // SetEndpointVisibility follows — unused.
+}
+
+/// Set `device_id` as the default audio endpoint for eConsole/eMultimedia/eCommunications via the
+/// undocumented `IPolicyConfig::SetDefaultEndpoint` (the call `mmsys.cpl` makes). Errs if any role
+/// fails.
+fn set_default_endpoint(device_id: &str) -> Result<()> {
+    use windows::core::{IUnknown, Interface, GUID, PCWSTR};
+    use windows::Win32::System::Com::{CoCreateInstance, CLSCTX_ALL};
+
+    // PolicyConfigClient coclass + IPolicyConfig (Win7+) IID.
+    const CLSID_POLICY_CONFIG: GUID = GUID::from_u128(0x870af99c_171d_4f9e_af0d_e63df40c2bc9);
+    const IID_IPOLICY_CONFIG: GUID = GUID::from_u128(0xf8679f50_850a_41cf_9c72_430f290290c8);
+
+    let wide: Vec<u16> = device_id.encode_utf16().chain(std::iter::once(0)).collect();
+
+    // SAFETY: CoCreateInstance with a valid CLSID returns an owned, refcounted IUnknown. We QI it for
+    // IPolicyConfig; on success (HRESULT ok + non-null pointer) we invoke its SetDefaultEndpoint slot
+    // through the documented vtable layout (3 IUnknown + 10 placeholder methods precede it) with a
+    // NUL-terminated UTF-16 id and an in-range ERole (0..=2), then Release the QI'd pointer. Every
+    // pointer is checked non-null before deref; `unk` is Released by its Drop on scope exit.
+    unsafe {
+        let unk: IUnknown = CoCreateInstance(&CLSID_POLICY_CONFIG, None, CLSCTX_ALL)
+            .map_err(|e| anyhow!("CoCreateInstance(PolicyConfig): {e}"))?;
+        let mut raw: *mut c_void = std::ptr::null_mut();
+        unk.query(&IID_IPOLICY_CONFIG, &mut raw)
+            .ok()
+            .map_err(|e| anyhow!("QueryInterface(IPolicyConfig): {e}"))?;
+        if raw.is_null() {
+            bail!("IPolicyConfig QueryInterface returned null");
+        }
+        let vtbl = *(raw as *const *const IPolicyConfigVtbl);
+        let mut result = Ok(());
+        for role in 0u32..=2 {
+            let hr = ((*vtbl).set_default_endpoint)(raw, PCWSTR(wide.as_ptr()), role);
+            if hr.is_err() {
+                result = hr
+                    .ok()
+                    .map_err(|e| anyhow!("SetDefaultEndpoint(role {role}): {e}"));
+            }
+        }
+        ((*vtbl).release)(raw);
+        result
+    }
+}

crates/punktfunk-host/src/audio/windows/wasapi_mic.rs

@@ -4,10 +4,12 @@
 //! **capture** endpoint then surfaces as a microphone that host apps can record from.
 //!
 //! Target device, by friendly-name substring (first match wins; override with `PUNKTFUNK_MIC_DEVICE`):
-//! "Steam Streaming Microphone" (ships with Steam Remote Play — exactly this purpose), VB-Audio
+//! VB-Audio "CABLE Input" (bundled by the installer — the preferred, dedicated mic target), the
-//! "CABLE Input", VoiceMeeter, or anything with "virtual" in the name. If none is present we
+//! "Steam Streaming Microphone", VoiceMeeter, or anything with "virtual" in the name.
-//! auto-install the Steam Streaming audio pair (see [`install_steam_audio_pair`]); failing that we
+//! [`super::audio_control`] sets the default playback to a DIFFERENT loopback-capable device so the
-//! return an error with install guidance and the host runs without mic passthrough.
+//! chosen mic is never the endpoint the loopback captures. If no candidate is present we auto-install
+//! the Steam Streaming audio pair (see [`install_steam_audio_pair`]); failing that we return an error
+//! with install guidance and the host runs without mic passthrough.
 //!
 //! **Anti-echo guard (the whole point of this being non-trivial).** The desktop-audio plane
 //! ([`super::wasapi_cap`]) loopback-captures the **default render endpoint**. WASAPI loopback
@@ -45,8 +47,8 @@ const MAX_QUEUE_BYTES: usize = (SAMPLE_RATE as usize * 80 / 1000) * BLOCK_ALIGN;
 /// Render-endpoint friendly-name substrings (lowercased) we can write into so the device's capture
 /// endpoint becomes a host mic. Ordered by preference.
 const CANDIDATES: &[&str] = &[
+    "cable input", // VB-Audio Virtual Cable — bundled by the installer; the preferred dedicated mic target
     "steam streaming microphone",
-    "cable input",
     "voicemeeter input",
     "voicemeeter aux input",
     "virtual",

crates/punktfunk-host/src/capture.rs

@@ -59,7 +59,7 @@ pub struct OutputFormat {
     /// Produce GPU-resident D3D11 frames (zero-copy for a GPU encoder — NVENC/AMF/QSV) rather than CPU
     /// staging. `false` **only** for the GPU-less software encoder.
     pub gpu: bool,
-    /// HDR: the capturer converts to 10-bit (IDD-push FP16 → `Rgb10a2`; the DDA secure-desktop HDR hint).
+    /// HDR: the capturer converts to 10-bit (IDD-push FP16 → `P010`, or `Rgb10a2` for a 4:4:4 source).
     /// `false` = 8-bit SDR.
     pub hdr: bool,
     /// Full-chroma 4:4:4 session: the capturer must keep full chroma — deliver packed **RGB**
@@ -380,23 +380,12 @@ pub fn capture_virtual_output(
         .map(|c| Box::new(c) as Box<dyn Capturer>)
 }
 
-/// `PUNKTFUNK_NO_WGC=1` forces the pure single-process DDA (Desktop Duplication) path everywhere: it
-/// skips WGC in [`capture_virtual_output`] AND bypasses the two-process secure-desktop relay (so even a
-/// SYSTEM host captures in-process via DDA, the way Apollo does — one capturer for the normal AND the
-/// secure desktop). For bringing DDA up to parity / validating it on its own; all the WGC code stays
-/// compiled and comes back the moment the flag is unset.
-#[cfg(target_os = "windows")]
-pub(crate) fn wgc_disabled() -> bool {
-    crate::config::config().no_wgc
-}
-
 #[cfg(target_os = "windows")]
 pub fn capture_virtual_output(
     vout: crate::vdisplay::VirtualOutput,
     want: OutputFormat,
-    capture: crate::session_plan::CaptureBackend,
+    _capture: crate::session_plan::CaptureBackend,
 ) -> Result<Box<dyn Capturer>> {
-    use crate::session_plan::CaptureBackend;
     let target = vout.win_capture.clone().ok_or_else(|| {
         anyhow::anyhow!(
             "SudoVDA target not yet an active display (needs a WDDM GPU to activate it)"
@@ -404,97 +393,36 @@ pub fn capture_virtual_output(
     })?;
     let pref = vout.preferred_mode;
     let keep = vout.keepalive;
-    // Full-chroma 4:4:4 needs a full-chroma RGB source. The IDD-push and WGC paths emit subsampled
+    // IDD direct-push is the sole Windows capture path: consume frames straight from the pf-vdisplay
-    // NV12/P010 by default, which can't reconstruct 4:4:4; route a 4:4:4 session to DDA, which delivers
+    // driver's shared ring (in-process, Session 0 — it captures the secure desktop too; no Desktop
-    // RGB (Bgra) when its `chroma_444` flag is set. (IDD-push/WGC 4:4:4 capture is a follow-up.)
+    // Duplication, no WGC helper). A FRESH monitor + ring is created per session: a REUSED monitor's
-    if want.chroma_444 && capture != CaptureBackend::Dda {
+    // swap-chain dies after ~2 sessions and can't be revived. The ring is always FP16 when the display
-        tracing::info!("4:4:4 session — using DDA capture (RGB source) instead of {capture:?}");
+    // is HDR (the driver composes the IDD in FP16); `want.hdr` proactively enables advanced color and
-        return dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
+    // selects the per-frame conversion (FP16 → P010 vs BGRA → NV12). `IddPushCapturer` takes the
-            .map(|c| Box::new(c) as Box<dyn Capturer>);
+    // keepalive (it owns the virtual display). There is NO fallback (DDA + the WGC relay were removed):
-    }
+    // if it can't open or the driver doesn't attach, the session fails cleanly and the client reconnects.
-    // P2 direct frame push (kill DDA): consume frames straight from the pf-vdisplay driver's shared
+    idd_push::IddPushCapturer::open(target, pref, want.hdr, keep)
-    // ring — no Desktop Duplication, no win32u reparenting hook. Resolved once in the `SessionPlan`
-    // (was re-derived from `config().idd_push` here); `IddPush` takes the keepalive (owns the virtual
-    // display) so there's no fall-through.
-    if capture == CaptureBackend::IddPush {
-        // Recreate the monitor + ring per session (fix-teardown): a FRESH monitor reliably gets a
-        // working IddCx swap-chain, whereas a REUSED monitor's swap-chain dies after ~2 sessions and
-        // the host can't revive it. The driver's recreate crash (target id resolved to 0) is fixed by
-        // stamping target_id onto the monitor context. The ring is always FP16 (the driver composes
-        // the IDD in FP16); `want_hdr` selects the per-frame conversion (FP16 → Rgb10a2 vs Bgra).
-        // If IDD-push can't open OR the driver doesn't attach to the ring within a few seconds (e.g. a
-        // hybrid-GPU render mismatch), fall back to DDA so the session is NEVER left black (audit §5.1).
-        // `open()` hands the keepalive back on failure so DDA can take ownership of the virtual display.
-        match idd_push::IddPushCapturer::open(target.clone(), pref, want.hdr, keep) {
-            Ok(c) => return Ok(Box::new(c) as Box<dyn Capturer>),
-            Err((e, keep)) => {
-                tracing::warn!(
-                    error = %format!("{e:#}"),
-                    "IDD-push open/attach failed — falling back to DDA"
-                );
-                return dxgi::DuplCapturer::open(
-                    target,
-                    pref,
-                    keep,
-                    want.gpu,
-                    false,
-                    want.chroma_444,
-                )
-                .map(|c| Box::new(c) as Box<dyn Capturer>);
-            }
-        }
-    }
-    // WGC (Windows.Graphics.Capture) is the default: it captures the COMPOSED desktop including the
-    // overlay/independent-flip planes DXGI Desktop Duplication misses (the frozen-HDR-animation bug),
-    // and has no ACCESS_LOST-on-overlay churn. DDA stays available via PUNKTFUNK_CAPTURE=dda and is
-    // the secure-desktop (lock/UAC) fallback (WGC can't capture those). `keep` is moved into the
-    // chosen backend (it owns the SudoVDA keepalive), so there's no open-time auto-fallback. The
-    // backend choice (`dda`/`dxgi`/`PUNKTFUNK_NO_WGC` → DDA, else WGC) is now resolved once in the plan.
-    if capture == CaptureBackend::Dda {
-        return dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
-            .map(|c| Box::new(c) as Box<dyn Capturer>);
-    }
-    // WGC default, with a watchdog'd DDA fallback. WGC's Direct3D11CaptureFramePool::CreateFreeThreaded
-    // intermittently HANGS on the headless SudoVDA (IddCx) display — a blocking call we can't error out
-    // of in place. So run WGC open on a dedicated thread and bound it: if it doesn't finish in time
-    // (hang) or errors, fall back to the reliable DDA path so the session is NEVER left black. WGC,
-    // when it opens, captures the composed desktop (overlay/MPO-correct HDR — fixes frozen animations);
-    // DDA is the safety net (+ the secure-desktop path). The encode thread is set MTA so the WGC
-    // objects built on the watchdog thread (also MTA) are usable here; the keepalive is handed to WGC
-    // only on success, else to DDA. A hung watchdog thread is abandoned (holds no keepalive).
-    // SAFETY: `RoInitialize` is a combase FFI call that initializes the WinRT apartment for the calling
-    // thread. It takes the `RO_INIT_MULTITHREADED` enum by value and borrows no memory, so there is no
-    // pointer/lifetime/aliasing obligation; it is safe on any thread and idempotent — a second call on a
-    // thread already in a compatible apartment returns S_FALSE / RPC_E_CHANGED_MODE, which we discard.
-    // Runs on the encode thread that goes on to use the WGC (WinRT) objects built by the watchdog thread.
-    unsafe {
-        let _ = windows::Win32::System::WinRT::RoInitialize(
-            windows::Win32::System::WinRT::RO_INIT_MULTITHREADED,
-        );
-    }
-    let (tx, rx) = std::sync::mpsc::channel();
-    let t = target.clone();
-    let _ = std::thread::Builder::new()
-        .name("wgc-open".into())
-        .spawn(move || {
-            let _ = tx.send(wgc::WgcCapturer::open(t, pref));
-        });
-    match rx.recv_timeout(std::time::Duration::from_secs(5)) {
-        Ok(Ok(mut c)) => {
-            c.attach_keepalive(keep);
-            Ok(Box::new(c) as Box<dyn Capturer>)
-        }
-        Ok(Err(e)) => {
-            tracing::warn!(error = %format!("{e:#}"), "WGC open failed — falling back to DDA");
-            dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
         .map(|c| Box::new(c) as Box<dyn Capturer>)
-        }
+        .map_err(|(e, _keep)| e.context("IDD-push capture open (no fallback)"))
-        Err(_) => {
+}
-            tracing::warn!("WGC open timed out (CreateFreeThreaded hang on the virtual display) — falling back to DDA");
+
-            dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
+/// Whether the active capturer can deliver a full-chroma (RGB) source for a 4:4:4 HEVC encode. The
-                .map(|c| Box::new(c) as Box<dyn Capturer>)
+/// negotiator gates 4:4:4 on this so the host honestly downgrades to 4:2:0 when the capturer can only
-        }
+/// produce subsampled frames. Linux (the portal capturer feeding CPU RGB → `yuv444p`) can; the Windows
-    }
+/// IDD-push path delivers subsampled NV12/P010 today, so full-chroma capture there is a follow-up.
+#[cfg(target_os = "linux")]
+pub(crate) fn capturer_supports_444() -> bool {
+    true
+}
+#[cfg(target_os = "windows")]
+pub(crate) fn capturer_supports_444() -> bool {
+    // IDD-push 4:4:4 (full-chroma RGB from the FP16 ring) is the next step; until then the sole Windows
+    // capturer delivers subsampled NV12/P010 only, so the host honestly negotiates 4:2:0.
+    false
+}
+#[cfg(not(any(target_os = "linux", target_os = "windows")))]
+pub(crate) fn capturer_supports_444() -> bool {
+    false
 }
 
 #[cfg(not(any(target_os = "linux", target_os = "windows")))]
@@ -506,14 +434,9 @@ pub fn capture_virtual_output(
     anyhow::bail!("virtual-output capture requires Linux or Windows")
 }
 
-// Goal-1 stage 6: the Windows backends live under `capture/windows/`, the Linux one under `capture/linux/`
+// Goal-1 stage 6: the Windows backend lives under `capture/windows/`, the Linux one under `capture/linux/`
-// (`#[path]` keeps the module names flat, so every `crate::capture::*` path is unchanged).
+// (`#[path]` keeps the module names flat, so every `crate::capture::*` path is unchanged). Windows capture
-#[cfg(target_os = "windows")]
+// is IDD direct-push only — DXGI Desktop Duplication (DDA) and the WGC two-process relay were removed.
-#[path = "capture/windows/composed_flip.rs"]
-pub mod composed_flip;
-#[cfg(target_os = "windows")]
-#[path = "capture/windows/desktop_watch.rs"]
-pub mod desktop_watch;
 #[cfg(target_os = "windows")]
 #[path = "capture/windows/dxgi.rs"]
 pub mod dxgi;
@@ -522,9 +445,3 @@ pub mod dxgi;
 pub mod idd_push;
 #[cfg(target_os = "linux")]
 mod linux;
-#[cfg(target_os = "windows")]
-#[path = "capture/windows/wgc.rs"]
-pub mod wgc;
-#[cfg(target_os = "windows")]
-#[path = "capture/windows/wgc_relay.rs"]
-pub mod wgc_relay;

crates/punktfunk-host/src/capture/windows/composed_flip.rs

@@ -1,217 +0,0 @@
-//! Force-composed-flip overlay (Windows) — make the secure (Winlogon: UAC / lock / login) desktop
-//! capturable by Desktop Duplication.
-//!
-//! The secure desktop's dialog/wallpaper presents via **fullscreen independent-flip / MPO**: it scans
-//! out directly, bypassing DWM composition, so `IDXGIOutputDuplication::AcquireNextFrame` returns
-//! `DXGI_ERROR_ACCESS_LOST` (born-lost) — there is no composed frame to hand out (the client sees
-//! black). Independent-flip requires the presenting app to own the ENTIRE output: putting ANY other
-//! visible window on that output disqualifies it, forcing DWM to **composite**, which DDA can then
-//! capture. So we keep a tiny, click-through, near-invisible **topmost layered window** alive on the
-//! *current input desktop* (which is Winlogon while the secure desktop is up). On a desktop switch the
-//! window is orphaned, so a dedicated thread tracks the input desktop and recreates it there.
-//!
-//! This is the non-input alternative to "tap a key to wake the lock screen": it needs no SendInput and
-//! no system-wide registry change (it does NOT disable MPO globally — it only nudges OUR output to
-//! composed while a session is live). Effectiveness can be build/driver-dependent; gated by
-//! `PUNKTFUNK_FORCE_COMPOSED` (default ON; set =0 to disable).
-
-// Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
-#![deny(clippy::undocumented_unsafe_blocks)]
-
-use std::sync::atomic::{AtomicBool, Ordering};
-use std::sync::Arc;
-use windows::core::w;
-use windows::Win32::Foundation::{HWND, LPARAM, LRESULT, WPARAM};
-use windows::Win32::System::LibraryLoader::GetModuleHandleW;
-use windows::Win32::System::StationsAndDesktops::{
-    CloseDesktop, GetUserObjectInformationW, OpenInputDesktop, SetThreadDesktop,
-    DESKTOP_ACCESS_FLAGS, DESKTOP_CONTROL_FLAGS, UOI_NAME,
-};
-use windows::Win32::UI::WindowsAndMessaging::{
-    CreateWindowExW, DefWindowProcW, DestroyWindow, DispatchMessageW, PeekMessageW, RegisterClassW,
-    SetLayeredWindowAttributes, SetWindowPos, ShowWindow, TranslateMessage, HWND_TOPMOST,
-    LWA_ALPHA, MSG, PM_REMOVE, SWP_NOACTIVATE, SWP_NOMOVE, SWP_NOSIZE, SW_SHOWNOACTIVATE,
-    WNDCLASSW, WS_EX_LAYERED, WS_EX_NOACTIVATE, WS_EX_TOOLWINDOW, WS_EX_TOPMOST, WS_EX_TRANSPARENT,
-    WS_POPUP,
-};
-
-/// A running force-composed-flip overlay. Drop signals the thread to tear down its window + exit.
-pub struct ForceComposedFlip {
-    stop: Arc<AtomicBool>,
-}
-
-impl ForceComposedFlip {
-    /// Start the overlay (no-op + `None` if disabled via `PUNKTFUNK_FORCE_COMPOSED=0`).
-    pub fn start() -> Option<Self> {
-        if std::env::var("PUNKTFUNK_FORCE_COMPOSED").as_deref() == Ok("0") {
-            tracing::info!("force-composed-flip overlay disabled (PUNKTFUNK_FORCE_COMPOSED=0)");
-            return None;
-        }
-        let stop = Arc::new(AtomicBool::new(false));
-        let st = stop.clone();
-        std::thread::Builder::new()
-            .name("composed-flip".into())
-            // SAFETY: `run` is this module's `unsafe fn` (it owns a desktop+window lifecycle via Win32
-            // FFI); it takes ownership of `st` (the stop `Arc<AtomicBool>`) and has no caller-side memory
-            // precondition. It is designed to own its thread for its whole duration — exactly the
-            // dedicated `composed-flip` thread spawned here.
-            .spawn(move || unsafe { run(st) })
-            .ok()?;
-        tracing::info!("force-composed-flip overlay started (Winlogon-aware)");
-        Some(ForceComposedFlip { stop })
-    }
-}
-
-impl Drop for ForceComposedFlip {
-    fn drop(&mut self) {
-        self.stop.store(true, Ordering::Relaxed);
-    }
-}
-
-extern "system" fn wndproc(hwnd: HWND, msg: u32, wp: WPARAM, lp: LPARAM) -> LRESULT {
-    // SAFETY: this is the window procedure the OS invokes with the window's own `hwnd` and a real
-    // message `(msg, wp, lp)`. `DefWindowProcW` performs default processing for exactly those
-    // parameters (all passed straight through by value); it borrows no Rust memory and is synchronous.
-    unsafe { DefWindowProcW(hwnd, msg, wp, lp) }
-}
-
-/// Read the current input-desktop name (e.g. "Default" / "Winlogon"); `None` if it can't be read.
-unsafe fn input_desktop_name() -> Option<String> {
-    let desk = OpenInputDesktop(
-        DESKTOP_CONTROL_FLAGS(0),
-        false,
-        DESKTOP_ACCESS_FLAGS(0x0001),
-    )
-    .ok()?;
-    let mut buf = [0u16; 64];
-    let mut needed = 0u32;
-    let ok = GetUserObjectInformationW(
-        windows::Win32::Foundation::HANDLE(desk.0),
-        UOI_NAME,
-        Some(buf.as_mut_ptr() as *mut _),
-        (buf.len() * 2) as u32,
-        Some(&mut needed),
-    )
-    .is_ok();
-    let _ = CloseDesktop(desk);
-    if !ok {
-        return None;
-    }
-    Some(
-        String::from_utf16_lossy(&buf)
-            .trim_end_matches('\u{0}')
-            .to_string(),
-    )
-}
-
-/// Create the tiny topmost layered click-through window on the CURRENT thread's desktop. Caller must
-/// have `SetThreadDesktop`'d to the target input desktop first.
-unsafe fn make_overlay() -> Option<HWND> {
-    let hinst = GetModuleHandleW(None).ok()?;
-    let class = w!("PunktfunkComposedFlip");
-    // RegisterClassW is idempotent-ish: a second register for the same name fails harmlessly; we
-    // ignore the result and rely on the class existing. (One process, so it registers once.)
-    let wc = WNDCLASSW {
-        lpfnWndProc: Some(wndproc),
-        hInstance: hinst.into(),
-        lpszClassName: class,
-        ..Default::default()
-    };
-    let atom = RegisterClassW(&wc);
-    if atom == 0 {
-        let e = windows::Win32::Foundation::GetLastError();
-        // 1410 = ERROR_CLASS_ALREADY_EXISTS is fine (re-register after a desktop switch).
-        if e.0 != 1410 {
-            tracing::warn!(err = e.0, "force-composed-flip: RegisterClassW failed");
-        }
-    }
-    let hwnd = match CreateWindowExW(
-        WS_EX_LAYERED | WS_EX_TRANSPARENT | WS_EX_TOPMOST | WS_EX_NOACTIVATE | WS_EX_TOOLWINDOW,
-        class,
-        w!(""),
-        WS_POPUP,
-        0,
-        0,
-        1,
-        1,
-        None,
-        None,
-        Some(hinst.into()),
-        None,
-    ) {
-        Ok(h) => h,
-        Err(e) => {
-            let le = windows::Win32::Foundation::GetLastError();
-            tracing::warn!(err = %format!("{e:?}"), last = le.0,
-                "force-composed-flip: CreateWindowExW failed");
-            return None;
-        }
-    };
-    // alpha=1: technically visible (so it disqualifies independent-flip) but imperceptible.
-    let _ = SetLayeredWindowAttributes(hwnd, windows::Win32::Foundation::COLORREF(0), 1, LWA_ALPHA);
-    let _ = ShowWindow(hwnd, SW_SHOWNOACTIVATE);
-    let _ = SetWindowPos(
-        hwnd,
-        Some(HWND_TOPMOST),
-        0,
-        0,
-        0,
-        0,
-        SWP_NOMOVE | SWP_NOSIZE | SWP_NOACTIVATE,
-    );
-    Some(hwnd)
-}
-
-unsafe fn run(stop: Arc<AtomicBool>) {
-    let mut cur_desktop: Option<String> = None;
-    let mut hwnd: Option<HWND> = None;
-    let mut ticks: u32 = 0;
-    while !stop.load(Ordering::Relaxed) {
-        // Follow the input desktop: if it changed (Default↔Winlogon), re-attach this thread and
-        // recreate the window there (a window is bound to the desktop it was created on).
-        let name = input_desktop_name();
-        if name != cur_desktop {
-            if let Some(h) = hwnd.take() {
-                let _ = DestroyWindow(h);
-            }
-            if let Ok(desk) = OpenInputDesktop(
-                DESKTOP_CONTROL_FLAGS(0),
-                false,
-                DESKTOP_ACCESS_FLAGS(0x1000_0000), // GENERIC_ALL (incl. DESKTOP_CREATEWINDOW=0x0002)
-            ) {
-                if SetThreadDesktop(desk).is_ok() {
-                    hwnd = make_overlay();
-                    tracing::info!(desktop = ?name, created = hwnd.is_some(),
-                        "force-composed-flip: overlay (re)created on input desktop");
-                }
-                // Leak `desk` while it's the thread desktop (closing the current thread desktop is UB).
-            }
-            cur_desktop = name;
-        }
-        // Re-assert topmost periodically (other windows on the secure desktop can push us down) and
-        // pump our message queue so the window stays responsive/composited.
-        if let Some(h) = hwnd {
-            let _ = SetWindowPos(
-                h,
-                Some(HWND_TOPMOST),
-                0,
-                0,
-                0,
-                0,
-                SWP_NOMOVE | SWP_NOSIZE | SWP_NOACTIVATE,
-            );
-            let mut msg = MSG::default();
-            while PeekMessageW(&mut msg, Some(h), 0, 0, PM_REMOVE).as_bool() {
-                let _ = TranslateMessage(&msg);
-                DispatchMessageW(&msg);
-            }
-        }
-        ticks = ticks.wrapping_add(1);
-        let _ = ticks;
-        std::thread::sleep(std::time::Duration::from_millis(200));
-    }
-    if let Some(h) = hwnd.take() {
-        let _ = DestroyWindow(h);
-    }
-    tracing::info!("force-composed-flip overlay stopped");
-}

crates/punktfunk-host/src/capture/windows/desktop_watch.rs

@@ -1,144 +0,0 @@
-//! Input-desktop watcher (Windows) — the authoritative "normal vs secure desktop" signal for the
-//! two-process secure-desktop design (design/archive/windows-secure-desktop.md).
-//!
-//! Windows switches the *input desktop* to "Winlogon" (the secure desktop) for UAC elevation, the
-//! lock screen and the login screen, and back to "Default" for the normal session. WGC captures only
-//! the normal desktop; DDA-as-SYSTEM captures the secure one. A dedicated thread polls the input
-//! desktop's NAME (WTS session notifications miss UAC entirely, so the name is the reliable signal)
-//! and publishes it as an atomic the capture mux + input path read.
-
-// Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
-#![deny(clippy::undocumented_unsafe_blocks)]
-
-use std::sync::atomic::{AtomicBool, AtomicU8, Ordering};
-use std::sync::Arc;
-use std::time::Duration;
-use windows::Win32::Foundation::HANDLE;
-use windows::Win32::System::StationsAndDesktops::{
-    CloseDesktop, GetUserObjectInformationW, OpenInputDesktop, DESKTOP_ACCESS_FLAGS,
-    DESKTOP_CONTROL_FLAGS, UOI_NAME,
-};
-
-/// The normal interactive desktop ("Default") — WGC capture applies.
-pub const DESKTOP_NORMAL: u8 = 0;
-/// The secure desktop ("Winlogon": UAC / lock / login) — DDA-as-SYSTEM capture applies.
-pub const DESKTOP_SECURE: u8 = 1;
-
-/// Polls the input-desktop name on its own thread and publishes [`DESKTOP_NORMAL`]/[`DESKTOP_SECURE`].
-pub struct DesktopWatcher {
-    state: Arc<AtomicU8>,
-    stop: Arc<AtomicBool>,
-}
-
-impl DesktopWatcher {
-    pub fn start() -> Self {
-        // Compute the CURRENT desktop synchronously before returning, so the first reader (the capture
-        // mux) sees the real state immediately. Otherwise a session that begins already on the secure
-        // desktop (e.g. a reconnect to a locked box) would read DESKTOP_NORMAL for the first poll
-        // interval and relay one stale normal-desktop frame — the "flash of the login screen" bug.
-        // SAFETY: `is_secure_desktop` is this module's `unsafe fn` — unsafe only because it calls Win32
-        // desktop FFI (`OpenInputDesktop`/`GetUserObjectInformationW`/`CloseDesktop`), with no caller
-        // precondition; it opens, names, and closes the input-desktop handle internally and is safe to
-        // call from any thread (here, on the thread running `DesktopWatcher::start`).
-        let initial = if unsafe { is_secure_desktop() } {
-            DESKTOP_SECURE
-        } else {
-            DESKTOP_NORMAL
-        };
-        let state = Arc::new(AtomicU8::new(initial));
-        let stop = Arc::new(AtomicBool::new(false));
-        let s = state.clone();
-        let st = stop.clone();
-        let _ = std::thread::Builder::new()
-            .name("desktop-watch".into())
-            .spawn(move || {
-                // Debounce: only publish a change after the raw reading has been stable for several
-                // polls. The input desktop flaps Default↔Winlogon transiently during a lock/UAC
-                // transition; publishing every flap makes the capture mux thrash (rebuild storms).
-                const STABLE_POLLS: u32 = 4; // ~80ms
-                let mut published = initial;
-                let mut candidate = initial;
-                let mut stable = 0u32;
-                while !st.load(Ordering::Relaxed) {
-                    // SAFETY: same as in `start` — `is_secure_desktop` is self-contained Win32 desktop
-                    // FFI with no caller precondition, called here on the dedicated `desktop-watch`
-                    // polling thread.
-                    let v = if unsafe { is_secure_desktop() } {
-                        DESKTOP_SECURE
-                    } else {
-                        DESKTOP_NORMAL
-                    };
-                    if v == candidate {
-                        stable = stable.saturating_add(1);
-                    } else {
-                        candidate = v;
-                        stable = 1;
-                    }
-                    if stable >= STABLE_POLLS && candidate != published {
-                        s.store(candidate, Ordering::Release);
-                        published = candidate;
-                        tracing::info!(
-                            desktop = if candidate == DESKTOP_SECURE {
-                                "Winlogon(secure)"
-                            } else {
-                                "Default"
-                            },
-                            "input desktop changed (debounced)"
-                        );
-                    }
-                    std::thread::sleep(Duration::from_millis(20));
-                }
-            });
-        DesktopWatcher { state, stop }
-    }
-
-    /// The shared atomic ([`DESKTOP_NORMAL`]/[`DESKTOP_SECURE`]) for the capture mux to read.
-    pub fn state(&self) -> Arc<AtomicU8> {
-        self.state.clone()
-    }
-
-    /// True when the secure (Winlogon) desktop is the input desktop right now.
-    pub fn is_secure(&self) -> bool {
-        self.state.load(Ordering::Acquire) == DESKTOP_SECURE
-    }
-}
-
-impl Drop for DesktopWatcher {
-    fn drop(&mut self) {
-        self.stop.store(true, Ordering::Relaxed);
-    }
-}
-
-/// True if the current input desktop is "Winlogon" (the secure desktop). Best-effort: if the desktop
-/// can't be opened or named, report not-secure (the safe default — keep WGC/normal capture).
-pub(crate) unsafe fn is_secure_desktop() -> bool {
-    let desk = match OpenInputDesktop(
-        DESKTOP_CONTROL_FLAGS(0),
-        false,
-        DESKTOP_ACCESS_FLAGS(DESKTOP_READOBJECTS),
-    ) {
-        Ok(d) => d,
-        Err(_) => return false,
-    };
-    let mut buf = [0u16; 64];
-    let mut needed = 0u32;
-    let ok = GetUserObjectInformationW(
-        HANDLE(desk.0),
-        UOI_NAME,
-        Some(buf.as_mut_ptr() as *mut _),
-        (buf.len() * 2) as u32,
-        Some(&mut needed),
-    )
-    .is_ok();
-    let _ = CloseDesktop(desk);
-    if !ok {
-        return false;
-    }
-    let name = String::from_utf16_lossy(&buf);
-    name.trim_end_matches('\u{0}')
-        .eq_ignore_ascii_case("Winlogon")
-}
-
-/// `DESKTOP_READOBJECTS` access right (the windows crate exposes it as a typed flag; we need the raw
-/// bit for `OpenInputDesktop`'s access mask).
-const DESKTOP_READOBJECTS: u32 = 0x0001;

crates/punktfunk-host/src/capture/windows/idd_push.rs

@@ -1,8 +1,8 @@
-//! P2 direct frame push (kill DDA) — HOST side. The pf-vdisplay driver runs in a restricted WUDFHost
+//! P2 direct frame push (kill DDA) — HOST side. The pf-vdisplay driver's WUDFHost canNOT create named
-//! token that canNOT create named kernel objects, so — exactly like the gamepad UMDF drivers
+//! kernel objects, so — exactly like the gamepad UMDF drivers (`inject/dualsense_windows.rs`) — the
-//! (`inject/dualsense_windows.rs`) — the HOST (privileged) CREATES the shared header + frame-ready
+//! HOST (privileged) CREATES the shared header + frame-ready event + ring of keyed-mutex textures
-//! event + ring of keyed-mutex textures (`Global\` names, permissive `D:(A;;GA;;;WD)` SDDL) on the
+//! (`Global\` names, scoped `D:(A;;GA;;;SY)(A;;GA;;;LS)` to SYSTEM + the driver's LocalService host —
-//! discrete render GPU, and the driver only OPENS them and copies frames in. We then consume the ring
+//! see `shared_object_sa`) on the discrete render GPU, and the driver only OPENS them and copies frames in. We then consume the ring
 //! straight into the zero-copy NVENC path — no DXGI Desktop Duplication, no `win32u` hook. Gated by
 //! `PUNKTFUNK_IDD_PUSH`. Driver counterpart: `packaging/windows/drivers/pf-vdisplay/src/
 //! frame_transport.rs`. The shared `SharedHeader` layout, `MAGIC`/`VERSION`/`RING_LEN`, the
@@ -236,13 +236,17 @@ pub struct IddPushCapturer {
 // ownership to one thread at a time with NO concurrent access; we do not (and must not) claim `Sync`.
 unsafe impl Send for IddPushCapturer {}
 
-/// Build a permissive (Everyone:GenericAll) `SECURITY_ATTRIBUTES` so the restricted WUDFHost driver
+/// Build a `SECURITY_ATTRIBUTES` granting GENERIC_ALL to **SYSTEM** (the host creates+publishes the
-/// can OPEN the host-created objects — the same `D:(A;;GA;;;WD)` SDDL the gamepad shared section uses.
+/// shared event + texture ring) and **LocalService** (the account the pf_vdisplay WUDFHost runs under,
-/// The returned `psd` backing must outlive `sa`; both are dropped when the process exits.
+/// which consumes them) — `D:(A;;GA;;;SY)(A;;GA;;;LS)`, the same scoping as the gamepad section. The
-unsafe fn permissive_sa() -> Result<(SECURITY_ATTRIBUTES, PSECURITY_DESCRIPTOR)> {
+/// old SDDL granted **Everyone** (`WD`), which let any local user open the `Global\pfvd-*` objects and
+/// read captured screen frames (security-review 2026-06-28 #5). Verified on the RTX box (2026-06-29):
+/// the WUDFHost token is `S-1-5-19` (LocalService), SYSTEM integrity, zero restricted SIDs — so SY+LS
+/// suffices for the driver and excludes normal user processes. `psd` must outlive `sa`.
+unsafe fn shared_object_sa() -> Result<(SECURITY_ATTRIBUTES, PSECURITY_DESCRIPTOR)> {
     let mut psd = PSECURITY_DESCRIPTOR::default();
     ConvertStringSecurityDescriptorToSecurityDescriptorW(
-        w!("D:(A;;GA;;;WD)"),
+        w!("D:(A;;GA;;;SY)(A;;GA;;;LS)"),
         SDDL_REVISION_1,
         &mut psd,
         None,
@@ -269,7 +273,7 @@ impl IddPushCapturer {
         h: u32,
         format: DXGI_FORMAT,
     ) -> Result<Vec<HostSlot>> {
-        let (sa, _psd) = permissive_sa()?;
+        let (sa, _psd) = shared_object_sa()?;
         let mut slots = Vec::new();
         for k in 0..RING_LEN {
             let desc = D3D11_TEXTURE2D_DESC {
@@ -375,7 +379,7 @@ impl IddPushCapturer {
         // SAFETY: one block over the whole ring setup; every operation in it is sound:
         // - `set_advanced_color`/`advanced_color_enabled` are `unsafe fn`s taking only a copy of the plain
         //   `u32` target id; they read/flip CCD display config and return owned values, borrowing nothing.
-        // - `CreateDXGIFactory1`, `EnumAdapterByLuid`, `make_device`, `permissive_sa`, `CreateFileMappingW`,
+        // - `CreateDXGIFactory1`, `EnumAdapterByLuid`, `make_device`, `shared_object_sa`, `CreateFileMappingW`,
         //   `MapViewOfFile`, `CreateEventW`, and `create_ring_slots` are all `?`-checked, so every returned
         //   interface/handle/view is non-error before use; `&sa`/`&adapter`/`&device`/the `&HSTRING` names
         //   are live borrows that outlive each synchronous call, and `sa.lpSecurityDescriptor` stays valid
@@ -421,7 +425,7 @@ impl IddPushCapturer {
                 .context("EnumAdapterByLuid(render adapter) for IDD push")?;
             let (device, context) = make_device(&adapter).context("make_device for IDD push")?;
 
-            let (sa, _psd) = permissive_sa()?;
+            let (sa, _psd) = shared_object_sa()?;
             let bytes = std::mem::size_of::<SharedHeader>().max(64);
 
             // Header.

crates/punktfunk-host/src/capture/windows/wgc.rs

@@ -1,816 +0,0 @@
-//! Windows.Graphics.Capture (WGC) capture backend — the HDR/animation-correct path.
-//!
-//! Why WGC over DXGI Desktop Duplication: DDA duplicates only the DWM-composed primary surface, so
-//! HDR desktop animations the OS routes onto hardware overlay / independent-flip / MPO planes (Start
-//! menu, Win11 Mica/acrylic, window resize) never enter the surface DDA reads — the stream shows a
-//! frozen desktop ("broken HDR animations"). Engaging WGC capture pulls that content back through DWM
-//! composition, so the surface WGC hands back contains the animations. WGC also has no
-//! ACCESS_LOST-on-overlay-flip churn.
-//!
-//! It reuses the rest of the pipeline UNCHANGED: the frame's GPU texture (the OS already composited
-//! the cursor into it — `IsCursorCaptureEnabled(true)`) goes through the same scRGB→BT.2020-PQ shader
-//! ([`super::dxgi::HdrConverter`]) into a host-owned `R10G10B10A2` texture (HDR) or is copied into a
-//! BGRA texture (SDR), which is handed to NVENC zero-copy (registered by pointer, encoded in place).
-//! Shares the D3D11 device with NVENC via `FramePayload::D3d11`.
-//!
-//! Limitation: WGC cannot capture the secure desktop (lock / UAC / login) — the caller falls back to
-//! the DDA backend ([`super::dxgi::DuplCapturer`]) for those (see capture.rs).
-
-// Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
-#![deny(clippy::undocumented_unsafe_blocks)]
-
-use super::dxgi::{
-    find_output, hdr_shader_p010_enabled, make_device, nudge_cursor_onto, D3d11Frame, HdrConverter,
-    HdrP010Converter, VideoConverter, WinCaptureTarget,
-};
-use super::{CapturedFrame, Capturer, FramePayload, PixelFormat};
-use anyhow::{bail, Context, Result};
-use std::collections::VecDeque;
-use std::sync::atomic::{AtomicU64, Ordering};
-use std::sync::{Arc, Condvar, Mutex};
-use std::time::{Duration, Instant};
-use windows::core::{IInspectable, Interface};
-use windows::Foundation::{TimeSpan, TypedEventHandler};
-use windows::Graphics::Capture::{
-    Direct3D11CaptureFrame, Direct3D11CaptureFramePool, GraphicsCaptureItem, GraphicsCaptureSession,
-};
-use windows::Graphics::DirectX::DirectXPixelFormat;
-use windows::Win32::Foundation::{CloseHandle, HANDLE};
-use windows::Win32::Graphics::Direct3D11::{
-    ID3D11Device, ID3D11DeviceContext, ID3D11RenderTargetView, ID3D11ShaderResourceView,
-    ID3D11Texture2D, D3D11_BIND_RENDER_TARGET, D3D11_BIND_SHADER_RESOURCE, D3D11_TEXTURE2D_DESC,
-    D3D11_USAGE_DEFAULT,
-};
-use windows::Win32::Graphics::Dxgi::Common::{
-    DXGI_COLOR_SPACE_RGB_FULL_G2084_NONE_P2020, DXGI_FORMAT_R10G10B10A2_UNORM,
-    DXGI_FORMAT_R16G16B16A16_FLOAT, DXGI_SAMPLE_DESC,
-};
-use windows::Win32::Graphics::Dxgi::{IDXGIDevice, IDXGIOutput6};
-use windows::Win32::Security::{ImpersonateLoggedOnUser, RevertToSelf};
-use windows::Win32::System::RemoteDesktop::{WTSGetActiveConsoleSessionId, WTSQueryUserToken};
-use windows::Win32::System::WinRT::Direct3D11::{
-    CreateDirect3D11DeviceFromDXGIDevice, IDirect3DDxgiInterfaceAccess,
-};
-use windows::Win32::System::WinRT::Graphics::Capture::IGraphicsCaptureItemInterop;
-use windows::Win32::System::WinRT::{RoInitialize, RO_INIT_MULTITHREADED};
-
-/// Output texture ring depth. The encode loop pipelines one frame deep (NVENC encodes frame N while
-/// the capturer produces N+1), so two live textures suffice; three gives headroom against a slow
-/// `lock_bitstream` and matches the WGC frame-pool depth.
-// Sized for the deep encode pipeline (`PUNKTFUNK_ENCODE_DEPTH`, default 4, clamped ≤ 6): up to DEPTH
-// frames are in flight in NVENC at once, so the HDR convert ring and the SDR held-frame set must each
-// keep DEPTH(+headroom) live textures, and the WGC pool needs spare buffers beyond what we hold.
-const OUT_RING: usize = 8;
-
-/// SDR zero-copy: how many recent WGC frames to keep alive so NVENC can encode the pool texture in
-/// place (no `CopyResource`). Each in-flight encode reads a distinct frame, so this must exceed the
-/// pipeline depth; the oldest is released once `HELD_FRAMES` newer ones exist.
-const HELD_FRAMES: usize = 8;
-/// WGC frame-pool buffer count. Must exceed `HELD_FRAMES` so the compositor always has free buffers
-/// to render into while we hold frames for in-place (zero-copy) SDR encode.
-const WGC_POOL_BUFFERS: i32 = 10;
-
-/// The host runs as SYSTEM (so the DDA secure-desktop path works), but WGC will NOT activate under
-/// the SYSTEM account (`CreateForMonitor` → 0x80070424). Impersonate the interactive console user
-/// for the WGC activation. Returns the user token (the caller reverts + closes it after activation)
-/// or `None` (no active user, or the host already runs AS the user — WTSQueryUserToken then fails and
-/// WGC works without impersonation). SYSTEM-only; harmless under a user-token host.
-unsafe fn impersonate_active_user() -> Option<HANDLE> {
-    let session = WTSGetActiveConsoleSessionId();
-    if session == 0xFFFF_FFFF {
-        return None;
-    }
-    let mut token = HANDLE::default();
-    if WTSQueryUserToken(session, &mut token).is_ok() {
-        if ImpersonateLoggedOnUser(token).is_ok() {
-            return Some(token);
-        }
-        let _ = CloseHandle(token);
-    }
-    None
-}
-
-/// RAII: reverts the WGC-activation impersonation when it drops (covers every `?` early-return).
-struct Deimpersonate(Option<HANDLE>);
-impl Drop for Deimpersonate {
-    fn drop(&mut self) {
-        if let Some(tok) = self.0.take() {
-            // SAFETY: `RevertToSelf` takes no arguments and undoes the thread impersonation set during
-            // WGC activation; `tok` is the impersonation token `HANDLE` from `impersonate_active_user`,
-            // owned by this `Deimpersonate` and closed exactly once here (taken out of the `Option`, so
-            // no double-close). Both are FFI calls borrowing no Rust memory.
-            unsafe {
-                let _ = RevertToSelf();
-                let _ = CloseHandle(tok);
-            }
-        }
-    }
-}
-
-/// Signal from the free-threaded FrameArrived callback to the encode thread: a monotonically
-/// increasing count of arrived frames + a condvar to wake `next_frame`. The encode thread tracks how
-/// many it has consumed; `TryGetNextFrame` is called exactly `available - consumed` times so we never
-/// hit the empty-pool ambiguity, and draining to the newest keeps latency at one frame.
-struct WgcSignal {
-    available: AtomicU64,
-    mtx: Mutex<()>,
-    cv: Condvar,
-}
-
-pub struct WgcCapturer {
-    device: ID3D11Device,
-    context: ID3D11DeviceContext,
-    // WGC objects — kept alive for the session's lifetime.
-    pool: Direct3D11CaptureFramePool,
-    session: GraphicsCaptureSession,
-    _item: GraphicsCaptureItem,
-    _frame_arrived_token: i64,
-    signal: Arc<WgcSignal>,
-    consumed: u64,
-
-    width: u32,
-    height: u32,
-    timeout_ms: u64,
-    first_frame: bool,
-
-    hdr: bool,
-    /// The source display's static HDR mastering metadata (ST.2086 + content light level), read from
-    /// `IDXGIOutput6::GetDesc1` at open when the output is HDR. Forwarded to the encoder (in-band SEI)
-    /// and the client (0xCE) by the stream loop. `None` when SDR. (The helper relay path also encodes,
-    /// so this is what gives the secure/normal-desktop HDR stream its mastering SEI.)
-    hdr_meta: Option<punktfunk_core::quic::HdrMeta>,
-    hdr_conv: Option<HdrConverter>,
-    fp16_src: Option<ID3D11Texture2D>,
-    fp16_srv: Option<ID3D11ShaderResourceView>,
-    /// `PUNKTFUNK_HDR_SHADER_P010` path: emit P010 (BT.2020 PQ 10-bit limited range) DIRECTLY from our
-    /// own shader (`HdrP010Converter`) so NVENC takes native P010 and skips its SM-side RGB→YUV CSC.
-    /// Gated by [`hdr_shader_p010_enabled`] AND `self.hdr`; `None`/empty when off → the existing R10 +
-    /// VideoProcessor paths run unchanged. `p010_disabled` latches a runtime failure (e.g. a driver
-    /// that rejects the planar plane RTV) so we fall back to the R10 path and stop retrying.
-    hdr_p010_conv: Option<HdrP010Converter>,
-    p010_out: Vec<ID3D11Texture2D>,
-    p010_idx: usize,
-    p010_disabled: bool,
-    /// Ring of host-owned output textures (BGRA for SDR, R10G10B10A2 for HDR), rotated per processed
-    /// frame. A ring — not one texture — is required because the encode loop is PIPELINED: NVENC
-    /// encodes frame N (in place, registered by pointer) while this capturer produces frame N+1, so
-    /// N+1 must land in a DIFFERENT texture or it clobbers the in-flight encode. (`fp16_src` stays
-    /// single: it's only touched within the D3D11 immediate context, whose op ordering already
-    /// serializes the convert's read against the next copy's write — NVENC's async engine read is the
-    /// only consumer that escapes that ordering, and it reads the ring output, never `fp16_src`.)
-    out_ring: Vec<ID3D11Texture2D>,
-    ring_idx: usize,
-    /// Video-processor RGB→YUV converter (off the 3D engine where possible) + its NV12/P010 output
-    /// ring. Preferred path: the OS-composited capture (cursor already in it) is converted DIRECTLY to
-    /// NVENC's native YUV — no `CopyResource`, no cursor draw, and NVENC skips its internal RGB→YUV.
-    /// `None`/error → falls back to the legacy SDR-zero-copy / HDR-shader paths.
-    video_conv: Option<VideoConverter>,
-    yuv_out: Vec<ID3D11Texture2D>,
-    yuv_idx: usize,
-    yuv_is_hdr: bool,
-    vp_disabled: bool,
-    /// SDR zero-copy: the recent WGC frames we hand to NVENC in place. Held so the pool doesn't
-    /// recycle the texture mid-encode; the oldest is released once `HELD_FRAMES` newer ones exist.
-    held: VecDeque<Direct3D11CaptureFrame>,
-    /// Last presentable GPU texture + format, repeated when no new frame arrived (static desktop).
-    last_present: Option<(ID3D11Texture2D, PixelFormat)>,
-
-    /// Owns the SudoVDA keepalive once attached (after WGC is confirmed open) — dropping the capturer
-    /// then REMOVEs the virtual output. `None` between open and attach so a WGC-open failure leaves
-    /// the keepalive with the caller for the DDA fallback.
-    _keepalive: Option<Box<dyn Send>>,
-}
-
-// SAFETY: like `DuplCapturer`. `WgcCapturer` holds D3D11 (free-threaded device/context) plus WGC WinRT
-// objects (`Direct3D11CaptureFramePool` etc., created free-threaded via `CreateFreeThreaded`). COM/WinRT
-// reference counting is interlocked, and the capturer is owned + used by exactly one encode thread,
-// moved to it once and never shared (no `Sync`), so transferring ownership across threads is sound. The
-// free-threaded `FrameArrived` callback touches only the `Arc<WgcSignal>` (itself `Send + Sync`), not
-// the capturer's COM fields.
-unsafe impl Send for WgcCapturer {}
-
-impl WgcCapturer {
-    /// Open WGC capture. Does NOT take the keepalive — the caller attaches it via
-    /// [`attach_keepalive`](Self::attach_keepalive) only after open succeeds, so a failure leaves the
-    /// keepalive with the caller to hand to the DDA fallback.
-    pub fn open(target: WinCaptureTarget, preferred: Option<(u32, u32, u32)>) -> Result<Self> {
-        // SAFETY: runs on the thread opening the WGC session. `RoInitialize` inits this thread's WinRT
-        // apartment (idempotent; result ignored). `impersonate_active_user()` and `find_output()` are
-        // this module's `unsafe fn`s whose contracts (call on the activating thread; pass a GDI name)
-        // are met, and the impersonation is reverted by `_deimp`'s Drop on every return path. Every
-        // COM/WinRT call thereafter operates on an object obtained + `?`-checked earlier in this same
-        // block on this single thread — the `IDXGIOutput1` from `find_output`, the device/context from
-        // `make_device`, the factory/interop/item/pool/session — and the `TypedEventHandler` closure
-        // captures an `Arc<WgcSignal>` (Send+Sync) by move. No raw pointers are dereferenced; borrowed
-        // locals outlive their synchronous calls.
-        unsafe {
-            // WGC is WinRT — the calling thread needs a COM/WinRT apartment for the GraphicsCaptureItem
-            // activation factory (RoGetActivationFactory). Initialize MTA; ignore "already initialized"
-            // / "changed mode" (another component on this thread may have init'd a compatible apartment).
-            let ro = RoInitialize(RO_INIT_MULTITHREADED);
-            // Impersonate the interactive user for the duration of WGC activation (host runs as
-            // SYSTEM; WGC won't activate under SYSTEM). Reverted by the guard's Drop on return. The
-            // WGC objects, once created, are accessed from the (SYSTEM) encode thread thereafter.
-            let imp = impersonate_active_user();
-            let _deimp = Deimpersonate(imp);
-            tracing::info!(ro_result = ?ro, impersonated = imp.is_some(), "WGC: RoInitialize(MTA)");
-            // The SudoVDA output appears a beat after the display is created — settle-retry like DDA.
-            let deadline = Instant::now() + Duration::from_millis(2000);
-            let (adapter, output) = loop {
-                if let Some(n) = crate::win_display::resolve_gdi_name(target.target_id) {
-                    if let Ok(found) = find_output(&n) {
-                        break found;
-                    }
-                }
-                if let Ok(found) = find_output(&target.gdi_name) {
-                    break found;
-                }
-                if Instant::now() >= deadline {
-                    bail!(
-                        "WGC: no DXGI output for SudoVDA target {} yet",
-                        target.target_id
-                    );
-                }
-                std::thread::sleep(Duration::from_millis(100));
-            };
-
-            let (device, context) = make_device(&adapter)?;
-            let od = output.GetDesc().context("output GetDesc")?;
-            let hmonitor = od.Monitor;
-
-            // HDR iff the output's colour space is BT.2020 PQ (G2084) — matches the DDA FP16 detection.
-            // From the same desc, read the source display's mastering metadata (ST.2086) when HDR.
-            let desc1 = output
-                .cast::<IDXGIOutput6>()
-                .ok()
-                .and_then(|o6| o6.GetDesc1().ok());
-            let hdr = desc1
-                .as_ref()
-                .map(|d1| d1.ColorSpace == DXGI_COLOR_SPACE_RGB_FULL_G2084_NONE_P2020)
-                .unwrap_or(false);
-            let hdr_meta = if hdr {
-                desc1.as_ref().map(|d| {
-                    crate::hdr::hdr_meta_from_display(
-                        (d.RedPrimary[0], d.RedPrimary[1]),
-                        (d.GreenPrimary[0], d.GreenPrimary[1]),
-                        (d.BluePrimary[0], d.BluePrimary[1]),
-                        (d.WhitePoint[0], d.WhitePoint[1]),
-                        d.MaxLuminance,
-                        d.MinLuminance,
-                        0, // MaxCLL: GetDesc1 has no content light level (Apollo zeroes it)
-                        0, // MaxFALL
-                    )
-                })
-            } else {
-                None
-            };
-
-            // Wrap our D3D11 device as a WinRT IDirect3DDevice so the frame pool allocates on it (the
-            // pool textures land on our device → CopyResource + NVENC are same-device, no readback).
-            let dxgi_device: IDXGIDevice = device.cast().context("ID3D11Device as IDXGIDevice")?;
-            let inspectable: IInspectable = CreateDirect3D11DeviceFromDXGIDevice(&dxgi_device)
-                .context("CreateDirect3D11DeviceFromDXGIDevice")?;
-            let d3d_device: windows::Graphics::DirectX::Direct3D11::IDirect3DDevice = inspectable
-                .cast()
-                .context("IInspectable as IDirect3DDevice")?;
-
-            tracing::info!(hdr, "WGC: device ready, creating capture item");
-            // GraphicsCaptureItem for the monitor (the SudoVDA output enumerates as a normal monitor).
-            let interop: IGraphicsCaptureItemInterop =
-                windows::core::factory::<GraphicsCaptureItem, IGraphicsCaptureItemInterop>()
-                    .context("GraphicsCaptureItem interop factory")?;
-            let item: GraphicsCaptureItem = interop
-                .CreateForMonitor(hmonitor)
-                .context("CreateForMonitor")?;
-            let size = item.Size().context("item Size")?;
-            let (width, height) = (size.Width.max(0) as u32, size.Height.max(0) as u32);
-            tracing::info!(
-                width,
-                height,
-                "WGC: capture item created, creating frame pool"
-            );
-
-            let pixel_format = if hdr {
-                DirectXPixelFormat::R16G16B16A16Float // scRGB FP16 — same surface DDA gives on HDR
-            } else {
-                DirectXPixelFormat::B8G8R8A8UIntNormalized
-            };
-            // Extra buffers: SDR zero-copy holds the last `HELD_FRAMES` frames (encoded in place), so
-            // the pool needs headroom beyond that for the producer to keep rendering at 240 Hz.
-            let pool = Direct3D11CaptureFramePool::CreateFreeThreaded(
-                &d3d_device,
-                pixel_format,
-                WGC_POOL_BUFFERS,
-                size,
-            )
-            .context("CreateFreeThreaded frame pool")?;
-
-            let signal = Arc::new(WgcSignal {
-                available: AtomicU64::new(0),
-                mtx: Mutex::new(()),
-                cv: Condvar::new(),
-            });
-            let sig = signal.clone();
-            let handler = TypedEventHandler::<Direct3D11CaptureFramePool, IInspectable>::new(
-                move |_pool, _arg| {
-                    sig.available.fetch_add(1, Ordering::Release);
-                    sig.cv.notify_one();
-                    Ok(())
-                },
-            );
-            let token = pool.FrameArrived(&handler).context("FrameArrived")?;
-
-            tracing::info!("WGC: creating capture session");
-            let session = pool
-                .CreateCaptureSession(&item)
-                .context("CreateCaptureSession")?;
-            // OS composites the cursor into the frame (HDR-correct, no manual composite pass).
-            let _ = session.SetIsCursorCaptureEnabled(true);
-            // Drop the yellow capture border (best-effort — older builds reject it).
-            let _ = session.SetIsBorderRequired(false);
-            // Lift the 60 Hz cap: allow up to the client's refresh (Win11 24H2+; below that this is a
-            // no-op and WGC caps ~60). 100 ns ticks per frame.
-            let refresh = preferred
-                .map(|(_, _, hz)| hz)
-                .filter(|&hz| hz > 0)
-                .unwrap_or(60);
-            let ticks = (10_000_000i64 / refresh.max(1) as i64).max(1);
-            let _ = session.SetMinUpdateInterval(TimeSpan { Duration: ticks });
-            tracing::info!("WGC: StartCapture");
-            session.StartCapture().context("StartCapture")?;
-            // WGC fires FrameArrived on CHANGE; a static desktop may never deliver the first frame
-            // (→ black, then the next_frame deadline ends the session). Nudge the cursor onto the
-            // output to force the first composition change, exactly like the DDA path does.
-            nudge_cursor_onto(&output);
-
-            let timeout_ms = (2000 / refresh.max(1) as u64).max(8);
-            tracing::info!(
-                width,
-                height,
-                hdr,
-                refresh,
-                "WGC capture started ({})",
-                if hdr {
-                    "HDR FP16→BT.2020 PQ"
-                } else {
-                    "SDR BGRA"
-                }
-            );
-
-            Ok(Self {
-                device,
-                context,
-                pool,
-                session,
-                _item: item,
-                _frame_arrived_token: token,
-                signal,
-                consumed: 0,
-                width,
-                height,
-                timeout_ms,
-                first_frame: true,
-                hdr,
-                hdr_meta,
-                hdr_conv: None,
-                fp16_src: None,
-                fp16_srv: None,
-                hdr_p010_conv: None,
-                p010_out: Vec::new(),
-                p010_idx: 0,
-                p010_disabled: false,
-                out_ring: Vec::new(),
-                ring_idx: 0,
-                video_conv: None,
-                yuv_out: Vec::new(),
-                yuv_idx: 0,
-                yuv_is_hdr: false,
-                vp_disabled: std::env::var_os("PUNKTFUNK_NO_VIDEO_PROCESSOR").is_some(),
-                held: VecDeque::new(),
-                last_present: None,
-                _keepalive: None,
-            })
-        }
-    }
-
-    /// Take ownership of the SudoVDA keepalive once the WGC session is confirmed open.
-    pub fn attach_keepalive(&mut self, keepalive: Box<dyn Send>) {
-        self._keepalive = Some(keepalive);
-    }
-
-    /// Block until a new frame arrives (cv), then drain `TryGetNextFrame` to the NEWEST queued frame
-    /// (skip stale → one-frame latency). Returns `None` on timeout (no new frame → caller repeats).
-    fn wait_and_drain(&mut self) -> Option<Direct3D11CaptureFrame> {
-        let wait_ms = if self.first_frame {
-            2000
-        } else {
-            self.timeout_ms
-        };
-        {
-            let mut g = self.signal.mtx.lock().unwrap();
-            while self.signal.available.load(Ordering::Acquire) <= self.consumed {
-                let (ng, res) = self
-                    .signal
-                    .cv
-                    .wait_timeout(g, Duration::from_millis(wait_ms))
-                    .unwrap();
-                g = ng;
-                if res.timed_out() {
-                    return None;
-                }
-            }
-        }
-        let target = self.signal.available.load(Ordering::Acquire);
-        let mut last = None;
-        while self.consumed < target {
-            if let Ok(f) = self.pool.TryGetNextFrame() {
-                last = Some(f);
-            }
-            self.consumed += 1;
-        }
-        last
-    }
-
-    unsafe fn ensure_fp16_src(&mut self) -> Result<()> {
-        if self.fp16_src.is_some() {
-            return Ok(());
-        }
-        let desc = tex_desc(
-            self.width,
-            self.height,
-            DXGI_FORMAT_R16G16B16A16_FLOAT,
-            (D3D11_BIND_RENDER_TARGET.0 | D3D11_BIND_SHADER_RESOURCE.0) as u32,
-        );
-        let mut t = None;
-        self.device
-            .CreateTexture2D(&desc, None, Some(&mut t))
-            .context("CreateTexture2D(wgc fp16 src)")?;
-        let t = t.context("fp16 src")?;
-        let mut srv = None;
-        self.device
-            .CreateShaderResourceView(&t, None, Some(&mut srv))?;
-        self.fp16_srv = Some(srv.context("fp16 srv")?);
-        self.fp16_src = Some(t);
-        Ok(())
-    }
-
-    /// Lazily allocate the HDR output texture ring (R10G10B10A2, the convert pass's render target →
-    /// NVENC input), `RENDER_TARGET`-bindable. SDR is zero-copy (encodes the WGC pool texture in
-    /// place) and uses no ring.
-    unsafe fn ensure_out_ring(
-        &mut self,
-        format: windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT,
-    ) -> Result<()> {
-        if !self.out_ring.is_empty() {
-            return Ok(());
-        }
-        let desc = tex_desc(
-            self.width,
-            self.height,
-            format,
-            D3D11_BIND_RENDER_TARGET.0 as u32,
-        );
-        for _ in 0..OUT_RING {
-            let mut t = None;
-            self.device
-                .CreateTexture2D(&desc, None, Some(&mut t))
-                .context("CreateTexture2D(wgc out ring)")?;
-            self.out_ring.push(t.context("wgc out ring tex")?);
-        }
-        Ok(())
-    }
-
-    /// Convert `input` (the OS-composited WGC pool texture: BGRA or scRGB FP16) → NVENC's native YUV
-    /// (NV12 / P010) on the video processor. Returns the YUV texture (from a ring so consecutive
-    /// encodes don't collide), or `None` to fall back to the legacy RGB paths.
-    unsafe fn convert_to_yuv(
-        &mut self,
-        input: &ID3D11Texture2D,
-        hdr: bool,
-    ) -> Option<ID3D11Texture2D> {
-        if self.vp_disabled {
-            return None;
-        }
-        if self.video_conv.is_none() || self.yuv_out.is_empty() || self.yuv_is_hdr != hdr {
-            self.video_conv = None;
-            self.yuv_out.clear();
-            self.yuv_idx = 0;
-            let vc = match VideoConverter::new(
-                &self.device,
-                &self.context,
-                self.width,
-                self.height,
-                hdr,
-            ) {
-                Ok(vc) => vc,
-                Err(e) => {
-                    tracing::warn!(error = %format!("{e:#}"),
-                        "WGC: video processor unavailable — falling back to RGB path");
-                    self.vp_disabled = true;
-                    return None;
-                }
-            };
-            let fmt = if hdr {
-                windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT_P010
-            } else {
-                windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT_NV12
-            };
-            let desc = tex_desc(
-                self.width,
-                self.height,
-                fmt,
-                D3D11_BIND_RENDER_TARGET.0 as u32,
-            );
-            for _ in 0..OUT_RING {
-                let mut t = None;
-                if self
-                    .device
-                    .CreateTexture2D(&desc, None, Some(&mut t))
-                    .is_err()
-                {
-                    tracing::warn!("WGC: CreateTexture2D(YUV) failed — falling back to RGB path");
-                    self.vp_disabled = true;
-                    self.yuv_out.clear();
-                    return None;
-                }
-                let Some(tex) = t else {
-                    self.vp_disabled = true;
-                    self.yuv_out.clear();
-                    return None;
-                };
-                self.yuv_out.push(tex);
-            }
-            self.video_conv = Some(vc);
-            self.yuv_is_hdr = hdr;
-            tracing::info!(
-                hdr,
-                "WGC: video-processor YUV path active ({})",
-                if hdr { "P010" } else { "NV12" }
-            );
-        }
-        let slot = self.yuv_idx;
-        self.yuv_idx = (self.yuv_idx + 1) % self.yuv_out.len();
-        let out = self.yuv_out[slot].clone();
-        if let Err(e) = self.video_conv.as_ref()?.convert(input, &out) {
-            tracing::warn!(error = %format!("{e:#}"),
-                "WGC: VideoProcessorBlt failed — falling back to RGB path");
-            self.vp_disabled = true;
-            self.video_conv = None;
-            self.yuv_out.clear();
-            return None;
-        }
-        Some(out)
-    }
-
-    /// `PUNKTFUNK_HDR_SHADER_P010` path: convert the OS-composited FP16 scRGB capture DIRECTLY to a
-    /// host-owned P010 texture (BT.2020 PQ, 10-bit limited range) via [`HdrP010Converter`] — two
-    /// shader passes writing the P010 planes. NVENC then takes native P010 and skips its internal
-    /// RGB→YUV CSC. Returns the next ring slot's P010 texture, or `Err` if the converter / a planar
-    /// plane RTV fails (the caller latches `p010_disabled` and falls back to the R10 path).
-    unsafe fn hdr_to_p010(&mut self, src: &ID3D11Texture2D) -> Result<ID3D11Texture2D> {
-        let slot = self.p010_idx;
-        // Lazily allocate the FP16 source (shared with the R10 path) + the P010 output ring.
-        self.ensure_fp16_src()?;
-        let fp16 = self.fp16_src.clone().context("fp16 src")?;
-        self.context.CopyResource(&fp16, src);
-        if self.p010_out.is_empty() {
-            let desc = tex_desc(
-                self.width,
-                self.height,
-                windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT_P010,
-                D3D11_BIND_RENDER_TARGET.0 as u32,
-            );
-            for _ in 0..OUT_RING {
-                let mut t = None;
-                self.device
-                    .CreateTexture2D(&desc, None, Some(&mut t))
-                    .context("CreateTexture2D(wgc p010 ring)")?;
-                self.p010_out.push(t.context("wgc p010 ring tex")?);
-            }
-        }
-        self.p010_idx = (self.p010_idx + 1) % self.p010_out.len();
-        let out = self.p010_out[slot].clone();
-        if self.hdr_p010_conv.is_none() {
-            self.hdr_p010_conv = Some(HdrP010Converter::new(&self.device)?);
-        }
-        let srv = self.fp16_srv.clone().context("fp16 srv")?;
-        self.hdr_p010_conv.as_ref().unwrap().convert(
-            &self.device,
-            &self.context,
-            &srv,
-            &out,
-            self.width,
-            self.height,
-        )?;
-        Ok(out)
-    }
-
-    fn process_frame(&mut self, frame: Direct3D11CaptureFrame) -> Result<CapturedFrame> {
-        // SAFETY: runs on the capturer's single owning thread. `frame` is a live
-        // `Direct3D11CaptureFrame` from `self.pool`; `frame.Surface().cast::<IDirect3DDxgiInterfaceAccess
-        // >().GetInterface()` yields the frame's backing `ID3D11Texture2D`, which belongs to
-        // `self.device` (the pool was created on it via `CreateDirect3D11DeviceFromDXGIDevice`). Every
-        // helper called here — `hdr_to_p010`, `convert_to_yuv`, `ensure_fp16_src`, `ensure_out_ring`,
-        // `HdrConverter::convert`, `CopyResource`, `CreateRenderTargetView` — operates on
-        // `self.device`/`self.context` and that same-device texture, so all resources share one device.
-        // The frame is held in `self.held` until its async GPU read completes for the zero-copy paths.
-        // Single-threaded immediate-context use; borrowed textures/SRVs/RTVs outlive each synchronous call.
-        unsafe {
-            let surface = frame.Surface().context("frame Surface")?;
-            let access: IDirect3DDxgiInterfaceAccess = surface
-                .cast()
-                .context("surface as IDirect3DDxgiInterfaceAccess")?;
-            let src: ID3D11Texture2D = access
-                .GetInterface()
-                .context("GetInterface ID3D11Texture2D")?;
-
-            // GATED P010-shader path (`PUNKTFUNK_HDR_SHADER_P010`): for HDR, emit P010 (BT.2020 PQ
-            // 10-bit limited range) DIRECTLY from our shader so NVENC takes native P010 and skips its
-            // SM-side RGB→YUV CSC. Runs BEFORE the R10 + VideoProcessor path. A converter/plane-RTV
-            // failure latches `p010_disabled` → we fall through to the unchanged R10 path for the rest
-            // of the session. Default OFF → none of this executes and behaviour is byte-for-byte as
-            // today.
-            if self.hdr && !self.p010_disabled && hdr_shader_p010_enabled() {
-                match self.hdr_to_p010(&src) {
-                    Ok(p010) => {
-                        // The P010 output is host-owned (the ring), and the FP16 CopyResource read
-                        // `src` synchronously on the immediate context before the shader passes — so we
-                        // do NOT need to hold `frame` past here (unlike the SDR/R10 in-place paths).
-                        // Dropping it returns the pool buffer to WGC immediately.
-                        drop(frame);
-                        self.last_present = Some((p010.clone(), PixelFormat::P010));
-                        return Ok(self.d3d11_frame(p010, PixelFormat::P010));
-                    }
-                    Err(e) => {
-                        tracing::warn!(error = %format!("{e:#}"),
-                            "WGC: HDR P010 shader path failed — disabling it, falling back to R10");
-                        self.p010_disabled = true;
-                        self.hdr_p010_conv = None;
-                        self.p010_out.clear();
-                    }
-                }
-            }
-
-            // Preferred path: convert the OS-composited capture (cursor already in it) DIRECTLY to
-            // NVENC's native YUV on the video processor — no CopyResource, no cursor draw, and NVENC
-            // skips its internal RGB→YUV (the contended 3D step). WGC's multi-buffer pool + held set
-            // means reading the pool texture directly does NOT serialize (unlike DDA's single-frame
-            // model). The frame is held until the async Blt finishes. (HDR: the video processor can't
-            // ingest FP16 scRGB, so the Blt fails and we fall back to the R10 path below; the
-            // `PUNKTFUNK_HDR_SHADER_P010` branch above is the off-the-SM HDR path.)
-            if let Some(yuv) = self.convert_to_yuv(&src, self.hdr) {
-                let fmt = if self.hdr {
-                    PixelFormat::P010
-                } else {
-                    PixelFormat::Nv12
-                };
-                self.last_present = Some((yuv.clone(), fmt));
-                let out = self.d3d11_frame(yuv, fmt);
-                self.held.push_back(frame);
-                while self.held.len() > HELD_FRAMES {
-                    self.held.pop_front();
-                }
-                return Ok(out);
-            }
-
-            // --- fallback (video processor unavailable) ---
-            if self.hdr {
-                // Next ring slot — the in-flight encode reads the slot we handed out last time, so
-                // this capture must land in a different one (see `out_ring`).
-                let slot = self.ring_idx;
-                self.ring_idx = (self.ring_idx + 1) % OUT_RING;
-                // FP16 (cursor already composited by the OS) → BT.2020 PQ 10-bit for NVENC.
-                self.ensure_fp16_src()?;
-                let fp16 = self.fp16_src.clone().context("fp16 src")?;
-                self.context.CopyResource(&fp16, &src);
-                self.ensure_out_ring(DXGI_FORMAT_R10G10B10A2_UNORM)?;
-                let out = self.out_ring[slot].clone();
-                if self.hdr_conv.is_none() {
-                    self.hdr_conv = Some(HdrConverter::new(&self.device)?);
-                }
-                let srv = self.fp16_srv.clone().context("fp16 srv")?;
-                let mut rtv: Option<ID3D11RenderTargetView> = None;
-                self.device
-                    .CreateRenderTargetView(&out, None, Some(&mut rtv))?;
-                let rtv = rtv.context("hdr10 rtv")?;
-                self.hdr_conv.as_ref().unwrap().convert(
-                    &self.context,
-                    &srv,
-                    &rtv,
-                    self.width,
-                    self.height,
-                );
-                self.last_present = Some((out.clone(), PixelFormat::Rgb10a2));
-                Ok(self.d3d11_frame(out, PixelFormat::Rgb10a2))
-            } else {
-                // SDR ZERO-COPY: hand NVENC the WGC pool texture DIRECTLY — no `CopyResource`. The
-                // per-frame copy otherwise queues on the graphics engine behind a GPU-saturating game
-                // and stalls `lock_bitstream` ~20 ms (NVENC sits idle waiting for its input). Encoding
-                // the pool texture in place removes that graphics-queue dependency (Apollo's model).
-                // We must keep the frame alive until its async encode finishes, so retain the last
-                // `HELD_FRAMES`; the pool has spare buffers so the producer never starves.
-                self.last_present = Some((src.clone(), PixelFormat::Bgra));
-                let out = self.d3d11_frame(src, PixelFormat::Bgra);
-                self.held.push_back(frame);
-                while self.held.len() > HELD_FRAMES {
-                    self.held.pop_front();
-                }
-                Ok(out)
-            }
-        }
-    }
-
-    fn d3d11_frame(&self, texture: ID3D11Texture2D, format: PixelFormat) -> CapturedFrame {
-        CapturedFrame {
-            width: self.width,
-            height: self.height,
-            pts_ns: now_ns(),
-            format,
-            payload: FramePayload::D3d11(D3d11Frame {
-                texture,
-                device: self.device.clone(),
-            }),
-        }
-    }
-}
-
-impl Capturer for WgcCapturer {
-    fn hdr_meta(&self) -> Option<punktfunk_core::quic::HdrMeta> {
-        self.hdr_meta
-    }
-
-    fn next_frame(&mut self) -> Result<CapturedFrame> {
-        let overall = Instant::now() + Duration::from_secs(20);
-        loop {
-            if let Some(frame) = self.wait_and_drain() {
-                self.first_frame = false;
-                return self.process_frame(frame);
-            }
-            // No new frame within the wait — repeat the last presented frame (static desktop).
-            if let Some((tex, fmt)) = &self.last_present {
-                return Ok(self.d3d11_frame(tex.clone(), *fmt));
-            }
-            if Instant::now() > overall {
-                bail!("no WGC frame within 20s (SudoVDA monitor not lit / no capture access?)");
-            }
-        }
-    }
-
-    fn try_latest(&mut self) -> Result<Option<CapturedFrame>> {
-        let target = self.signal.available.load(Ordering::Acquire);
-        if target <= self.consumed {
-            return Ok(None);
-        }
-        let mut last = None;
-        while self.consumed < target {
-            if let Ok(f) = self.pool.TryGetNextFrame() {
-                last = Some(f);
-            }
-            self.consumed += 1;
-        }
-        match last {
-            Some(frame) => self.process_frame(frame).map(Some),
-            None => Ok(None),
-        }
-    }
-    // set_active: the trait default (no-op) is correct — WGC keeps its session running across the
-    // active/idle gate (cheap; the frame pool just recycles), like the DDA duplication.
-}
-
-impl Drop for WgcCapturer {
-    fn drop(&mut self) {
-        let _ = self.session.Close();
-        let _ = self.pool.Close();
-        // _keepalive drops after, REMOVEing the SudoVDA monitor.
-    }
-}
-
-fn tex_desc(
-    width: u32,
-    height: u32,
-    format: windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT,
-    bind: u32,
-) -> D3D11_TEXTURE2D_DESC {
-    D3D11_TEXTURE2D_DESC {
-        Width: width,
-        Height: height,
-        MipLevels: 1,
-        ArraySize: 1,
-        Format: format,
-        SampleDesc: DXGI_SAMPLE_DESC {
-            Count: 1,
-            Quality: 0,
-        },
-        Usage: D3D11_USAGE_DEFAULT,
-        BindFlags: bind,
-        CPUAccessFlags: 0,
-        MiscFlags: 0,
-    }
-}
-
-fn now_ns() -> u64 {
-    std::time::SystemTime::now()
-        .duration_since(std::time::UNIX_EPOCH)
-        .map(|d| d.as_nanos() as u64)
-        .unwrap_or(0)
-}

crates/punktfunk-host/src/capture/windows/wgc_relay.rs

@@ -1,484 +0,0 @@
-//! Host-side WGC helper relay (Windows two-process secure-desktop design,
-//! design/archive/windows-secure-desktop.md — step 4).
-//!
-//! WGC won't activate under the SYSTEM account, so the SYSTEM host can't capture the normal desktop
-//! itself. Instead it spawns `punktfunk-host wgc-helper` in the **interactive user session** (so WGC works)
-//! via `CreateProcessAsUserW`, with the helper's **stdout** redirected to an anonymous pipe the host
-//! reads. The helper ships framed Annex-B access units; this module parses them back into AUs the
-//! host relays onto the live QUIC session (same `EncodedFrame` flow, just sourced over a pipe instead
-//! of a local encoder). A second pipe carries a tiny **control** channel to the helper (stdin: force
-//! keyframe), and the helper's **stderr** is forwarded line-by-line into host tracing so its logs are
-//! visible from the SYSTEM host's console.
-//!
-//! Wire framing (must match `wgc_helper::write_au`): per AU
-//! `[u32 magic "PFAU" LE][u32 len LE][u64 pts_ns LE][u8 keyframe][len bytes data]`.
-
-// Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
-#![deny(clippy::undocumented_unsafe_blocks)]
-
-use crate::capture::dxgi::WinCaptureTarget;
-use anyhow::{bail, Context, Result};
-use std::io::{BufRead, BufReader, Read};
-use std::sync::mpsc::{Receiver, SyncSender};
-use std::sync::Mutex;
-use windows::core::PWSTR;
-use windows::Win32::Foundation::SetHandleInformation;
-use windows::Win32::Foundation::{CloseHandle, HANDLE};
-use windows::Win32::Foundation::{HANDLE_FLAGS, HANDLE_FLAG_INHERIT};
-use windows::Win32::Security::{
-    DuplicateTokenEx, SecurityImpersonation, TokenPrimary, SECURITY_ATTRIBUTES, TOKEN_ALL_ACCESS,
-};
-use windows::Win32::System::Environment::{CreateEnvironmentBlock, DestroyEnvironmentBlock};
-use windows::Win32::System::Pipes::CreatePipe;
-use windows::Win32::System::RemoteDesktop::{WTSGetActiveConsoleSessionId, WTSQueryUserToken};
-use windows::Win32::System::Threading::{
-    CreateProcessAsUserW, TerminateProcess, CREATE_NO_WINDOW, CREATE_UNICODE_ENVIRONMENT,
-    PROCESS_INFORMATION, STARTF_USESTDHANDLES, STARTUPINFOW,
-};
-
-/// Must match [`crate::wgc_helper`]'s `AU_MAGIC` ("PFAU").
-const AU_MAGIC: u32 = 0x5046_4155;
-
-/// One access unit relayed from the helper, in the helper's (= the host's, same machine) monotonic
-/// clock — `pts_ns` is directly comparable to the host's `now_ns()`.
-pub struct RelayAu {
-    pub data: Vec<u8>,
-    pub pts_ns: u64,
-    pub keyframe: bool,
-}
-
-/// A running USER-session WGC helper whose AUs the SYSTEM host relays. Drop kills the child + closes
-/// the pipes; the reader threads then end on the broken pipe.
-pub struct HelperRelay {
-    proc: HANDLE,
-    thread: HANDLE,
-    /// Host write end of the helper's stdin — control commands (force keyframe). Mutex so the relay
-    /// can be shared while the encode thread requests keyframes.
-    stdin_w: Mutex<HANDLE>,
-    /// Parsed AUs from the helper's stdout reader thread.
-    rx: Receiver<RelayAu>,
-}
-
-// SAFETY: every field is itself `Send`: the `proc`/`thread` `HANDLE`s are process-global kernel
-// handle values (plain integers valid from any thread, owned for the relay's lifetime and closed once
-// on Drop), `stdin_w` is a `Mutex<HANDLE>`, and `rx` is an mpsc `Receiver<RelayAu>` (which is `Send`).
-// The relay is moved to one thread and owned there, so transferring it across threads is sound.
-unsafe impl Send for HelperRelay {}
-// NOTE: `HelperRelay` is deliberately NOT `Sync`. Its `rx: Receiver<RelayAu>` is `!Sync` (std mpsc
-// is single-consumer), and the relay is only ever a single-owner local in the punktfunk1 two-process
-// mux loop — never shared by `&` across threads — so `Sync` is neither sound nor needed. (A prior
-// `unsafe impl Sync` here asserted more than the fields support; removed.)
-
-/// Control byte on the helper's stdin: force the next encoded frame to be an IDR (client decode
-/// recovery). Mirrors `enc.request_keyframe()` in the single-process path.
-const CTL_KEYFRAME: u8 = 0x01;
-
-impl HelperRelay {
-    /// Spawn the helper in the interactive user session and start relaying its AUs. `target` is the
-    /// SudoVDA output the host already created (captured by GDI name only — the helper never touches
-    /// display topology). `(w, h, hz)` is the negotiated mode; `bitrate_kbps` the negotiated bitrate.
-    pub fn spawn(
-        target: &WinCaptureTarget,
-        mode: (u32, u32, u32),
-        bitrate_kbps: u32,
-        bit_depth: u8,
-    ) -> Result<HelperRelay> {
-        let exe = std::env::current_exe().context("current_exe for helper spawn")?;
-        let exe = exe.to_string_lossy().into_owned();
-        let (w, h, hz) = mode;
-        // CreateProcessAsUserW takes a single mutable command line (argv[0] = exe).
-        let cmdline = format!(
-            "\"{exe}\" wgc-helper --gdi \"{}\" --target-id {} --mode {w}x{h}x{hz} --bitrate {bitrate_kbps} --bit-depth {bit_depth}",
-            target.gdi_name, target.target_id
-        );
-        tracing::info!(cmd = %cmdline, "spawning WGC helper in user session");
-
-        // SAFETY: `spawn_inner` is an `unsafe fn` only because it drives raw Win32 token/pipe/process
-        // FFI; it imposes no caller-side memory precondition beyond valid arguments. `cmdline` is a live
-        // `&str` borrowed for the synchronous call and `(w, h, hz)` are plain `u32`s. It validates its
-        // own runtime requirements (active console session, SYSTEM token) and returns `Err` otherwise.
-        unsafe { spawn_inner(&cmdline, w, h, hz) }
-    }
-
-    /// Receive the next relayed AU. Distinguishes a `Timeout` (helper slow/stalled — keep waiting)
-    /// from `Disconnected` (helper exited → its stdout closed → reader thread ended → channel
-    /// dropped), which returns *immediately* and means the relay must stop, not spin.
-    pub fn recv_timeout(
-        &self,
-        dur: std::time::Duration,
-    ) -> Result<RelayAu, std::sync::mpsc::RecvTimeoutError> {
-        self.rx.recv_timeout(dur)
-    }
-
-    /// Non-blocking receive — used to drain stale buffered AUs (encoded while the secure desktop was
-    /// the live source) before resuming the relay. `Ok` while AUs remain, `Err` once empty.
-    pub fn try_recv(&self) -> Result<RelayAu, std::sync::mpsc::TryRecvError> {
-        self.rx.try_recv()
-    }
-
-    /// Ask the helper's encoder for an IDR on the next frame (client decode recovery). Best-effort:
-    /// a write failure means the helper is gone — the caller's recv loop will see the disconnect.
-    pub fn request_keyframe(&self) {
-        let h = self.stdin_w.lock().unwrap();
-        let mut written = 0u32;
-        // SAFETY: `*h` is the host's write end of the helper's stdin pipe — a live `HANDLE` owned by
-        // this `HelperRelay` (held under the `stdin_w` Mutex, locked here), closed only in Drop.
-        // `WriteFile` reads the 1-byte `&[CTL_KEYFRAME]` buffer and writes the byte count into
-        // `written`; both are live locals that outlive the synchronous call. A failure (helper gone) is
-        // discarded as documented.
-        unsafe {
-            let _ = windows::Win32::Storage::FileSystem::WriteFile(
-                *h,
-                Some(&[CTL_KEYFRAME]),
-                Some(&mut written),
-                None,
-            );
-        }
-    }
-}
-
-impl Drop for HelperRelay {
-    fn drop(&mut self) {
-        // SAFETY: `self.proc`/`self.thread` are the child process/thread `HANDLE`s from
-        // `CreateProcessAsUserW`, and `stdin_w` is the host's pipe write end — all owned by this
-        // `HelperRelay` and closed exactly once here in Drop (no double-close). `TerminateProcess` and
-        // the three `CloseHandle`s are FFI calls taking those handles by value, borrowing no Rust memory.
-        unsafe {
-            // Terminate the child first so its WGC capture + NVENC session tear down, then close our
-            // handles (the reader threads end on the resulting broken pipe).
-            let _ = TerminateProcess(self.proc, 1);
-            let _ = CloseHandle(*self.stdin_w.lock().unwrap());
-            let _ = CloseHandle(self.proc);
-            let _ = CloseHandle(self.thread);
-        }
-        tracing::info!("WGC helper relay torn down");
-    }
-}
-
-/// Inheritable anonymous pipe (read, write). The caller marks whichever end the host keeps as
-/// non-inheritable so the child only inherits its own end.
-unsafe fn make_pipe() -> Result<(HANDLE, HANDLE)> {
-    let mut read = HANDLE::default();
-    let mut write = HANDLE::default();
-    let sa = SECURITY_ATTRIBUTES {
-        nLength: std::mem::size_of::<SECURITY_ATTRIBUTES>() as u32,
-        lpSecurityDescriptor: std::ptr::null_mut(),
-        bInheritHandle: true.into(),
-    };
-    CreatePipe(&mut read, &mut write, Some(&sa), 0).context("CreatePipe")?;
-    Ok((read, write))
-}
-
-/// Mark a handle non-inheritable (the host keeps it; the child must not get a copy).
-unsafe fn no_inherit(h: HANDLE) {
-    let _ = SetHandleInformation(h, HANDLE_FLAG_INHERIT.0, HANDLE_FLAGS(0));
-}
-
-/// Build a child environment block: the target session's block (so DLL/PATH/SystemRoot resolve) with
-/// this process's `PUNKTFUNK_*` vars overlaid, so the child runs with the SAME settings this process
-/// has (`PUNKTFUNK_ENCODER=nvenc`, `PUNKTFUNK_ZEROCOPY`, …) instead of the target shell's. Returns a
-/// UTF-16, double-null-terminated block suitable for `CREATE_UNICODE_ENVIRONMENT`. Shared by the WGC
-/// helper spawn (here) and the Windows service launching the host into the active session.
-pub(crate) unsafe fn merged_env_block(user_block: *const u16) -> Vec<u16> {
-    // Parse the user block ("VAR=VALUE\0" … "\0") into entries.
-    let mut entries: Vec<String> = Vec::new();
-    if !user_block.is_null() {
-        let mut p = user_block;
-        loop {
-            let mut len = 0isize;
-            while *p.offset(len) != 0 {
-                len += 1;
-            }
-            if len == 0 {
-                break; // the trailing empty string = end of block
-            }
-            let slice = std::slice::from_raw_parts(p, len as usize);
-            entries.push(String::from_utf16_lossy(slice));
-            p = p.offset(len + 1);
-        }
-    }
-    // Overlay "our" settings — PUNKTFUNK_* and RUST_LOG — dropping whatever the target block had.
-    let is_ours = |k: &str| k.starts_with("PUNKTFUNK_") || k == "RUST_LOG";
-    entries.retain(|e| !is_ours(e.split('=').next().unwrap_or("")));
-    for (k, v) in std::env::vars().filter(|(k, _)| is_ours(k)) {
-        entries.push(format!("{k}={v}"));
-    }
-    // Serialize back to a UTF-16 double-null-terminated block.
-    let mut block: Vec<u16> = Vec::new();
-    for e in entries {
-        block.extend(e.encode_utf16());
-        block.push(0);
-    }
-    block.push(0);
-    block
-}
-
-unsafe fn spawn_inner(cmdline: &str, w: u32, h: u32, hz: u32) -> Result<HelperRelay> {
-    // The user token of the active console session (requires the host to be SYSTEM).
-    let session = WTSGetActiveConsoleSessionId();
-    if session == 0xFFFF_FFFF {
-        bail!("no active console session (WTSGetActiveConsoleSessionId)");
-    }
-    let mut user_token = HANDLE::default();
-    WTSQueryUserToken(session, &mut user_token)
-        .context("WTSQueryUserToken (host must run as SYSTEM)")?;
-
-    // A primary token for CreateProcessAsUserW.
-    let mut primary = HANDLE::default();
-    let dup = DuplicateTokenEx(
-        user_token,
-        TOKEN_ALL_ACCESS,
-        None,
-        SecurityImpersonation,
-        TokenPrimary,
-        &mut primary,
-    );
-    let _ = CloseHandle(user_token);
-    dup.context("DuplicateTokenEx(TokenPrimary)")?;
-
-    // The user's environment block (PATH, USERPROFILE, SystemRoot → DLL resolution), MERGED with the
-    // host's PUNKTFUNK_* vars. CreateProcessAsUserW would otherwise give the helper the *user's* env
-    // only, dropping PUNKTFUNK_ENCODER=nvenc / PUNKTFUNK_ZEROCOPY/… that the host runs with — so the
-    // helper would fall back to the software (H.264-only) encoder. We parse the user block, strip any
-    // PUNKTFUNK_* it has, append the host's, and pass the merged block.
-    let mut env_block: *mut core::ffi::c_void = std::ptr::null_mut();
-    let _ = CreateEnvironmentBlock(&mut env_block, Some(primary), false);
-    let merged_env = merged_env_block(env_block as *const u16);
-    if !env_block.is_null() {
-        let _ = DestroyEnvironmentBlock(env_block);
-    }
-
-    // Three pipes: stdout (helper→host AUs), stdin (host→helper control), stderr (helper→host logs).
-    let (out_r, out_w) = make_pipe().context("stdout pipe")?;
-    let (in_r, in_w) = make_pipe().context("stdin pipe")?;
-    let (err_r, err_w) = make_pipe().context("stderr pipe")?;
-    // The host keeps out_r / in_w / err_r — none inheritable; the child inherits out_w/in_r/err_w.
-    no_inherit(out_r);
-    no_inherit(in_w);
-    no_inherit(err_r);
-
-    let mut si = STARTUPINFOW {
-        cb: std::mem::size_of::<STARTUPINFOW>() as u32,
-        dwFlags: STARTF_USESTDHANDLES,
-        hStdInput: in_r,
-        hStdOutput: out_w,
-        hStdError: err_w,
-        ..Default::default()
-    };
-    // WGC needs the interactive desktop.
-    let mut desktop: Vec<u16> = "winsta0\\default\0".encode_utf16().collect();
-    si.lpDesktop = PWSTR(desktop.as_mut_ptr());
-
-    let mut cmd: Vec<u16> = cmdline.encode_utf16().chain(std::iter::once(0)).collect();
-    let mut pi = PROCESS_INFORMATION::default();
-
-    let created = CreateProcessAsUserW(
-        Some(primary),
-        None,
-        Some(PWSTR(cmd.as_mut_ptr())),
-        None,
-        None,
-        true, // inherit handles (the child's std ends)
-        CREATE_UNICODE_ENVIRONMENT | CREATE_NO_WINDOW,
-        Some(merged_env.as_ptr() as *const core::ffi::c_void),
-        None,
-        &si,
-        &mut pi,
-    );
-
-    // Clean up regardless of outcome: the child now owns its inherited ends; close our copies.
-    let _ = CloseHandle(out_w);
-    let _ = CloseHandle(in_r);
-    let _ = CloseHandle(err_w);
-    let _ = CloseHandle(primary);
-
-    if let Err(e) = created {
-        let _ = CloseHandle(out_r);
-        let _ = CloseHandle(in_w);
-        let _ = CloseHandle(err_r);
-        return Err(e).context("CreateProcessAsUserW(wgc-helper)");
-    }
-    tracing::info!(pid = pi.dwProcessId, mode = %format!("{w}x{h}@{hz}"), "WGC helper spawned");
-
-    // The helper does the WGC capture + NVENC encode, but it runs under the user's UAC-FILTERED token
-    // (no SE_INC_BASE_PRIORITY), so it can't raise its OWN GPU scheduling-priority class — under a
-    // GPU-saturating game NVENC then gets starved (the "240→40 fps in-game collapse"). The SYSTEM host
-    // holds the privilege, so stamp the HIGH GPU priority class onto the child here, right after spawn
-    // (the process-level class applies to the GPU contexts the helper creates afterwards).
-    crate::capture::dxgi::set_child_gpu_priority_class(pi.hProcess);
-
-    // stderr → host tracing, line by line.
-    let err_handle = HandleReader(err_r);
-    std::thread::Builder::new()
-        .name("wgc-helper-log".into())
-        .spawn(move || {
-            let r = BufReader::new(err_handle);
-            for line in r.lines() {
-                match line {
-                    Ok(l) if !l.trim().is_empty() => tracing::info!(target: "wgc_helper", "{l}"),
-                    Ok(_) => {}
-                    Err(_) => break,
-                }
-            }
-        })
-        .ok();
-
-    // stdout → parsed AUs. Bounded so a stalled relay applies backpressure (the pipe then fills and
-    // the helper blocks on write — the same backpressure the single-process channel gives).
-    let (tx, rx) = std::sync::mpsc::sync_channel::<RelayAu>(3);
-    let out_handle = HandleReader(out_r);
-    std::thread::Builder::new()
-        .name("wgc-helper-au".into())
-        .spawn(move || au_reader(out_handle, tx))
-        .ok();
-
-    Ok(HelperRelay {
-        proc: pi.hProcess,
-        thread: pi.hThread,
-        stdin_w: Mutex::new(in_w),
-        rx,
-    })
-}
-
-/// Parse the AU framing off the helper's stdout and forward each AU. Ends (returns) when the pipe
-/// breaks (helper exit) or the channel's receiver is dropped (relay torn down).
-fn au_reader(mut r: HandleReader, tx: SyncSender<RelayAu>) {
-    loop {
-        let mut hdr = [0u8; 4 + 4 + 8 + 1];
-        if r.read_exact(&mut hdr).is_err() {
-            break;
-        }
-        let magic = u32::from_le_bytes([hdr[0], hdr[1], hdr[2], hdr[3]]);
-        if magic != AU_MAGIC {
-            tracing::error!(
-                magic = format!("{magic:#x}"),
-                "WGC helper AU stream desync — aborting relay"
-            );
-            break;
-        }
-        let len = u32::from_le_bytes([hdr[4], hdr[5], hdr[6], hdr[7]]) as usize;
-        let pts_ns = u64::from_le_bytes([
-            hdr[8], hdr[9], hdr[10], hdr[11], hdr[12], hdr[13], hdr[14], hdr[15],
-        ]);
-        let keyframe = hdr[16] != 0;
-        // Bound the allocation — a corrupt length must not OOM the host. 64 MiB is far above any real
-        // AU (a 5K keyframe is a few MB).
-        if len > 64 * 1024 * 1024 {
-            tracing::error!(len, "WGC helper AU length implausible — aborting relay");
-            break;
-        }
-        let mut data = vec![0u8; len];
-        if r.read_exact(&mut data).is_err() {
-            break;
-        }
-        if tx
-            .send(RelayAu {
-                data,
-                pts_ns,
-                keyframe,
-            })
-            .is_err()
-        {
-            break; // relay dropped
-        }
-    }
-}
-
-/// Minimal `Read` over a Win32 pipe HANDLE (the windows crate doesn't impl `Read` on HANDLE).
-struct HandleReader(HANDLE);
-// SAFETY: `HandleReader` owns a single pipe `HANDLE` (a process-global kernel handle value, valid from
-// any thread). It is moved into the dedicated reader thread and used only there (and closed once on
-// Drop), never shared — so transferring ownership across threads is sound.
-unsafe impl Send for HandleReader {}
-impl Read for HandleReader {
-    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
-        let mut read = 0u32;
-        // SAFETY: `self.0` is the live read end of an anonymous pipe owned by this `HandleReader`
-        // (closed only in Drop). `ReadFile` fills the caller-provided `buf` (writing at most `buf.len()`
-        // bytes) and stores the count in `read`; both outlive the synchronous call. A broken pipe
-        // surfaces as `Err` and is mapped to EOF below.
-        let ok = unsafe {
-            windows::Win32::Storage::FileSystem::ReadFile(self.0, Some(buf), Some(&mut read), None)
-        };
-        match ok {
-            Ok(()) => Ok(read as usize),
-            // A broken pipe (helper exited) reads as ERROR_BROKEN_PIPE → report EOF (0).
-            Err(_) => Ok(0),
-        }
-    }
-}
-impl Drop for HandleReader {
-    fn drop(&mut self) {
-        // SAFETY: `self.0` is the pipe `HANDLE` this `HandleReader` owns; `CloseHandle` (an FFI call
-        // taking the handle by value) is invoked exactly once here in Drop, so there is no double-close.
-        unsafe {
-            let _ = CloseHandle(self.0);
-        }
-    }
-}
-
-/// Is this process running as the LOCAL SYSTEM account? Used to decide whether the two-process
-/// secure-desktop path applies (only SYSTEM can `WTSQueryUserToken` + capture the Winlogon desktop).
-pub fn running_as_system() -> bool {
-    use windows::Win32::Security::{GetTokenInformation, TokenUser, TOKEN_QUERY, TOKEN_USER};
-    use windows::Win32::System::Threading::{GetCurrentProcess, OpenProcessToken};
-    // SAFETY: `OpenProcessToken(GetCurrentProcess(), TOKEN_QUERY, &mut token)` opens the current-process
-    // token (the pseudo-handle is always valid) into `token`, which is closed once before each return.
-    // The first `GetTokenInformation` (null buffer) queries the required `len`; `buf` is then a
-    // `Vec<u8>` of exactly `len` bytes and the second call fills it, so `&*(buf.as_ptr() as *const
-    // TOKEN_USER)` reads a `TOKEN_USER` the kernel just wrote into a sufficiently-sized buffer (the
-    // variable-length SID it points at also lies within `buf`, which outlives the borrow).
-    // `is_local_system_sid` is this module's `unsafe fn`, given that in-buffer `PSID`. Safe on any thread.
-    unsafe {
-        let mut token = HANDLE::default();
-        if OpenProcessToken(GetCurrentProcess(), TOKEN_QUERY, &mut token).is_err() {
-            return false;
-        }
-        let mut len = 0u32;
-        let _ = GetTokenInformation(token, TokenUser, None, 0, &mut len);
-        if len == 0 {
-            let _ = CloseHandle(token);
-            return false;
-        }
-        let mut buf = vec![0u8; len as usize];
-        let ok = GetTokenInformation(
-            token,
-            TokenUser,
-            Some(buf.as_mut_ptr() as *mut _),
-            len,
-            &mut len,
-        )
-        .is_ok();
-        let _ = CloseHandle(token);
-        if !ok {
-            return false;
-        }
-        let tu = &*(buf.as_ptr() as *const TOKEN_USER);
-        // The well-known LocalSystem SID is S-1-5-18.
-        is_local_system_sid(tu.User.Sid)
-    }
-}
-
-/// True iff `sid` is S-1-5-18 (LocalSystem).
-unsafe fn is_local_system_sid(sid: windows::Win32::Security::PSID) -> bool {
-    use windows::Win32::Security::{
-        GetSidIdentifierAuthority, GetSidSubAuthority, GetSidSubAuthorityCount, IsValidSid,
-    };
-    if !IsValidSid(sid).as_bool() {
-        return false;
-    }
-    let auth = GetSidIdentifierAuthority(sid);
-    if auth.is_null() {
-        return false;
-    }
-    // NT Authority = {0,0,0,0,0,5}.
-    let a = (*auth).Value;
-    if a != [0, 0, 0, 0, 0, 5] {
-        return false;
-    }
-    let count = *GetSidSubAuthorityCount(sid);
-    if count != 1 {
-        return false;
-    }
-    *GetSidSubAuthority(sid, 0) == 18 // SECURITY_LOCAL_SYSTEM_RID
-}

crates/punktfunk-host/src/config.rs

@@ -6,8 +6,8 @@
 //!
 //! **Goal-1 stages 1–2** (`design/windows-host-rewrite.md` §2.2): stage 1 stood this up; stage 2 migrated the
 //! genuinely-constant operator/dispatch knobs onto it (the dispatch-disagreement bug class: `idd_push`,
-//! `capture_backend`, `encoder_pref`, `render_adapter`, `no_wgc`, the vdisplay backend select — plus the
+//! `encoder_pref`, `render_adapter`, the vdisplay backend select — plus the plan-named
-//! plan-named `secure_dda`/`idd_depth`/`zerocopy`/`ten_bit`/`four_four_four` and the multi-site `perf`/`compositor`/
+//! `idd_depth`/`zerocopy`/`ten_bit`/`four_four_four` and the multi-site `perf`/`compositor`/
 //! `video_source`/`gamepad`). `SessionPlan` (stage 3) consumes it as the single owner of the
 //! capture/topology/encoder decision.
 //!
@@ -36,27 +36,17 @@ use std::sync::OnceLock;
 /// derived `Debug` impl, so the parser can stay a single platform-neutral function.
 #[derive(Debug, Clone, Default)]
 pub struct HostConfig {
-    /// `PUNKTFUNK_IDD_PUSH` — capture from the pf-vdisplay driver's shared ring (in-process Session-0
+    /// `PUNKTFUNK_IDD_PUSH` — IDD direct-push monitor mode (the per-session monitor + ring recreate and
-    /// capture; no WGC helper). **Value-aware** (`0`/`false`/`no`/`off`/empty ⇒ off, else on); unset ⇒ off.
+    /// the discrete-render-GPU pin in [`crate::vdisplay::manager`]). IDD-push is the sole Windows capture
-    /// The installer's default `host.env` sets it on, so a fresh install runs the validated IDD-push path
+    /// path (DXGI Desktop Duplication and the WGC relay were removed), so this should stay on — the
-    /// (it falls back to DDA if the driver can't attach — see [`crate::capture`]). NOT a bare presence flag
+    /// installer's `host.env` sets it. **Value-aware** (`0`/`false`/`no`/`off`/empty ⇒ off, else on);
-    /// (so an operator can turn it OFF in `host.env` with `=0`, which a `var_os` presence check can't).
+    /// unset ⇒ off. NOT a bare presence flag (so an operator can turn it OFF with `=0`).
     pub idd_push: bool,
     /// `PUNKTFUNK_ENCODER` — explicit encoder-backend override (lowercased; empty = auto-detect by GPU vendor).
     pub encoder_pref: String,
-    /// `PUNKTFUNK_NO_HELPER` — never spawn the user-session WGC helper.
-    pub no_helper: bool,
-    /// `PUNKTFUNK_FORCE_HELPER` — force the WGC helper even when not running as SYSTEM.
-    pub force_helper: bool,
-    /// `PUNKTFUNK_NO_WGC` — force the pure single-process DDA path (skip WGC and the two-process relay).
-    pub no_wgc: bool,
-    /// `PUNKTFUNK_CAPTURE` — explicit Windows capture-backend override (lowercased; `dda`/`dxgi` vs the WGC default).
-    pub capture_backend: String,
     /// `PUNKTFUNK_RENDER_ADAPTER` — discrete render-GPU pin by description substring (`Some` even when empty:
     /// the empty string still counts as "set" for the presence checks, and the value reader filters it).
     pub render_adapter: Option<String>,
-    /// `PUNKTFUNK_SECURE_DDA` — enable the experimental DDA-on-secure-desktop (Winlogon/UAC) mux leg.
-    pub secure_dda: bool,
     /// `PUNKTFUNK_IDD_DEPTH` — IDD-push pipeline depth override (default 2; the call site clamps to its `OUT_RING`).
     pub idd_depth: usize,
     /// `PUNKTFUNK_ZEROCOPY` — opt into the Windows D3D11 zero-copy encode path (presence semantics; see module docs).
@@ -103,14 +93,7 @@ impl HostConfig {
             encoder_pref: std::env::var("PUNKTFUNK_ENCODER")
                 .unwrap_or_default()
                 .to_ascii_lowercase(),
-            no_helper: flag("PUNKTFUNK_NO_HELPER"),
-            force_helper: flag("PUNKTFUNK_FORCE_HELPER"),
-            no_wgc: flag("PUNKTFUNK_NO_WGC"),
-            capture_backend: std::env::var("PUNKTFUNK_CAPTURE")
-                .unwrap_or_default()
-                .to_ascii_lowercase(),
             render_adapter: val("PUNKTFUNK_RENDER_ADAPTER"),
-            secure_dda: flag("PUNKTFUNK_SECURE_DDA"),
             idd_depth: val("PUNKTFUNK_IDD_DEPTH")
                 .and_then(|s| s.parse::<usize>().ok())
                 .unwrap_or(2),

crates/punktfunk-host/src/gamestream/gamepad.rs

@@ -66,6 +66,13 @@ pub const BTN_A: u32 = 0x1000;
 pub const BTN_B: u32 = 0x2000;
 pub const BTN_X: u32 = 0x4000;
 pub const BTN_Y: u32 = 0x8000;
+// Extended buttons in the `buttonFlags2 << 16` namespace (mirror `punktfunk_core::input::gamepad`):
+// the four back-grip paddles. `decode` already merges `buttonFlags2 << 16` into `buttons`, but the
+// injector map dropped these bits — Sunshine/Moonlight paddle clients were silently no-op'd.
+pub const BTN_PADDLE1: u32 = 0x0001_0000;
+pub const BTN_PADDLE2: u32 = 0x0002_0000;
+pub const BTN_PADDLE3: u32 = 0x0004_0000;
+pub const BTN_PADDLE4: u32 = 0x0008_0000;
 
 /// Decode one decrypted control plaintext into a controller event, if it is one. Mouse,
 /// keyboard, keepalives etc. yield `None` (they're handled by [`super::input::decode`]).

crates/punktfunk-host/src/gamestream/pairing.rs

@@ -101,6 +101,10 @@ struct Session {
     server_challenge: [u8; 16],
     /// The client's phase-3 hash, recomputed + checked in phase 4.
     client_hash: Vec<u8>,
+    /// Set once phase 3 has produced the RSA-signed serversecret. A repeated phase 3 is refused so a
+    /// peer past phase 1 can't loop phase2/phase3 to harvest many signing-time samples (a passive
+    /// timing-oracle amplifier vs. the rsa-crate Marvin side-channel; see `.cargo/audit.toml`).
+    responded: bool,
 }
 
 pub struct Pairing {
@@ -155,6 +159,7 @@ impl Pairing {
                 serversecret: [0; 16],
                 server_challenge: [0; 16],
                 client_hash: Vec::new(),
+                responded: false,
             },
         );
         tracing::info!(
@@ -216,6 +221,14 @@ impl Pairing {
             bail!("short challenge response");
         }
         s.client_hash = client_hash[..32].to_vec();
+        // Sign the serversecret exactly ONCE per ceremony: refuse a repeated phase 3 so a peer that
+        // cleared phase 1 (operator PIN) can't replay it to collect many RSA signing-time samples
+        // (timing-oracle amplifier vs. RUSTSEC-2023-0071; see `.cargo/audit.toml`). A legit client
+        // signs once. The session stays for phase 4 (the cert-pin step) but won't re-sign.
+        if s.responded {
+            bail!("serverchallengeresp already answered for this pairing session");
+        }
+        s.responded = true;
         let sig: Signature = id.signing_key.sign(&s.serversecret);
         let mut secret = Vec::with_capacity(16 + 256);
         secret.extend_from_slice(&s.serversecret);

crates/punktfunk-host/src/gamestream/stream.rs

@@ -213,6 +213,17 @@ fn open_gs_virtual_source(
     let compositor = if let Some(c) = app.and_then(|a| a.compositor) {
         c
     } else {
+        // Windows has a single virtual-display backend (pf-vdisplay); `vdisplay::open` ignores the
+        // compositor arg there, so short-circuit the Linux session-detection state machine with a
+        // placeholder — mirrors `punktfunk1::resolve_compositor`. Without this, the Linux `detect()`
+        // below bails on Windows ("could not detect compositor … XDG_CURRENT_DESKTOP=''"), which
+        // killed the GameStream video thread → black screen (the native plane was already guarded).
+        #[cfg(target_os = "windows")]
+        {
+            crate::vdisplay::Compositor::Kwin
+        }
+        #[cfg(not(target_os = "windows"))]
+        {
             let active = crate::vdisplay::detect_active_session();
             crate::vdisplay::apply_session_env(&active);
             let c = crate::vdisplay::compositor_for_kind(active.kind)
@@ -221,6 +232,7 @@ fn open_gs_virtual_source(
                 .context("detect compositor")?;
             crate::vdisplay::apply_input_env(c);
             c
+        }
     };
     let mut vd = crate::vdisplay::open(compositor).context("open virtual display")?;
     // Carry the resolved launch command on the backend instance (per-session) rather than a

crates/punktfunk-host/src/inject.rs

@@ -491,6 +491,31 @@ pub mod gamepad;
 #[cfg(target_os = "windows")]
 #[path = "inject/windows/gamepad_raii.rs"]
 mod gamepad_raii;
+/// Linux: virtual Steam Deck via UHID — the kernel `hid-steam` driver binds it as a real Deck.
+#[cfg(target_os = "linux")]
+#[path = "inject/linux/steam_controller.rs"]
+pub mod steam_controller;
+/// Linux: virtual Steam Deck via the USB gadget subsystem (`raw_gadget` + `dummy_hcd`) — the only
+/// virtual-Deck transport Steam Input promotes (presents the controller on USB interface 2).
+/// SteamOS-host only (needs `dummy_hcd` + `raw_gadget`).
+#[cfg(target_os = "linux")]
+#[path = "inject/linux/steam_gadget.rs"]
+pub mod steam_gadget;
+/// Transport-independent Steam Controller / Steam Deck HID contract (descriptor, byte-exact Deck
+/// serializer, XInput/rich mappers, rumble parser), used by the Linux UHID backend ([`steam_controller`]).
+#[cfg(target_os = "linux")]
+#[path = "inject/proto/steam_proto.rs"]
+pub mod steam_proto;
+/// Pure fallback-remap policy (Steam-only inputs onto a non-Steam backend) + the Deck motion rescale.
+#[cfg(target_os = "linux")]
+#[path = "inject/proto/steam_remap.rs"]
+pub mod steam_remap;
+/// Linux: virtual Steam Deck over **USB/IP** (`vhci_hcd`) — the shippable, Secure-Boot-clean,
+/// Steam-Input-promotable virtual-Deck transport on non-SteamOS hosts (Bazzite/generic), where
+/// `dummy_hcd`/`raw_gadget` aren't built. In-tree + signed; no module build, no MOK.
+#[cfg(target_os = "linux")]
+#[path = "inject/linux/steam_usbip.rs"]
+pub mod steam_usbip;
 /// Stub — virtual gamepads need Linux uinput or the Windows UMDF drivers; events are dropped elsewhere.
 #[cfg(not(any(target_os = "linux", target_os = "windows")))]
 pub mod gamepad {

crates/punktfunk-host/src/inject/linux/dualsense.rs

@@ -182,6 +182,9 @@ pub struct DualSenseManager {
     last_write: Vec<Instant>,
     /// Pad creation failed (e.g. /dev/uhid permissions) — warn once, drop events.
     broken: bool,
+    /// Fallback policy for the Steam back grips a client may send (the DualSense has no back-button
+    /// HID slot). `PUNKTFUNK_STEAM_REMAP=paddles=…`; default drop.
+    remap: crate::inject::steam_remap::RemapConfig,
 }
 
 impl Default for DualSenseManager {
@@ -198,6 +201,7 @@ impl DualSenseManager {
             last_rumble: vec![(0, 0); MAX_PADS],
             last_write: vec![Instant::now(); MAX_PADS],
             broken: false,
+            remap: crate::inject::steam_remap::RemapConfig::from_env(),
         }
     }
 
@@ -229,8 +233,12 @@ impl DualSenseManager {
                 // Merge buttons/sticks/triggers from the frame, preserving touch + motion (those
                 // come on the rich-input plane and must survive a button-only frame).
                 let prev = self.state[idx];
+                // Steam back grips have no DualSense slot — fold them onto standard buttons per the
+                // configured policy (default drop) so they aren't silently lost.
+                let buttons =
+                    crate::inject::steam_remap::fold_paddles(f.buttons, self.remap.paddles);
                 let mut s = DsState::from_gamepad(
-                    f.buttons,
+                    buttons,
                     f.ls_x,
                     f.ls_y,
                     f.rs_x,
@@ -252,7 +260,9 @@ impl DualSenseManager {
     /// arrived first); they're dropped if the pad isn't present.
     pub fn apply_rich(&mut self, rich: RichInput) {
         let idx = match rich {
-            RichInput::Touchpad { pad, .. } | RichInput::Motion { pad, .. } => pad as usize,
+            RichInput::Touchpad { pad, .. }
+            | RichInput::Motion { pad, .. }
+            | RichInput::TouchpadEx { pad, .. } => pad as usize,
         };
         if idx >= MAX_PADS || self.pads[idx].is_none() {
             return;
@@ -280,6 +290,26 @@ impl DualSenseManager {
                 self.state[idx].gyro = gyro;
                 self.state[idx].accel = accel;
             }
+            RichInput::TouchpadEx {
+                surface,
+                finger,
+                touch,
+                x,
+                y,
+                ..
+            } => {
+                // A Steam right/single pad maps onto the one DualSense touchpad (signed centre-0 →
+                // 0..=65535); surface 1 (the Steam left pad) has no DualSense equivalent.
+                if surface != 1 {
+                    let slot = (finger as usize).min(1);
+                    let n = |v: i16| ((v as i32) + 32768) as u32;
+                    let t = &mut self.state[idx].touch[slot];
+                    t.active = touch;
+                    t.id = slot as u8;
+                    t.x = (n(x) * (DS_TOUCH_W - 1) as u32 / u16::MAX as u32) as u16;
+                    t.y = (n(y) * (DS_TOUCH_H - 1) as u32 / u16::MAX as u32) as u16;
+                }
+            }
         }
         self.write(idx);
     }

crates/punktfunk-host/src/inject/linux/dualshock4.rs

@@ -367,6 +367,9 @@ pub struct DualShock4Manager {
     last_write: Vec<Instant>,
     /// Pad creation failed (e.g. /dev/uhid permissions) — warn once, drop events.
     broken: bool,
+    /// Fallback policy for the Steam back grips a client may send (the DS4 has no back-button HID
+    /// slot). `PUNKTFUNK_STEAM_REMAP=paddles=…`; default drop.
+    remap: crate::inject::steam_remap::RemapConfig,
 }
 
 impl Default for DualShock4Manager {
@@ -384,6 +387,7 @@ impl DualShock4Manager {
             last_led: vec![None; MAX_PADS],
             last_write: vec![Instant::now(); MAX_PADS],
             broken: false,
+            remap: crate::inject::steam_remap::RemapConfig::from_env(),
         }
     }
 
@@ -416,8 +420,12 @@ impl DualShock4Manager {
                 // Merge buttons/sticks/triggers, preserving touch + motion (those arrive on the
                 // rich-input plane and must survive a button-only frame).
                 let prev = self.state[idx];
+                // Steam back grips have no DS4 slot — fold them onto standard buttons per the
+                // configured policy (default drop) so they aren't silently lost.
+                let buttons =
+                    crate::inject::steam_remap::fold_paddles(f.buttons, self.remap.paddles);
                 let mut s = DsState::from_gamepad(
-                    f.buttons,
+                    buttons,
                     f.ls_x,
                     f.ls_y,
                     f.rs_x,
@@ -439,7 +447,9 @@ impl DualShock4Manager {
     /// pad isn't present.
     pub fn apply_rich(&mut self, rich: RichInput) {
         let idx = match rich {
-            RichInput::Touchpad { pad, .. } | RichInput::Motion { pad, .. } => pad as usize,
+            RichInput::Touchpad { pad, .. }
+            | RichInput::Motion { pad, .. }
+            | RichInput::TouchpadEx { pad, .. } => pad as usize,
         };
         if idx >= MAX_PADS || self.pads[idx].is_none() {
             return;
@@ -466,6 +476,26 @@ impl DualShock4Manager {
                 self.state[idx].gyro = gyro;
                 self.state[idx].accel = accel;
             }
+            RichInput::TouchpadEx {
+                surface,
+                finger,
+                touch,
+                x,
+                y,
+                ..
+            } => {
+                // A Steam right/single pad maps onto the one DS4 touchpad (signed centre-0 →
+                // 0..=65535); surface 1 (the Steam left pad) has no DS4 equivalent.
+                if surface != 1 {
+                    let slot = (finger as usize).min(1);
+                    let n = |v: i16| ((v as i32) + 32768) as u32;
+                    let t = &mut self.state[idx].touch[slot];
+                    t.active = touch;
+                    t.id = slot as u8;
+                    t.x = (n(x) * (DS4_TOUCH_W - 1) as u32 / u16::MAX as u32) as u16;
+                    t.y = (n(y) * (DS4_TOUCH_H - 1) as u32 / u16::MAX as u32) as u16;
+                }
+            }
         }
         self.write(idx);
     }

crates/punktfunk-host/src/inject/linux/gamepad.rs

@@ -69,9 +69,16 @@ const BTN_START: u16 = 0x13b;
 const BTN_MODE: u16 = 0x13c;
 const BTN_THUMBL: u16 = 0x13d;
 const BTN_THUMBR: u16 = 0x13e;
+// Xbox-Elite paddle codes (the xpad convention SDL / Steam Input recognize). A client's back grips —
+// and the GameStream `buttonFlags2` paddle bits, which were silently dropped before — land here, so
+// the virtual X-Box pad exposes paddles like an Elite controller. PADDLE1/2/3/4 = R4/L4/R5/L5.
+const BTN_TRIGGER_HAPPY5: u16 = 0x2c4;
+const BTN_TRIGGER_HAPPY6: u16 = 0x2c5;
+const BTN_TRIGGER_HAPPY7: u16 = 0x2c6;
+const BTN_TRIGGER_HAPPY8: u16 = 0x2c7;
 
 /// `(GameStream button bit, evdev key code)` — D-pad is emitted as HAT axes instead.
-const BUTTON_MAP: [(u32, u16); 11] = [
+const BUTTON_MAP: [(u32, u16); 15] = [
     (gamepad::BTN_A, BTN_SOUTH),
     (gamepad::BTN_B, BTN_EAST),
     (gamepad::BTN_X, BTN_NORTH),
@@ -83,6 +90,10 @@ const BUTTON_MAP: [(u32, u16); 11] = [
     (gamepad::BTN_GUIDE, BTN_MODE),
     (gamepad::BTN_LS_CLK, BTN_THUMBL),
     (gamepad::BTN_RS_CLK, BTN_THUMBR),
+    (gamepad::BTN_PADDLE1, BTN_TRIGGER_HAPPY5),
+    (gamepad::BTN_PADDLE2, BTN_TRIGGER_HAPPY6),
+    (gamepad::BTN_PADDLE3, BTN_TRIGGER_HAPPY7),
+    (gamepad::BTN_PADDLE4, BTN_TRIGGER_HAPPY8),
 ];
 
 /// The USB identity a virtual uinput pad presents. SDL/Steam/Proton key their built-in mapping off

crates/punktfunk-host/src/inject/linux/kwin_fake_input.rs

@@ -7,9 +7,14 @@
 //! which the libei/portal path cannot. We connect as an ordinary Wayland client on the KWin session's
 //! `$WAYLAND_DISPLAY` and translate events into fake-input requests; keyboard keys are raw Linux
 //! evdev codes that KWin resolves through the session's own keymap (no keymap upload, unlike the wlr
-//! virtual-keyboard path), and absolute pointer/touch coordinates are global compositor space — which
+//! virtual-keyboard path), and absolute pointer/touch coordinates are global compositor space.
-//! on a headless box (single per-session virtual output at the origin, scale 1) equals the streamed
+//!
-//! output's pixels.
+//! Global compositor space is *logical* pixels (post display-scaling), which only equals the streamed
+//! output's physical pixels at scale 1. Under a fractional/integer scale the logical edge sits at
+//! `physical / scale`, so feeding the raw streamed pixel coordinate lands the cursor `scale×` too far
+//! toward the bottom-right (top-left stays put). We therefore track each output's logical geometry
+//! (position + size) via `xdg-output` and map the normalized client position into the matching
+//! output's logical rectangle — the same shape the libei backend uses with its EI region.
 
 #![allow(clippy::all, dead_code, non_camel_case_types, non_snake_case, unused)]
 // Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
@@ -18,8 +23,14 @@
 use super::{gs_button_to_evdev, vk_to_evdev, InputEvent, InputInjector};
 use anyhow::{Context, Result};
 use punktfunk_core::input::InputKind;
+use std::time::{Duration, Instant};
+use wayland_client::protocol::wl_output::{self, WlOutput};
 use wayland_client::protocol::wl_registry::{self, WlRegistry};
-use wayland_client::{Connection, Dispatch, EventQueue, Proxy, QueueHandle};
+use wayland_client::{Connection, Dispatch, EventQueue, Proxy, QueueHandle, WEnum};
+use wayland_protocols::xdg::xdg_output::zv1::client::{
+    zxdg_output_manager_v1::ZxdgOutputManagerV1,
+    zxdg_output_v1::{self, ZxdgOutputV1},
+};
 
 // Generate the client bindings for the vendored protocol XML inline (no build.rs), exactly like the
 // KWin virtual-output backend. Path is relative to CARGO_MANIFEST_DIR.
@@ -48,10 +59,39 @@ const AXIS_HORIZONTAL: u32 = 1;
 /// `code` value marking a horizontal scroll event (mirrors `gamestream::input` / the wlr backend).
 const SCROLL_HORIZONTAL: u32 = 1;
 
+/// One tracked output: its physical mode (to match the streamed resolution) and its logical geometry
+/// (the global-compositor-space rectangle absolute coordinates are addressed in). `logical_w == 0`
+/// means xdg-output hasn't reported its size yet.
+struct OutputTrack {
+    /// Registry global id — also the dispatch user-data, so events route back to this entry.
+    name: u32,
+    wl_output: WlOutput,
+    xdg_output: Option<ZxdgOutputV1>,
+    /// Physical pixel mode from `wl_output.mode` (the `current` mode); matched against the streamed WxH.
+    mode_w: i32,
+    mode_h: i32,
+    /// Logical (post-scale) geometry from `xdg-output`.
+    logical_x: i32,
+    logical_y: i32,
+    logical_w: i32,
+    logical_h: i32,
+}
+
 /// Registry-bound globals (the Wayland dispatch state).
 #[derive(Default)]
 struct State {
     fake: Option<FakeInput>,
+    xdg_mgr: Option<ZxdgOutputManagerV1>,
+    outputs: Vec<OutputTrack>,
+}
+
+impl State {
+    /// Create the `xdg_output` for a tracked output once both it and the manager exist.
+    fn ensure_xdg_output(o: &mut OutputTrack, mgr: &ZxdgOutputManagerV1, qh: &QueueHandle<State>) {
+        if o.xdg_output.is_none() {
+            o.xdg_output = Some(mgr.get_xdg_output(&o.wl_output, qh, o.name));
+        }
+    }
 }
 
 impl Dispatch<WlRegistry, ()> for State {
@@ -63,15 +103,57 @@ impl Dispatch<WlRegistry, ()> for State {
         _: &Connection,
         qh: &QueueHandle<Self>,
     ) {
-        if let wl_registry::Event::Global {
+        match event {
+            wl_registry::Event::Global {
                 name,
                 interface,
                 version,
-        } = event
+            } => match interface.as_str() {
-        {
+                "org_kde_kwin_fake_input" => {
-            if interface == "org_kde_kwin_fake_input" {
                     state.fake = Some(registry.bind(name, version.min(MAX_VERSION), qh, ()));
                 }
+                "wl_output" => {
+                    // v1 carries `mode` (all we need); bind no higher than the proxy's max (4).
+                    let wl_output: WlOutput = registry.bind(name, version.min(4), qh, name);
+                    let mut o = OutputTrack {
+                        name,
+                        wl_output,
+                        xdg_output: None,
+                        mode_w: 0,
+                        mode_h: 0,
+                        logical_x: 0,
+                        logical_y: 0,
+                        logical_w: 0,
+                        logical_h: 0,
+                    };
+                    if let Some(mgr) = state.xdg_mgr.clone() {
+                        State::ensure_xdg_output(&mut o, &mgr, qh);
+                    }
+                    state.outputs.push(o);
+                }
+                "zxdg_output_manager_v1" => {
+                    let mgr: ZxdgOutputManagerV1 = registry.bind(name, version.min(3), qh, ());
+                    // Outputs bound before the manager have no xdg_output yet — create them now.
+                    for o in state.outputs.iter_mut() {
+                        State::ensure_xdg_output(o, &mgr, qh);
+                    }
+                    state.xdg_mgr = Some(mgr);
+                }
+                _ => {}
+            },
+            wl_registry::Event::GlobalRemove { name } => {
+                state.outputs.retain(|o| {
+                    if o.name == name {
+                        if let Some(x) = &o.xdg_output {
+                            x.destroy();
+                        }
+                        false
+                    } else {
+                        true
+                    }
+                });
+            }
+            _ => {}
         }
     }
 }
@@ -89,13 +171,86 @@ impl Dispatch<FakeInput, ()> for State {
     }
 }
 
+impl Dispatch<WlOutput, u32> for State {
+    fn event(
+        state: &mut Self,
+        _: &WlOutput,
+        event: wl_output::Event,
+        name: &u32,
+        _: &Connection,
+        _: &QueueHandle<Self>,
+    ) {
+        // Only the *current* mode matters — a real monitor also advertises its other supported modes.
+        if let wl_output::Event::Mode {
+            flags: WEnum::Value(flags),
+            width,
+            height,
+            ..
+        } = event
+        {
+            if flags.contains(wl_output::Mode::Current) {
+                if let Some(o) = state.outputs.iter_mut().find(|o| o.name == *name) {
+                    o.mode_w = width;
+                    o.mode_h = height;
+                }
+            }
+        }
+    }
+}
+
+impl Dispatch<ZxdgOutputV1, u32> for State {
+    fn event(
+        state: &mut Self,
+        _: &ZxdgOutputV1,
+        event: zxdg_output_v1::Event,
+        name: &u32,
+        _: &Connection,
+        _: &QueueHandle<Self>,
+    ) {
+        if let Some(o) = state.outputs.iter_mut().find(|o| o.name == *name) {
+            match event {
+                zxdg_output_v1::Event::LogicalPosition { x, y } => {
+                    o.logical_x = x;
+                    o.logical_y = y;
+                }
+                zxdg_output_v1::Event::LogicalSize { width, height } => {
+                    o.logical_w = width;
+                    o.logical_h = height;
+                }
+                _ => {}
+            }
+        }
+    }
+}
+
+// The manager has no events.
+impl Dispatch<ZxdgOutputManagerV1, ()> for State {
+    fn event(
+        _: &mut Self,
+        _: &ZxdgOutputManagerV1,
+        _: <ZxdgOutputManagerV1 as Proxy>::Event,
+        _: &(),
+        _: &Connection,
+        _: &QueueHandle<Self>,
+    ) {
+    }
+}
+
 pub struct KwinFakeInjector {
     conn: Connection,
     queue: EventQueue<State>,
     state: State,
     fake: FakeInput,
+    /// When output geometry was last re-read; throttles the per-event roundtrip (see `refresh_geometry`).
+    last_refresh: Option<Instant>,
 }
 
+/// How often the fake_input backend re-reads output geometry from the compositor. Output add/remove
+/// (a new session's virtual output) and live scale/resolution changes are infrequent, so a lazy
+/// poll on the injector's own thread is plenty and adds at most one local-socket roundtrip twice a
+/// second — versus a blocking roundtrip on every single mouse-move event.
+const GEO_REFRESH: Duration = Duration::from_millis(500);
+
 impl KwinFakeInjector {
     pub fn open() -> Result<Self> {
         let conn = Connection::connect_to_env()
@@ -122,13 +277,77 @@ impl KwinFakeInjector {
             .context("fake_input authenticate roundtrip")?;
         conn.flush().ok();
 
-        tracing::info!("KWin fake_input ready (headless keyboard/mouse/touch — no portal)");
+        // Settle output geometry (wl_output + xdg-output were bound during the registry roundtrip
-        Ok(Self {
+        // above; their logical_size arrives on a follow-up roundtrip). Best-effort — falls back to
+        // scale-1 mapping if xdg-output is absent.
+        let mut injector = Self {
             conn,
             queue,
             state,
             fake,
-        })
+            last_refresh: None,
+        };
+        injector.refresh_geometry();
+        tracing::info!(
+            outputs = injector.state.outputs.len(),
+            "KWin fake_input ready (headless keyboard/mouse/touch — no portal)"
+        );
+        Ok(injector)
+    }
+
+    /// Re-read output geometry, throttled to [`GEO_REFRESH`]. A `roundtrip` both flushes any pending
+    /// `get_xdg_output` requests and reads the geometry events back. A wl_output that *appeared* this
+    /// round only gets its xdg_output created mid-dispatch, so its `logical_size` lands on a later
+    /// roundtrip — keep going (bounded) until every output is settled.
+    fn refresh_geometry(&mut self) {
+        let now = Instant::now();
+        if let Some(t) = self.last_refresh {
+            if now.duration_since(t) < GEO_REFRESH {
+                return;
+            }
+        }
+        self.last_refresh = Some(now);
+        for _ in 0..3 {
+            if self.queue.roundtrip(&mut self.state).is_err() {
+                return;
+            }
+            let pending =
+                self.state.xdg_mgr.is_some() && self.state.outputs.iter().any(|o| o.logical_w == 0);
+            if !pending {
+                break;
+            }
+        }
+    }
+
+    /// Resolve the logical (global-compositor-space) rectangle to map a normalized client position
+    /// into. Prefer the output whose physical mode matches the streamed `phys_w`×`phys_h` (the
+    /// per-session virtual output); fall back to the sole output, then — if xdg-output is unavailable
+    /// — to the streamed pixels at the origin (the pre-scaling behavior, correct at scale 1).
+    fn logical_target(&self, phys_w: i32, phys_h: i32) -> (f64, f64, f64, f64) {
+        let usable = || {
+            self.state
+                .outputs
+                .iter()
+                .filter(|o| o.logical_w > 0 && o.logical_h > 0)
+        };
+        let chosen = usable()
+            .find(|o| o.mode_w == phys_w && o.mode_h == phys_h)
+            .or_else(|| {
+                let mut it = usable();
+                match (it.next(), it.next()) {
+                    (Some(only), None) => Some(only),
+                    _ => None,
+                }
+            });
+        match chosen {
+            Some(o) => (
+                o.logical_x as f64,
+                o.logical_y as f64,
+                o.logical_w as f64,
+                o.logical_h as f64,
+            ),
+            None => (0.0, 0.0, phys_w as f64, phys_h as f64),
+        }
     }
 }
 
@@ -139,12 +358,17 @@ impl InputInjector for KwinFakeInjector {
                 self.fake.pointer_motion(event.x as f64, event.y as f64);
             }
             InputKind::MouseMoveAbs => {
-                let w = (event.flags >> 16) & 0xffff;
+                let w = ((event.flags >> 16) & 0xffff) as i32;
-                let h = event.flags & 0xffff;
+                let h = (event.flags & 0xffff) as i32;
                 if w > 0 && h > 0 {
-                    let x = event.x.clamp(0, w as i32) as f64;
+                    self.refresh_geometry();
-                    let y = event.y.clamp(0, h as i32) as f64;
+                    let (lx, ly, lw, lh) = self.logical_target(w, h);
-                    self.fake.pointer_motion_absolute(x, y);
+                    // Normalize in the streamed (physical) pixel space, then place inside the output's
+                    // logical rectangle — so display scaling no longer offsets the cursor.
+                    let nx = (event.x as f64 / w as f64).clamp(0.0, 1.0);
+                    let ny = (event.y as f64 / h as f64).clamp(0.0, 1.0);
+                    self.fake
+                        .pointer_motion_absolute(lx + nx * lw, ly + ny * lh);
                 }
             }
             InputKind::MouseButtonDown | InputKind::MouseButtonUp => {
@@ -179,11 +403,15 @@ impl InputInjector for KwinFakeInjector {
             // Touch: id = event.code, coords in the client surface w×h packed into flags (same
             // absolute mapping as MouseMoveAbs). Each event is its own frame.
             InputKind::TouchDown | InputKind::TouchMove => {
-                let w = (event.flags >> 16) & 0xffff;
+                let w = ((event.flags >> 16) & 0xffff) as i32;
-                let h = event.flags & 0xffff;
+                let h = (event.flags & 0xffff) as i32;
                 if w > 0 && h > 0 {
-                    let x = event.x.clamp(0, w as i32) as f64;
+                    self.refresh_geometry();
-                    let y = event.y.clamp(0, h as i32) as f64;
+                    let (lx, ly, lw, lh) = self.logical_target(w, h);
+                    let nx = (event.x as f64 / w as f64).clamp(0.0, 1.0);
+                    let ny = (event.y as f64 / h as f64).clamp(0.0, 1.0);
+                    let x = lx + nx * lw;
+                    let y = ly + ny * lh;
                     if event.kind == InputKind::TouchDown {
                         self.fake.touch_down(event.code, x, y);
                     } else {

crates/punktfunk-host/src/inject/linux/steam_controller.rs

@@ -0,0 +1,575 @@
+//! Virtual Steam Deck controller via UHID — the Steam analogue of the virtual DualSense
+//! ([`super::dualsense`]). A UHID device with Valve VID `28DE` / Deck PID `1205` is bound by the
+//! kernel `hid-steam` driver, which exposes a full Steam Deck gamepad evdev (incl. the four back
+//! grips) **plus** a separate IMU evdev, and — when Steam runs on the host — is re-grabbed by Steam
+//! Input with native glyphs + trackpad/gyro/back-button bindings.
+//!
+//! The transport-independent contract (descriptor, byte-exact serializer, the `XInput`/rich
+//! mappers, the rumble parser) lives in [`super::steam_proto`]; this module is the `/dev/uhid`
+//! plumbing + the two Steam-specific lifecycle quirks the DualSense path lacks:
+//!
+//! 1. **`gamepad_mode` entry.** `steam_do_deck_input_event` early-returns under the default
+//!    `lizard_mode` until `gamepad_mode` is toggled on — which the kernel only does when the `b9.6`
+//!    Steam/menu-right button is held ~450 ms with no hidraw client open. So on the first pad we
+//!    best-effort clear `lizard_mode` via sysfs (needs root; bypasses the gate entirely) AND every
+//!    pad pulses `b9.6` for [`MODE_ENTER`] at creation. After that an **anti-toggle guard** caps any
+//!    continuous `b9.6` (a long in-game Start-hold) below the kernel's 450 ms threshold so play can
+//!    never accidentally flip `gamepad_mode` back off.
+//! 2. **`UHID_SET_REPORT`.** Steam feedback (`0xEB` rumble) + the kernel's settings/serial writes
+//!    arrive as FEATURE set-reports that MUST be answered `err = 0`, or the kernel stalls ~5 s per
+//!    command (the DualSense backend only services GET_REPORT + OUTPUT).
+
+use super::steam_proto::{
+    btn, parse_steam_output, serial_reply, serialize_deck_state, SteamState, STEAMDECK_PRODUCT,
+    STEAMDECK_RDESC, STEAM_REPORT_LEN, STEAM_VENDOR,
+};
+use crate::gamestream::gamepad::{GamepadEvent, MAX_PADS};
+use anyhow::{Context, Result};
+use punktfunk_core::quic::{HidOutput, RichInput};
+use std::fs::{File, OpenOptions};
+use std::io::{Read, Write};
+use std::os::unix::fs::OpenOptionsExt;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::time::{Duration, Instant};
+
+// /dev/uhid event ABI — same layout as the DualSense backend.
+const UHID_PATH: &str = "/dev/uhid";
+const UHID_DESTROY: u32 = 1;
+const UHID_OUTPUT: u32 = 6;
+const UHID_GET_REPORT: u32 = 9;
+const UHID_GET_REPORT_REPLY: u32 = 10;
+const UHID_CREATE2: u32 = 11;
+const UHID_INPUT2: u32 = 12;
+const UHID_SET_REPORT: u32 = 13;
+const UHID_SET_REPORT_REPLY: u32 = 14;
+const HID_MAX_DESCRIPTOR_SIZE: usize = 4096;
+const UHID_EVENT_SIZE: usize = 4 + 4372;
+const BUS_USB: u16 = 0x03;
+
+/// Hold the `b9.6` mode-switch this long at creation to toggle `gamepad_mode` on (the kernel needs
+/// ~450 ms continuous; give margin).
+const MODE_ENTER: Duration = Duration::from_millis(650);
+/// Cap continuous `b9.6` (Start) below the kernel's 450 ms mode-switch threshold: after this long
+/// we insert a one-frame release so an in-game long-Start-hold can't toggle `gamepad_mode` off.
+const MENU_HOLD_CAP: Duration = Duration::from_millis(350);
+
+fn put_cstr(ev: &mut [u8], off: usize, cap: usize, s: &str) {
+    let n = s.len().min(cap - 1);
+    ev[off..off + n].copy_from_slice(&s.as_bytes()[..n]);
+}
+
+/// Best-effort, once per process: clear `hid_steam`'s `lizard_mode` so `steam_do_deck_input_event`
+/// stops gating on `gamepad_mode` (gamepad events then always flow). Needs root; on failure the
+/// per-pad `b9.6` pulse + guard handle it instead.
+fn try_clear_lizard_mode() {
+    static TRIED: AtomicBool = AtomicBool::new(false);
+    if TRIED.swap(true, Ordering::Relaxed) {
+        return;
+    }
+    match std::fs::write("/sys/module/hid_steam/parameters/lizard_mode", "N") {
+        Ok(()) => {
+            tracing::info!("cleared hid_steam lizard_mode (Steam Deck gamepad events always flow)")
+        }
+        Err(e) => tracing::debug!(
+            error = %e,
+            "could not clear hid_steam lizard_mode (no root?) — using the gamepad_mode pulse + guard"
+        ),
+    }
+}
+
+/// A virtual Steam Deck backed by `/dev/uhid`. Dropping it destroys the device (the kernel tears
+/// down the bound `hid-steam` interface + both evdevs).
+pub struct SteamDeckPad {
+    fd: File,
+    seq: u32,
+    created: Instant,
+    /// When `b9.6` started being continuously held in our OUTPUT (anti-toggle guard); `None` = not.
+    menu_hold_since: Option<Instant>,
+}
+
+impl SteamDeckPad {
+    pub fn open(index: u8) -> Result<SteamDeckPad> {
+        try_clear_lizard_mode();
+        let fd = OpenOptions::new()
+            .read(true)
+            .write(true)
+            .custom_flags(libc::O_NONBLOCK)
+            .open(UHID_PATH)
+            .with_context(|| {
+                format!("open {UHID_PATH} (is the uhid udev rule installed + are you in 'input'?)")
+            })?;
+        let mut pad = SteamDeckPad {
+            fd,
+            seq: 0,
+            created: Instant::now(),
+            menu_hold_since: None,
+        };
+        pad.send_create2(index).context("UHID_CREATE2 Steam Deck")?;
+        Ok(pad)
+    }
+
+    fn send_create2(&mut self, index: u8) -> Result<()> {
+        let mut ev = [0u8; UHID_EVENT_SIZE];
+        ev[0..4].copy_from_slice(&UHID_CREATE2.to_ne_bytes());
+        put_cstr(&mut ev, 4, 128, &format!("Punktfunk Steam Deck {index}")); // name[128]
+        put_cstr(&mut ev, 132, 64, &format!("punktfunk/steam/{index}")); // phys[64]
+        put_cstr(&mut ev, 196, 64, &format!("punktfunk-steam-{index}")); // uniq[64]
+        ev[260..262].copy_from_slice(&(STEAMDECK_RDESC.len() as u16).to_ne_bytes()); // rd_size
+        ev[262..264].copy_from_slice(&BUS_USB.to_ne_bytes()); // bus
+        ev[264..268].copy_from_slice(&STEAM_VENDOR.to_ne_bytes());
+        ev[268..272].copy_from_slice(&STEAMDECK_PRODUCT.to_ne_bytes());
+        ev[272..276].copy_from_slice(&0x0100u32.to_ne_bytes()); // version
+        ev[276..280].copy_from_slice(&0u32.to_ne_bytes()); // country
+        ev[280..280 + STEAMDECK_RDESC.len()].copy_from_slice(STEAMDECK_RDESC);
+        self.fd.write_all(&ev).context("write UHID_CREATE2")?;
+        Ok(())
+    }
+
+    /// Serialize `st` (with the gamepad-mode entry overlay + anti-toggle guard applied) and write it.
+    pub fn write_state(&mut self, st: &SteamState) -> Result<()> {
+        self.seq = self.seq.wrapping_add(1);
+        let mut s = *st;
+        s.buttons = self.effective_buttons(st.buttons);
+        let mut r = [0u8; STEAM_REPORT_LEN];
+        serialize_deck_state(&mut r, &s, self.seq);
+
+        let mut ev = [0u8; UHID_EVENT_SIZE];
+        ev[0..4].copy_from_slice(&UHID_INPUT2.to_ne_bytes());
+        ev[4..6].copy_from_slice(&(r.len() as u16).to_ne_bytes()); // input2.size
+        ev[6..6 + r.len()].copy_from_slice(&r); // input2.data
+        self.fd.write_all(&ev).context("write UHID_INPUT2")?;
+        Ok(())
+    }
+
+    /// True while still pulsing the mode-switch at creation (the caller force-writes during this).
+    fn in_mode_entry(&self) -> bool {
+        self.created.elapsed() < MODE_ENTER
+    }
+
+    /// During mode entry, force `b9.6` held (override). Afterwards, pass the real buttons through but
+    /// drop `b9.6` for one frame whenever it's been continuously held past [`MENU_HOLD_CAP`].
+    fn effective_buttons(&mut self, mut buttons: u64) -> u64 {
+        if self.in_mode_entry() {
+            return btn::STEAM_MENU_RIGHT;
+        }
+        if buttons & btn::MENU != 0 {
+            let now = Instant::now();
+            match self.menu_hold_since {
+                None => self.menu_hold_since = Some(now),
+                Some(since) if now.duration_since(since) >= MENU_HOLD_CAP => {
+                    buttons &= !btn::MENU; // one-frame release resets the kernel's mode-switch timer
+                    self.menu_hold_since = None;
+                }
+                Some(_) => {}
+            }
+        } else {
+            self.menu_hold_since = None;
+        }
+        buttons
+    }
+
+    /// Service the device, non-blocking: answer the kernel's GET_REPORT (serial) + SET_REPORT
+    /// (settings / rumble — ack `err=0`) and parse any rumble feedback (`0xEB`, on either the
+    /// SET_REPORT or OUTPUT path) into `(low, high)` for the universal rumble plane.
+    pub fn service(&mut self) -> Option<(u16, u16)> {
+        let mut rumble = None;
+        let mut ev = [0u8; UHID_EVENT_SIZE];
+        while let Ok(n) = self.fd.read(&mut ev) {
+            if n < UHID_EVENT_SIZE {
+                break;
+            }
+            match u32::from_ne_bytes([ev[0], ev[1], ev[2], ev[3]]) {
+                UHID_OUTPUT => {
+                    let size = u16::from_ne_bytes([ev[4100], ev[4101]]) as usize;
+                    let end = 4 + size.min(HID_MAX_DESCRIPTOR_SIZE);
+                    if let Some(r) = parse_steam_output(&ev[4..end]).rumble {
+                        rumble = Some(r);
+                    }
+                }
+                UHID_GET_REPORT => {
+                    let id = u32::from_ne_bytes([ev[4], ev[5], ev[6], ev[7]]);
+                    let _ = self.reply_get_report(id, &serial_reply("PUNKTFUNK01"));
+                }
+                UHID_SET_REPORT => {
+                    let id = u32::from_ne_bytes([ev[4], ev[5], ev[6], ev[7]]);
+                    // SET_REPORT data: [report-id 0, cmd, …] at ev[12..]. Surface rumble, then ack.
+                    let end = (12 + 16).min(UHID_EVENT_SIZE);
+                    if let Some(r) = parse_steam_output(&ev[12..end]).rumble {
+                        rumble = Some(r);
+                    }
+                    let _ = self.reply_set_report(id);
+                }
+                _ => {} // Start/Stop/Open/Close — ignore
+            }
+        }
+        rumble
+    }
+
+    fn reply_get_report(&mut self, id: u32, data: &[u8]) -> Result<()> {
+        let mut ev = [0u8; UHID_EVENT_SIZE];
+        ev[0..4].copy_from_slice(&UHID_GET_REPORT_REPLY.to_ne_bytes());
+        ev[4..8].copy_from_slice(&id.to_ne_bytes());
+        ev[8..10].copy_from_slice(&0u16.to_ne_bytes()); // err 0
+        ev[10..12].copy_from_slice(&(data.len() as u16).to_ne_bytes());
+        ev[12..12 + data.len()].copy_from_slice(data);
+        self.fd.write_all(&ev).context("UHID_GET_REPORT_REPLY")?;
+        Ok(())
+    }
+
+    fn reply_set_report(&mut self, id: u32) -> Result<()> {
+        let mut ev = [0u8; UHID_EVENT_SIZE];
+        ev[0..4].copy_from_slice(&UHID_SET_REPORT_REPLY.to_ne_bytes());
+        ev[4..8].copy_from_slice(&id.to_ne_bytes());
+        ev[8..10].copy_from_slice(&0u16.to_ne_bytes()); // err 0 (ack)
+        self.fd.write_all(&ev).context("UHID_SET_REPORT_REPLY")?;
+        Ok(())
+    }
+}
+
+impl Drop for SteamDeckPad {
+    fn drop(&mut self) {
+        let mut ev = [0u8; UHID_EVENT_SIZE];
+        ev[0..4].copy_from_slice(&UHID_DESTROY.to_ne_bytes());
+        let _ = self.fd.write_all(&ev);
+    }
+}
+
+/// All virtual Steam Deck pads of a session — the Steam analogue of
+/// [`DualSenseManager`](super::dualsense::DualSenseManager), selected with `PUNKTFUNK_GAMEPAD=steamdeck`.
+/// Button/stick frames arrive via [`handle`](Self::handle); the right trackpad + motion via
+/// [`apply_rich`](Self::apply_rich); [`pump`](Self::pump) services the kernel handshake + routes
+/// rumble back; [`heartbeat`](Self::heartbeat) keeps the pad alive (and drives the mode-entry pulse).
+/// The transport a manager pad drives. UHID is universal but Steam Input won't promote it (a UHID
+/// device has no USB interface number, `Interface: -1`); the USB **gadget** (`raw_gadget`, SteamOS)
+/// and **usbip** (`vhci_hcd`, universal) both present the controller on USB interface 2, which Steam
+/// Input *does* promote. Selected per-pad by [`open_transport`].
+enum DeckTransport {
+    Uhid(SteamDeckPad),
+    Gadget(crate::inject::steam_gadget::SteamDeckGadget),
+    Usbip(crate::inject::steam_usbip::SteamDeckUsbip),
+}
+
+impl DeckTransport {
+    fn write_state(&mut self, st: &SteamState) {
+        match self {
+            DeckTransport::Uhid(p) => {
+                let _ = p.write_state(st);
+            }
+            DeckTransport::Gadget(g) => g.write_state(st),
+            DeckTransport::Usbip(u) => u.write_state(st),
+        }
+    }
+    fn service(&mut self) -> Option<(u16, u16)> {
+        match self {
+            DeckTransport::Uhid(p) => p.service(),
+            DeckTransport::Gadget(g) => g.service().rumble,
+            DeckTransport::Usbip(u) => u.service().rumble,
+        }
+    }
+    fn in_mode_entry(&self) -> bool {
+        match self {
+            // Only the UHID pad needs the gamepad-mode entry pulse: the promoted transports are
+            // read raw via hidraw by Steam Input, which bypasses the kernel's evdev mode gate.
+            DeckTransport::Uhid(p) => p.in_mode_entry(),
+            DeckTransport::Gadget(_) | DeckTransport::Usbip(_) => false,
+        }
+    }
+}
+
+/// Open the best Steam-Input-promotable Deck transport available, in preference order:
+/// **`raw_gadget` (SteamOS validated fast-path) → `usbip`/`vhci_hcd` (universal, Secure-Boot-clean)
+/// → UHID (universal, but `Interface: -1` so Steam Input won't promote it).** Each rung degrades to
+/// the next on failure, so a host lacking the gadget modules still gets a *promotable* Deck via
+/// usbip, and one lacking both still gets a working (if non-promoted) UHID pad.
+fn open_transport(idx: u8) -> Result<DeckTransport> {
+    use crate::inject::{steam_gadget, steam_usbip};
+    // 1. raw_gadget — the validated SteamOS fast-path (default on there).
+    if steam_gadget::gadget_preferred() {
+        steam_gadget::ensure_modules();
+        match steam_gadget::SteamDeckGadget::open(idx) {
+            Ok(g) => {
+                tracing::info!(
+                    index = idx,
+                    "virtual Steam Deck created (USB gadget — Steam Input recognizes it)"
+                );
+                return Ok(DeckTransport::Gadget(g));
+            }
+            Err(e) => {
+                tracing::warn!(error = %format!("{e:#}"), "USB-gadget Deck unavailable — trying usbip")
+            }
+        }
+    }
+    // 2. usbip/vhci_hcd — the universal, in-tree, Secure-Boot-clean transport (default on elsewhere).
+    if steam_usbip::usbip_preferred() {
+        match steam_usbip::SteamDeckUsbip::open(idx) {
+            Ok(u) => return Ok(DeckTransport::Usbip(u)),
+            Err(e) => {
+                tracing::warn!(error = %format!("{e:#}"), "usbip Deck unavailable — falling back to UHID")
+            }
+        }
+    }
+    // 3. UHID — universal fallback (works everywhere; Steam Input won't promote it).
+    let p = SteamDeckPad::open(idx)?;
+    tracing::info!(
+        index = idx,
+        "virtual Steam Deck created (UHID hid-steam — not Steam-Input-promoted)"
+    );
+    Ok(DeckTransport::Uhid(p))
+}
+
+pub struct SteamControllerManager {
+    pads: Vec<Option<DeckTransport>>,
+    state: Vec<SteamState>,
+    last_rumble: Vec<(u16, u16)>,
+    last_write: Vec<Instant>,
+    broken: bool,
+}
+
+impl Default for SteamControllerManager {
+    fn default() -> SteamControllerManager {
+        SteamControllerManager::new()
+    }
+}
+
+impl SteamControllerManager {
+    pub fn new() -> SteamControllerManager {
+        SteamControllerManager {
+            pads: (0..MAX_PADS).map(|_| None).collect(),
+            state: vec![SteamState::neutral(); MAX_PADS],
+            last_rumble: vec![(0, 0); MAX_PADS],
+            last_write: vec![Instant::now(); MAX_PADS],
+            broken: false,
+        }
+    }
+
+    pub fn handle(&mut self, ev: &GamepadEvent) {
+        match ev {
+            GamepadEvent::Arrival { index, kind, .. } => {
+                tracing::info!(index, kind, "controller arrival (Steam Deck)");
+                self.ensure(*index as usize);
+            }
+            GamepadEvent::State(f) => {
+                let idx = f.index as usize;
+                if idx >= MAX_PADS {
+                    return;
+                }
+                for (i, slot) in self.pads.iter_mut().enumerate() {
+                    if slot.is_some() && f.active_mask & (1 << i) == 0 {
+                        tracing::info!(index = i, "controller unplugged (Steam Deck)");
+                        *slot = None;
+                        self.state[i] = SteamState::neutral();
+                        self.last_rumble[i] = (0, 0);
+                    }
+                }
+                if f.active_mask & (1 << idx) == 0 {
+                    return;
+                }
+                self.ensure(idx);
+                // Merge buttons/sticks/triggers, preserving the rich-plane fields (trackpad + motion
+                // arrive separately and must survive a button-only frame).
+                let prev = self.state[idx];
+                let mut s = SteamState::from_gamepad(
+                    f.buttons,
+                    f.ls_x,
+                    f.ls_y,
+                    f.rs_x,
+                    f.rs_y,
+                    f.left_trigger,
+                    f.right_trigger,
+                );
+                s.rpad_x = prev.rpad_x;
+                s.rpad_y = prev.rpad_y;
+                s.lpad_x = prev.lpad_x;
+                s.lpad_y = prev.lpad_y;
+                s.gyro = prev.gyro;
+                s.accel = prev.accel;
+                s.buttons |= prev.buttons & (btn::RPAD_TOUCH | btn::LPAD_TOUCH);
+                self.state[idx] = s;
+                self.write(idx);
+            }
+        }
+    }
+
+    /// Apply a rich client→host event (right trackpad / motion) to an existing pad.
+    pub fn apply_rich(&mut self, rich: RichInput) {
+        let idx = match rich {
+            RichInput::Touchpad { pad, .. }
+            | RichInput::Motion { pad, .. }
+            | RichInput::TouchpadEx { pad, .. } => pad as usize,
+        };
+        if idx >= MAX_PADS || self.pads[idx].is_none() {
+            return;
+        }
+        self.state[idx].apply_rich(rich);
+        self.write(idx);
+    }
+
+    fn write(&mut self, idx: usize) {
+        let st = self.state[idx];
+        if let Some(pad) = self.pads[idx].as_mut() {
+            pad.write_state(&st);
+        }
+        self.last_write[idx] = Instant::now();
+    }
+
+    /// Re-emit each live pad's current report when silent past `max_gap`, and force a steady stream
+    /// while a pad is still pulsing its gamepad-mode entry (so the `b9.6` toggle completes even with
+    /// no game input).
+    pub fn heartbeat(&mut self, max_gap: Duration) {
+        let now = Instant::now();
+        for i in 0..self.pads.len() {
+            let Some(pad) = self.pads[i].as_ref() else {
+                continue;
+            };
+            if pad.in_mode_entry() || now.duration_since(self.last_write[i]) >= max_gap {
+                self.write(i);
+            }
+        }
+    }
+
+    fn ensure(&mut self, idx: usize) {
+        if idx >= MAX_PADS || self.pads[idx].is_some() || self.broken {
+            return;
+        }
+        match open_transport(idx as u8) {
+            Ok(t) => {
+                self.pads[idx] = Some(t);
+                self.state[idx] = SteamState::neutral();
+                self.last_rumble[idx] = (0, 0);
+                self.last_write[idx] = Instant::now();
+            }
+            Err(e) => {
+                tracing::error!(error = %format!("{e:#}"), "virtual Steam Deck creation failed — controller input disabled");
+                self.broken = true;
+            }
+        }
+    }
+
+    /// Service every pad: answer the kernel handshake and forward rumble on the universal plane.
+    /// `rumble` fires `(index, low, high)` only on a level change. The Steam Deck has no rich
+    /// host→client feedback plane (no lightbar / adaptive triggers), so `hidout` goes unused.
+    pub fn pump(&mut self, mut rumble: impl FnMut(u16, u16, u16), _hidout: impl FnMut(HidOutput)) {
+        for i in 0..self.pads.len() {
+            let Some(pad) = self.pads[i].as_mut() else {
+                continue;
+            };
+            if let Some(r) = pad.service() {
+                if self.last_rumble[i] != r {
+                    self.last_rumble[i] = r;
+                    rumble(i as u16, r.0, r.1);
+                }
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Find the evdev node for a kernel input device by exact name (e.g. `"Steam Deck"`).
+    fn find_node(name: &str) -> Option<String> {
+        let devs = std::fs::read_to_string("/proc/bus/input/devices").ok()?;
+        for block in devs.split("\n\n") {
+            if !block
+                .lines()
+                .any(|l| l.trim() == format!("N: Name=\"{name}\""))
+            {
+                continue;
+            }
+            for l in block.lines() {
+                if let Some(h) = l.strip_prefix("H: Handlers=") {
+                    if let Some(ev) = h.split_whitespace().find(|t| t.starts_with("event")) {
+                        return Some(format!("/dev/input/{ev}"));
+                    }
+                }
+            }
+        }
+        None
+    }
+
+    /// Read the evdev's current key bitmap (`EVIOCGKEY`) and test whether `code` is down.
+    fn key_is_down(node: &str, code: u16) -> bool {
+        use std::os::unix::io::AsRawFd;
+        let Ok(f) = std::fs::File::open(node) else {
+            return false;
+        };
+        let mut bits = [0u8; 96];
+        const EVIOCGKEY: libc::c_ulong = (2 << 30) | (96 << 16) | (0x45 << 8) | 0x18;
+        // SAFETY: EVIOCGKEY copies the current key-state bitmap of the evdev behind the valid fd
+        // `f` into `bits`; 96 bytes covers KEY_MAX/8, so the kernel never writes past the buffer.
+        let rc = unsafe { libc::ioctl(f.as_raw_fd(), EVIOCGKEY, bits.as_mut_ptr()) };
+        rc >= 0 && (bits[(code / 8) as usize] >> (code % 8)) & 1 == 1
+    }
+
+    /// Read the current value of an absolute axis (`EVIOCGABS`) — the first `i32` of `input_absinfo`.
+    fn abs_value(node: &str, abs: u16) -> Option<i32> {
+        use std::os::unix::io::AsRawFd;
+        let f = std::fs::File::open(node).ok()?;
+        let mut info = [0u8; 24]; // struct input_absinfo { value, min, max, fuzz, flat, resolution }
+        let req: libc::c_ulong =
+            (2 << 30) | (24 << 16) | (0x45 << 8) | (0x40 + abs as libc::c_ulong);
+        // SAFETY: EVIOCGABS fills the 24-byte input_absinfo for the valid evdev fd `f`; we read only
+        // the leading i32 `value`. The buffer is exactly sizeof(input_absinfo), so no overflow.
+        let rc = unsafe { libc::ioctl(f.as_raw_fd(), req, info.as_mut_ptr()) };
+        (rc >= 0).then(|| i32::from_ne_bytes([info[0], info[1], info[2], info[3]]))
+    }
+
+    /// On-box smoke test for the real backend: a `SteamDeckPad` must bind `hid-steam` (creating both
+    /// the gamepad + IMU evdevs), enter `gamepad_mode` via the creation pulse, and land a held button
+    /// on the evdev (`BTN_A`, code 0x130) — proving the entry overlay + byte-exact serialize path —
+    /// then tear the device down on drop. Touches `/dev/uhid`, so it is `#[ignore]`d in CI; run on a
+    /// box with `hid-steam` + `input`-group access: `cargo test -p punktfunk-host -- --ignored`.
+    #[test]
+    #[ignore = "creates a real /dev/uhid device; needs hid-steam + the input group"]
+    fn backend_binds_and_input_flows() {
+        use punktfunk_core::input::gamepad as gs;
+        const BTN_A: u16 = 0x130;
+        const ABS_HAT0X: u16 = 0x10; // left trackpad X
+        let mut pad = SteamDeckPad::open(0).expect("open SteamDeckPad (/dev/uhid + input group?)");
+        // Drive the full M3 wire path: build state through `from_gamepad` (BTN_A + the L4 back grip)
+        // and `apply_rich` (a left-pad TouchpadEx contact), then hold it past MODE_ENTER (the b9.6
+        // pulse), servicing the handshake.
+        let mut st = SteamState::from_gamepad(gs::BTN_A | gs::BTN_PADDLE2, 0, 0, 0, 0, 0, 0);
+        st.apply_rich(RichInput::TouchpadEx {
+            pad: 0,
+            surface: 1,
+            finger: 0,
+            touch: true,
+            click: false,
+            x: -8000,
+            y: 9000,
+            pressure: 0,
+        });
+        let start = Instant::now();
+        while start.elapsed() < Duration::from_millis(1200) {
+            let _ = pad.service();
+            pad.write_state(&st).expect("write_state");
+            std::thread::sleep(Duration::from_millis(4));
+        }
+        let devs = std::fs::read_to_string("/proc/bus/input/devices").unwrap_or_default();
+        assert!(devs.contains("Steam Deck"), "gamepad evdev not created");
+        assert!(
+            devs.contains("Steam Deck Motion Sensors"),
+            "IMU evdev not created"
+        );
+        let node = find_node("Steam Deck").expect("gamepad evdev node");
+        assert!(
+            key_is_down(&node, BTN_A),
+            "BTN_A not down — gamepad_mode entry or serialize failed"
+        );
+        // The left trackpad contact (TouchpadEx surface 1, gated on LPAD_TOUCH) reaches ABS_HAT0X.
+        assert_eq!(
+            abs_value(&node, ABS_HAT0X),
+            Some(-8000),
+            "left trackpad (TouchpadEx surface 1) did not reach ABS_HAT0X"
+        );
+        drop(pad);
+        std::thread::sleep(Duration::from_millis(200));
+        let devs = std::fs::read_to_string("/proc/bus/input/devices").unwrap_or_default();
+        assert!(
+            !devs.contains("Steam Deck Motion Sensors"),
+            "device not torn down on drop"
+        );
+    }
+}

crates/punktfunk-host/src/inject/linux/steam_gadget.rs

@@ -0,0 +1,558 @@
+//! Virtual Steam Deck via the USB **gadget** subsystem (`raw_gadget` + `dummy_hcd`) — the only
+//! virtual-Deck transport Steam Input recognizes.
+//!
+//! The UHID [`super::steam_controller::SteamDeckPad`] binds the kernel `hid-steam` driver, but Steam's
+//! own controller driver filters the Deck's controller to USB **interface 2**, and a UHID device has no
+//! USB interface number (`Interface: -1`), so Steam enumerates it but never promotes it. This backend
+//! instead presents a *real* 3-interface USB Deck (mouse = interface 0, keyboard = 1, **controller =
+//! 2**) on a `dummy_hcd` loopback UDC, driven from userspace via `/dev/raw-gadget` so we can answer
+//! every control transfer (including the HID feature reports `f_hid` can't). Proven on a real Deck:
+//! hid-steam binds it, Steam reserves an XInput slot and emits an X-Box pad. Descriptors are captured
+//! verbatim from a physical Deck; see `packaging/linux/steam-deck-gadget/` for the original PoC + the
+//! USB-stack gotchas. **SteamOS-host only** (needs `dummy_hcd` + `raw_gadget`, which SteamOS ships).
+//!
+//! The transport here is self-contained (libc + std); the report bytes it streams are produced by
+//! [`super::steam_proto`] in the wrapping backend.
+
+use anyhow::{bail, Context, Result};
+use std::mem::size_of;
+use std::os::fd::RawFd;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{Arc, Mutex};
+use std::thread::JoinHandle;
+
+// ---- raw_gadget UAPI (mirrors linux/usb/raw_gadget.h; inlined like the C PoC) ----
+const UDC_NAME_MAX: usize = 128;
+
+#[repr(C)]
+struct UsbRawInit {
+    driver_name: [u8; UDC_NAME_MAX],
+    device_name: [u8; UDC_NAME_MAX],
+    speed: u8,
+}
+
+// usb_raw_event { u32 type; u32 length; u8 data[]; } — we read it into a fixed buffer.
+const EVENT_HDR: usize = 8; // type + length
+const EVENT_BUF: usize = EVENT_HDR + 64; // setup packet (8) fits easily
+
+// usb_raw_ep_io { u16 ep; u16 flags; u32 length; u8 data[]; }
+const EPIO_HDR: usize = 8;
+
+// usb_endpoint_descriptor is 9 bytes in the kernel (audio bRefresh/bSynchAddress); EP_ENABLE wants it.
+#[repr(C, packed)]
+#[derive(Clone, Copy, Default)]
+struct UsbEndpointDescriptor {
+    b_length: u8,
+    b_descriptor_type: u8,
+    b_endpoint_address: u8,
+    bm_attributes: u8,
+    w_max_packet_size: u16,
+    b_interval: u8,
+    b_refresh: u8,
+    b_synch_address: u8,
+}
+
+const fn ioc(dir: u64, nr: u64, size: usize) -> libc::c_ulong {
+    ((dir << 30) | ((size as u64) << 16) | ((b'U' as u64) << 8) | nr) as libc::c_ulong
+}
+const IOCTL_INIT: libc::c_ulong = ioc(1, 0, size_of::<UsbRawInit>());
+const IOCTL_RUN: libc::c_ulong = ioc(0, 1, 0);
+const IOCTL_EVENT_FETCH: libc::c_ulong = ioc(2, 2, EVENT_HDR); // size is the header; kernel copies more
+const IOCTL_EP0_WRITE: libc::c_ulong = ioc(1, 3, EPIO_HDR);
+const IOCTL_EP0_READ: libc::c_ulong = ioc(2 | 1, 4, EPIO_HDR); // _IOWR
+const IOCTL_EP_ENABLE: libc::c_ulong = ioc(1, 5, size_of::<UsbEndpointDescriptor>());
+const IOCTL_EP_WRITE: libc::c_ulong = ioc(1, 7, EPIO_HDR);
+const IOCTL_CONFIGURE: libc::c_ulong = ioc(0, 9, 0);
+const IOCTL_VBUS_DRAW: libc::c_ulong = ioc(1, 10, 4);
+const IOCTL_EP0_STALL: libc::c_ulong = ioc(0, 12, 0);
+
+const USB_RAW_EVENT_CONNECT: u32 = 1;
+const USB_RAW_EVENT_CONTROL: u32 = 2;
+const USB_SPEED_HIGH: u8 = 3;
+
+// Captured-from-hardware Deck descriptors + the `0x83`/`0xAE` feature contract live in the shared
+// [`super::steam_proto`] module (single source of truth, also used by the usbip transport).
+use super::steam_proto::{
+    deck_serial, deck_unit_id, feature_reply, neutral_deck_report, RDESC_DECK_CTRL as RDESC_CTRL,
+    RDESC_DECK_KBD as RDESC_KBD, RDESC_DECK_MOUSE as RDESC_MOUSE,
+};
+
+const DEV_DESC: [u8; 18] = [
+    18, 1, 0x00, 0x02, // bLength, DEVICE, bcdUSB 2.00
+    0, 0, 0, 64, // class/sub/proto, bMaxPacketSize0
+    0xDE, 0x28, 0x05, 0x12, // idVendor 28DE, idProduct 1205
+    0x00, 0x03, // bcdDevice 3.00
+    1, 2, 3, 1, // iManufacturer, iProduct, iSerial, bNumConfigurations
+];
+
+const HID_DT: u8 = 0x21;
+const HID_RPT_DT: u8 = 0x22;
+
+/// Assemble the 84-byte config descriptor: config + 3×(interface + HID + 7-byte endpoint).
+fn build_config() -> Vec<u8> {
+    let mut c = Vec::with_capacity(84);
+    // config descriptor (wTotalLength patched after)
+    c.extend_from_slice(&[9, 2, 84, 0, 3, 1, 0, 0x80, 250]);
+    // helper closures
+    let iface = |n: u8, sub: u8, proto: u8| [9u8, 4, n, 0, 1, 3, sub, proto, 0];
+    let hid = |rlen: u16, country: u8| {
+        [
+            9u8,
+            HID_DT,
+            0x10,
+            0x01,
+            country,
+            1,
+            HID_RPT_DT,
+            (rlen & 0xff) as u8,
+            (rlen >> 8) as u8,
+        ]
+    };
+    let ep = |addr: u8, mps: u16| [7u8, 5, addr, 0x03, (mps & 0xff) as u8, (mps >> 8) as u8, 4];
+    // interface 0: mouse, EP 0x81
+    c.extend_from_slice(&iface(0, 0, 2));
+    c.extend_from_slice(&hid(RDESC_MOUSE.len() as u16, 0));
+    c.extend_from_slice(&ep(0x81, 8));
+    // interface 1: keyboard (boot), EP 0x82
+    c.extend_from_slice(&iface(1, 1, 1));
+    c.extend_from_slice(&hid(RDESC_KBD.len() as u16, 0));
+    c.extend_from_slice(&ep(0x82, 8));
+    // interface 2: controller, EP 0x83, bCountryCode 33
+    c.extend_from_slice(&iface(2, 0, 0));
+    c.extend_from_slice(&hid(RDESC_CTRL.len() as u16, 33));
+    c.extend_from_slice(&ep(0x83, 64));
+    debug_assert_eq!(c.len(), 84);
+    c
+}
+
+fn string_desc(idx: u8, serial: &str) -> Vec<u8> {
+    if idx == 0 {
+        return vec![4, 3, 0x09, 0x04]; // LANGID en-US
+    }
+    let s: &str = match idx {
+        1 => "Valve Software",
+        2 => "Steam Deck Controller",
+        3 => serial,
+        _ => "",
+    };
+    let mut v = vec![(2 + s.len() * 2) as u8, 3];
+    for ch in s.encode_utf16() {
+        v.push((ch & 0xff) as u8);
+        v.push((ch >> 8) as u8);
+    }
+    v
+}
+
+// ---- ioctl wrappers (the only unsafe surface for the raw_gadget UAPI; documented once) ----
+fn ioctl_ptr<T>(fd: RawFd, req: libc::c_ulong, arg: *const T) -> i32 {
+    // SAFETY: `fd` is our open /dev/raw-gadget descriptor; `arg` points to a correctly-sized,
+    // initialized argument for `req` (a raw_gadget UAPI struct or an owned usb_raw_ep_io buffer)
+    // that lives for the duration of the call. `ioctl` is variadic, so passing a thin pointer is ABI-correct.
+    unsafe { libc::ioctl(fd, req as _, arg) as i32 }
+}
+fn ioctl_mut<T>(fd: RawFd, req: libc::c_ulong, arg: *mut T) -> i32 {
+    // SAFETY: as `ioctl_ptr`, but `arg` is a writable buffer the kernel fills for `req` (EVENT_FETCH / EP0_READ).
+    unsafe { libc::ioctl(fd, req as _, arg) as i32 }
+}
+fn ioctl_val(fd: RawFd, req: libc::c_ulong, val: libc::c_ulong) -> i32 {
+    // SAFETY: `req` (VBUS_DRAW) takes an integer argument by value; `fd` is our descriptor.
+    unsafe { libc::ioctl(fd, req as _, val) as i32 }
+}
+fn ioctl_none(fd: RawFd, req: libc::c_ulong) -> i32 {
+    // SAFETY: `req` (RUN / CONFIGURE / EP0_STALL) takes no argument, but raw_gadget rejects a non-zero
+    // `value` with EINVAL — pass an explicit 0 (an omitted vararg would be an indeterminate register).
+    unsafe { libc::ioctl(fd, req as _, 0) as i32 }
+}
+
+// ---- low-level ep0 helpers (operate on the shared fd) ----
+fn ep0_write(fd: RawFd, data: &[u8]) -> i32 {
+    let mut buf = vec![0u8; EPIO_HDR + data.len()];
+    buf[0..2].copy_from_slice(&0u16.to_ne_bytes()); // ep 0
+    buf[4..8].copy_from_slice(&(data.len() as u32).to_ne_bytes());
+    buf[EPIO_HDR..].copy_from_slice(data);
+    ioctl_ptr(fd, IOCTL_EP0_WRITE, buf.as_ptr())
+}
+fn ep0_read(fd: RawFd, len: usize) -> (i32, Vec<u8>) {
+    let mut buf = vec![0u8; EPIO_HDR + len.max(1)];
+    buf[4..8].copy_from_slice(&(len as u32).to_ne_bytes());
+    let r = ioctl_mut(fd, IOCTL_EP0_READ, buf.as_mut_ptr());
+    let n = if r > 0 { r as usize } else { 0 };
+    (r, buf[EPIO_HDR..EPIO_HDR + n.min(len.max(1))].to_vec())
+}
+/// Complete a no-data OUT control (status stage is an IN, handled by a zero-length read).
+fn ep0_ack(fd: RawFd) {
+    ep0_read(fd, 0);
+}
+fn ep0_stall(fd: RawFd) {
+    ioctl_none(fd, IOCTL_EP0_STALL);
+}
+
+/// Owns the `/dev/raw-gadget` fd; closing it tears the device down.
+struct GadgetFd(RawFd);
+impl Drop for GadgetFd {
+    fn drop(&mut self) {
+        // SAFETY: `self.0` is the fd we opened in `SteamDeckGadget::open` and own uniquely here.
+        unsafe { libc::close(self.0) };
+    }
+}
+
+/// A virtual Steam Deck presented over the USB gadget subsystem. Dropping it stops the threads and
+/// closes the gadget (the kernel tears down the device).
+pub struct SteamDeckGadget {
+    report: Arc<Mutex<[u8; 64]>>,
+    feedback: Arc<Mutex<super::steam_proto::SteamFeedback>>,
+    running: Arc<AtomicBool>,
+    threads: Vec<JoinHandle<()>>,
+    _fd: Arc<GadgetFd>,
+    seq: u32,
+}
+
+impl SteamDeckGadget {
+    /// Bind a virtual Deck on a fresh `dummy_hcd` UDC. `index` only varies the serial. Requires
+    /// `dummy_hcd` + `raw_gadget` loaded and write access to `/dev/raw-gadget` (root on SteamOS).
+    pub fn open(index: u8) -> Result<SteamDeckGadget> {
+        // SAFETY: opening a constant NUL-terminated device path with O_RDWR; returns a fd or -1.
+        let fd = unsafe { libc::open(c"/dev/raw-gadget".as_ptr(), libc::O_RDWR) };
+        if fd < 0 {
+            bail!(
+                "open /dev/raw-gadget ({}) — is raw_gadget+dummy_hcd loaded and are we root?",
+                std::io::Error::last_os_error()
+            );
+        }
+        let fd = Arc::new(GadgetFd(fd));
+        let raw = fd.0;
+
+        // INIT against the dummy UDC, then RUN.
+        // SAFETY: `UsbRawInit` is a plain-old-data struct (byte arrays + u8); all-zero is a valid value.
+        let mut init: UsbRawInit = unsafe { std::mem::zeroed() };
+        copy_cstr(&mut init.driver_name, "dummy_udc");
+        copy_cstr(&mut init.device_name, "dummy_udc.0");
+        init.speed = USB_SPEED_HIGH;
+        if ioctl_ptr(raw, IOCTL_INIT, &init as *const _) < 0 {
+            bail!("raw_gadget INIT: {}", std::io::Error::last_os_error());
+        }
+        if ioctl_none(raw, IOCTL_RUN) < 0 {
+            bail!("raw_gadget RUN: {}", std::io::Error::last_os_error());
+        }
+
+        let serial = deck_serial(index);
+        let unit_id = deck_unit_id(index); // "PF" + index — a synthetic per-instance device id
+        let report = Arc::new(Mutex::new(neutral_deck_report()));
+        let feedback = Arc::new(Mutex::new(Default::default()));
+        let running = Arc::new(AtomicBool::new(true));
+        let ctrl_ep = Arc::new(std::sync::atomic::AtomicI32::new(-1));
+        let configured = Arc::new(AtomicBool::new(false));
+
+        // Control thread: enumerate + answer every control transfer.
+        let control = {
+            let fd = fd.clone();
+            let running = running.clone();
+            let ctrl_ep = ctrl_ep.clone();
+            let configured = configured.clone();
+            let feedback = feedback.clone();
+            std::thread::Builder::new()
+                .name("pf-deck-gadget-ctrl".into())
+                .spawn(move || {
+                    control_loop(fd, running, ctrl_ep, configured, feedback, serial, unit_id)
+                })
+                .context("spawn gadget control thread")?
+        };
+        // Stream thread: push the current report on the controller interrupt-IN endpoint.
+        let stream = {
+            let fd = fd.clone();
+            let running = running.clone();
+            let ctrl_ep = ctrl_ep.clone();
+            let configured = configured.clone();
+            let report = report.clone();
+            std::thread::Builder::new()
+                .name("pf-deck-gadget-stream".into())
+                .spawn(move || stream_loop(fd, running, ctrl_ep, configured, report))
+                .context("spawn gadget stream thread")?
+        };
+
+        Ok(SteamDeckGadget {
+            report,
+            feedback,
+            running,
+            threads: vec![control, stream],
+            _fd: fd,
+            seq: 0,
+        })
+    }
+
+    /// Serialize `st` into the 64-byte Deck state report streamed to the kernel.
+    pub fn write_state(&mut self, st: &super::steam_proto::SteamState) {
+        self.seq = self.seq.wrapping_add(1);
+        let mut r = [0u8; 64];
+        super::steam_proto::serialize_deck_state(&mut r, st, self.seq);
+        if let Ok(mut g) = self.report.lock() {
+            *g = r;
+        }
+    }
+
+    /// Drain any feedback (rumble) the kernel/Steam wrote to the device.
+    pub fn service(&mut self) -> super::steam_proto::SteamFeedback {
+        self.feedback
+            .lock()
+            .map(|mut f| std::mem::take(&mut *f))
+            .unwrap_or_default()
+    }
+}
+
+impl Drop for SteamDeckGadget {
+    fn drop(&mut self) {
+        self.running.store(false, Ordering::SeqCst);
+        for t in self.threads.drain(..) {
+            let _ = t.join();
+        }
+    }
+}
+
+fn copy_cstr(dst: &mut [u8], s: &str) {
+    let n = s.len().min(dst.len() - 1);
+    dst[..n].copy_from_slice(&s.as_bytes()[..n]);
+}
+
+fn control_loop(
+    fd: Arc<GadgetFd>,
+    running: Arc<AtomicBool>,
+    ctrl_ep: Arc<std::sync::atomic::AtomicI32>,
+    configured: Arc<AtomicBool>,
+    feedback: Arc<Mutex<super::steam_proto::SteamFeedback>>,
+    serial: String,
+    unit_id: u32,
+) {
+    let raw = fd.0;
+    let cfg = build_config();
+    let mut last_set: Vec<u8> = Vec::new();
+    let mut evbuf = [0u8; EVENT_BUF];
+    while running.load(Ordering::SeqCst) {
+        // EVENT_FETCH: type(4) length(4) data[].
+        evbuf[4..8].copy_from_slice(&(8u32).to_ne_bytes()); // request setup-sized payload
+        let r = ioctl_mut(raw, IOCTL_EVENT_FETCH, evbuf.as_mut_ptr());
+        if r < 0 {
+            if running.load(Ordering::SeqCst) {
+                // transient; brief backoff
+                std::thread::sleep(std::time::Duration::from_millis(2));
+            }
+            continue;
+        }
+        let etype = u32::from_ne_bytes([evbuf[0], evbuf[1], evbuf[2], evbuf[3]]);
+        match etype {
+            USB_RAW_EVENT_CONNECT => {}
+            USB_RAW_EVENT_CONTROL => {
+                let s = &evbuf[EVENT_HDR..EVENT_HDR + 8];
+                let ctrl = Setup {
+                    bm_request_type: s[0],
+                    b_request: s[1],
+                    w_value: u16::from_le_bytes([s[2], s[3]]),
+                    w_index: u16::from_le_bytes([s[4], s[5]]),
+                    w_length: u16::from_le_bytes([s[6], s[7]]),
+                };
+                handle_control(
+                    raw,
+                    &ctrl,
+                    &cfg,
+                    &serial,
+                    unit_id,
+                    &ctrl_ep,
+                    &configured,
+                    &mut last_set,
+                    &feedback,
+                );
+            }
+            _ => {}
+        }
+    }
+}
+
+struct Setup {
+    bm_request_type: u8,
+    b_request: u8,
+    w_value: u16,
+    w_index: u16,
+    w_length: u16,
+}
+
+#[allow(clippy::too_many_arguments)]
+fn handle_control(
+    raw: RawFd,
+    ctrl: &Setup,
+    cfg: &[u8],
+    serial: &str,
+    unit_id: u32,
+    ctrl_ep: &std::sync::atomic::AtomicI32,
+    configured: &AtomicBool,
+    last_set: &mut Vec<u8>,
+    feedback: &Mutex<super::steam_proto::SteamFeedback>,
+) {
+    let idx = (ctrl.w_index & 0xff) as u8;
+    let type_class = ctrl.bm_request_type & 0x60;
+    let wl = ctrl.w_length as usize;
+    if type_class == 0x00 {
+        // standard
+        match ctrl.b_request {
+            0x06 => {
+                // GET_DESCRIPTOR
+                let dt = (ctrl.w_value >> 8) as u8;
+                let di = (ctrl.w_value & 0xff) as u8;
+                let resp: Vec<u8> = match dt {
+                    1 => DEV_DESC.to_vec(),
+                    2 => cfg.to_vec(),
+                    3 => string_desc(di, serial),
+                    HID_RPT_DT => match idx {
+                        0 => RDESC_MOUSE.to_vec(),
+                        1 => RDESC_KBD.to_vec(),
+                        _ => RDESC_CTRL.to_vec(),
+                    },
+                    HID_DT => {
+                        // re-emit the interface's HID descriptor from the config blob (best effort)
+                        hid_desc_for(cfg, idx)
+                    }
+                    _ => {
+                        ep0_stall(raw);
+                        return;
+                    }
+                };
+                let n = resp.len().min(wl);
+                ep0_write(raw, &resp[..n]);
+            }
+            0x09 => {
+                // SET_CONFIGURATION
+                ioctl_val(raw, IOCTL_VBUS_DRAW, 0x32);
+                ioctl_none(raw, IOCTL_CONFIGURE);
+                enable_endpoints(raw, ctrl_ep);
+                ep0_ack(raw);
+                configured.store(true, Ordering::SeqCst);
+            }
+            0x0b => ep0_ack(raw), // SET_INTERFACE
+            0x00 => {
+                let st = 0u16;
+                ep0_write(raw, &st.to_le_bytes());
+            }
+            _ => ep0_stall(raw),
+        }
+    } else if type_class == 0x20 {
+        // HID class
+        match ctrl.b_request {
+            0x01 => {
+                // GET_REPORT — serve the Deck feature reply for the last requested command.
+                let resp = feature_reply(last_set, serial, unit_id);
+                let n = resp.len().min(wl);
+                ep0_write(raw, &resp[..n]);
+            }
+            0x09 => {
+                // SET_REPORT — read the host's data; remember it + extract feedback.
+                let (r, data) = ep0_read(raw, wl);
+                if r > 0 {
+                    *last_set = data.clone();
+                    // parse_steam_output expects [report-id(0), cmd, …]; EP0 OUT data is [cmd, …].
+                    let mut framed = Vec::with_capacity(data.len() + 1);
+                    framed.push(0);
+                    framed.extend_from_slice(&data);
+                    let fb = super::steam_proto::parse_steam_output(&framed);
+                    if fb.rumble.is_some() {
+                        if let Ok(mut g) = feedback.lock() {
+                            *g = fb;
+                        }
+                    }
+                }
+            }
+            0x0a | 0x0b => ep0_ack(raw), // SET_IDLE / SET_PROTOCOL
+            0x03 => {
+                ep0_write(raw, &[0u8]);
+            } // GET_PROTOCOL
+            _ => ep0_stall(raw),
+        }
+    } else {
+        ep0_stall(raw);
+    }
+}
+
+fn hid_desc_for(cfg: &[u8], idx: u8) -> Vec<u8> {
+    // The HID descriptors live right after each interface descriptor in the config blob.
+    // Offsets: cfg(9) | i0(9) h0(9) e0(7) | i1(9) h1(9) e1(7) | i2(9) h2(9) e2(7)
+    let off = match idx {
+        0 => 9 + 9,
+        1 => 9 + 25 + 9,
+        _ => 9 + 50 + 9,
+    };
+    cfg.get(off..off + 9)
+        .map(|s| s.to_vec())
+        .unwrap_or_default()
+}
+
+fn enable_endpoints(raw: RawFd, ctrl_ep: &std::sync::atomic::AtomicI32) {
+    let mk = |addr: u8, mps: u16| UsbEndpointDescriptor {
+        b_length: 7,
+        b_descriptor_type: 5,
+        b_endpoint_address: addr,
+        bm_attributes: 0x03,
+        w_max_packet_size: mps,
+        b_interval: 4,
+        ..Default::default()
+    };
+    let e0 = mk(0x81, 8);
+    let e1 = mk(0x82, 8);
+    let e2 = mk(0x83, 64);
+    ioctl_ptr(raw, IOCTL_EP_ENABLE, &e0 as *const _);
+    ioctl_ptr(raw, IOCTL_EP_ENABLE, &e1 as *const _);
+    let h2 = ioctl_ptr(raw, IOCTL_EP_ENABLE, &e2 as *const _);
+    ctrl_ep.store(h2, Ordering::SeqCst);
+}
+
+fn stream_loop(
+    fd: Arc<GadgetFd>,
+    running: Arc<AtomicBool>,
+    ctrl_ep: Arc<std::sync::atomic::AtomicI32>,
+    configured: Arc<AtomicBool>,
+    report: Arc<Mutex<[u8; 64]>>,
+) {
+    let raw = fd.0;
+    while running.load(Ordering::SeqCst) {
+        let ep = ctrl_ep.load(Ordering::SeqCst);
+        if configured.load(Ordering::SeqCst) && ep >= 0 {
+            let r = report
+                .lock()
+                .map(|g| *g)
+                .unwrap_or_else(|_| neutral_deck_report());
+            let mut buf = [0u8; EPIO_HDR + 64];
+            buf[0..2].copy_from_slice(&(ep as u16).to_ne_bytes());
+            buf[4..8].copy_from_slice(&(64u32).to_ne_bytes());
+            buf[EPIO_HDR..].copy_from_slice(&r);
+            // Blocks until the host polls the interrupt-IN endpoint; that's fine on its own thread.
+            ioctl_ptr(raw, IOCTL_EP_WRITE, buf.as_ptr());
+        }
+        std::thread::sleep(std::time::Duration::from_millis(8));
+    }
+}
+
+/// Best-effort load of the gadget modules (SteamOS ships `dummy_hcd` + `raw_gadget`). Failures are
+/// ignored — the caller falls back to UHID if `/dev/raw-gadget` is then still unusable.
+pub fn ensure_modules() {
+    for m in ["dummy_hcd", "raw_gadget"] {
+        let _ = std::process::Command::new("modprobe").arg(m).status();
+    }
+}
+
+/// Whether to prefer the USB-gadget Deck over the UHID `SteamDeckPad` — the only transport Steam Input
+/// promotes (validated glass-to-glass on a Deck). Defaults **on for SteamOS** hosts (which ship the
+/// gadget modules + run Steam Input); off elsewhere, where the universal UHID path stays the default.
+/// `PUNKTFUNK_STEAM_GADGET=1`/`0` forces it on/off. A Deck-as-host with a *physical* Deck never reaches
+/// here: `resolve_gamepad`'s conflict gate degrades `SteamDeck` → DualSense before the manager is built.
+pub fn gadget_preferred() -> bool {
+    if let Ok(v) = std::env::var("PUNKTFUNK_STEAM_GADGET") {
+        return v == "1" || v.eq_ignore_ascii_case("true");
+    }
+    is_steamos()
+}
+
+/// True on SteamOS-class hosts (`/etc/os-release` `ID=steamos`, or `ID_LIKE` naming it).
+fn is_steamos() -> bool {
+    std::fs::read_to_string("/etc/os-release")
+        .map(|s| {
+            s.lines()
+                .any(|l| l == "ID=steamos" || (l.starts_with("ID_LIKE=") && l.contains("steamos")))
+        })
+        .unwrap_or(false)
+}

crates/punktfunk-host/src/inject/linux/steam_usbip.rs

@@ -0,0 +1,733 @@
+//! Virtual Steam Deck over **USB/IP** (`vhci_hcd`) — the shippable, Secure-Boot-clean, universal
+//! alternative to [`super::steam_gadget`] (`raw_gadget` + `dummy_hcd`, SteamOS-only).
+//!
+//! Like the gadget, this presents a *real* 3-interface USB Steam Deck (mouse = interface 0, keyboard
+//! = 1, **controller = 2**) — the interface-2 layout Steam's own driver filters on, so Steam Input
+//! promotes it (a UHID Deck, `Interface: -1`, never is). Unlike the gadget it needs no out-of-tree
+//! module: `vhci_hcd` is in-tree + signed on SteamOS, Bazzite, and ~every distro, loads under Secure
+//! Boot, and needs no MOK. A userspace [`usbip_sim`] server emulates the Deck; the local `vhci_hcd`
+//! attaches it. **Validated on Bazzite**: `vhci_hcd` enumerates the 3-interface Deck, `hid-steam`
+//! binds it, and Steam reserves an XInput slot — identical recognition to the gadget.
+//!
+//! The device model + the USB/IP protocol come from the vendored [`usbip_sim`] crate (the upstream
+//! `usbip` crate trimmed of its libusb host mode); the captured descriptors + the `0x83`/`0xAE`
+//! feature contract come from the shared [`super::steam_proto`] (one source of truth with the gadget).
+//!
+//! **Attach** is in-process by default (no external `usbip` CLI dependency — the production goal): we
+//! run the emulation server on a loopback TCP port, connect to it ourselves, perform the
+//! `OP_REQ_IMPORT` handshake, then hand the connected socket fd to `vhci_hcd` via its sysfs `attach`
+//! file. If anything in that path fails we fall back to the widely-packaged `usbip` CLI; if *that*
+//! also fails, [`open`](SteamDeckUsbip::open) returns `Err` and the caller degrades to UHID.
+
+use super::steam_proto::{
+    deck_serial, deck_unit_id, feature_reply, neutral_deck_report, parse_steam_output,
+    SteamFeedback, SteamState, RDESC_DECK_CTRL, RDESC_DECK_KBD, RDESC_DECK_MOUSE,
+};
+use anyhow::{bail, Context, Result};
+use std::any::Any;
+use std::collections::HashSet;
+use std::io::{Read, Write};
+use std::net::TcpStream;
+use std::os::fd::AsRawFd;
+use std::path::{Path, PathBuf};
+use std::process::Command;
+use std::sync::{Arc, Mutex};
+use std::thread::JoinHandle;
+use std::time::{Duration, Instant};
+use usbip_sim::{
+    Direction, SetupPacket, UsbDevice, UsbEndpoint, UsbInterface, UsbInterfaceHandler, UsbIpServer,
+    Version,
+};
+
+const STEAM_VENDOR: u16 = 0x28DE;
+const STEAMDECK_PRODUCT: u16 = 0x1205;
+/// The single device's USB/IP bus id (one device per server, so the fixed default is fine).
+const BUS_ID: &str = "0-0-0";
+/// The usbip default TCP port — the server must listen here for the `usbip` CLI fallback to attach.
+const USBIP_TCP_PORT: u16 = 3240;
+
+/// Build the 9-byte HID class descriptor inserted between the interface and endpoint descriptors.
+fn hid_desc(report_len: usize, country: u8) -> Vec<u8> {
+    let l = report_len as u16;
+    #[rustfmt::skip]
+    let d = vec![0x09, 0x21, 0x10, 0x01, country, 1, 0x22, (l & 0xff) as u8, (l >> 8) as u8];
+    d
+}
+
+/// The Deck **controller** interface (vendor HID, interface 2): answers the HID feature reports
+/// (descriptor / `0x83` attributes / `0xAE` serial), streams the current 64-byte state on the
+/// interrupt-IN endpoint, and surfaces rumble written via SET_REPORT.
+#[derive(Debug)]
+struct ControllerHandler {
+    /// The current 64-byte Deck input report, shared with [`SteamDeckUsbip::write_state`].
+    report: Arc<Mutex<[u8; 64]>>,
+    /// Rumble extracted from the kernel's SET_REPORTs, drained by [`SteamDeckUsbip::service`].
+    feedback: Arc<Mutex<SteamFeedback>>,
+    /// The host's last SET_REPORT command (drives [`feature_reply`]).
+    last_set: Vec<u8>,
+    serial: String,
+    unit_id: u32,
+}
+
+impl UsbInterfaceHandler for ControllerHandler {
+    fn get_class_specific_descriptor(&self) -> Vec<u8> {
+        hid_desc(RDESC_DECK_CTRL.len(), 33)
+    }
+    fn handle_urb(
+        &mut self,
+        _interface: &UsbInterface,
+        ep: UsbEndpoint,
+        _len: u32,
+        setup: SetupPacket,
+        req: &[u8],
+    ) -> std::io::Result<Vec<u8>> {
+        if ep.is_ep0() {
+            Ok(match (setup.request_type, setup.request) {
+                // GET report descriptor (standard, interface recipient).
+                (0x81, 0x06) if (setup.value >> 8) == 0x22 => RDESC_DECK_CTRL.to_vec(),
+                // HID GET_REPORT (feature) — the Deck `0x83`/`0xAE` contract.
+                (0xA1, 0x01) => feature_reply(&self.last_set, &self.serial, self.unit_id).to_vec(),
+                // HID SET_REPORT — remember the command (for the next feature reply) + surface rumble.
+                (0x21, 0x09) => {
+                    self.last_set = req.to_vec();
+                    // `parse_steam_output` expects `[report-id(0), cmd, …]`; EP0 OUT data is `[cmd, …]`.
+                    let mut framed = Vec::with_capacity(req.len() + 1);
+                    framed.push(0);
+                    framed.extend_from_slice(req);
+                    let fb = parse_steam_output(&framed);
+                    if fb.rumble.is_some() {
+                        if let Ok(mut g) = self.feedback.lock() {
+                            *g = fb;
+                        }
+                    }
+                    vec![]
+                }
+                (0x21, 0x0A) | (0x21, 0x0B) => vec![], // SET_IDLE / SET_PROTOCOL
+                _ => vec![],
+            })
+        } else if let Direction::In = ep.direction() {
+            // Interrupt-IN poll: return the current report. The vendored sim paces interrupt-IN by
+            // bInterval (vhci_hcd does NOT throttle the server side), so this isn't a busy spin.
+            let r = self
+                .report
+                .lock()
+                .map(|g| *g)
+                .unwrap_or_else(|_| neutral_deck_report());
+            Ok(r.to_vec())
+        } else {
+            Ok(vec![])
+        }
+    }
+    fn as_any(&mut self) -> &mut dyn Any {
+        self
+    }
+}
+
+/// A minimal idle HID interface (mouse / keyboard) — serves only its report descriptor.
+#[derive(Debug)]
+struct IdleHidHandler {
+    report_desc: Vec<u8>,
+}
+impl UsbInterfaceHandler for IdleHidHandler {
+    fn get_class_specific_descriptor(&self) -> Vec<u8> {
+        hid_desc(self.report_desc.len(), 0)
+    }
+    fn handle_urb(
+        &mut self,
+        _i: &UsbInterface,
+        ep: UsbEndpoint,
+        _l: u32,
+        setup: SetupPacket,
+        _req: &[u8],
+    ) -> std::io::Result<Vec<u8>> {
+        if ep.is_ep0() && setup.request == 0x06 && (setup.value >> 8) == 0x22 {
+            Ok(self.report_desc.clone())
+        } else {
+            Ok(vec![])
+        }
+    }
+    fn as_any(&mut self) -> &mut dyn Any {
+        self
+    }
+}
+
+fn boxed(
+    h: impl UsbInterfaceHandler + Send + 'static,
+) -> Arc<Mutex<Box<dyn UsbInterfaceHandler + Send>>> {
+    Arc::new(Mutex::new(Box::new(h)))
+}
+fn ep(addr: u8, mps: u16) -> UsbEndpoint {
+    UsbEndpoint {
+        address: addr,
+        attributes: 0x03, // interrupt
+        max_packet_size: mps,
+        interval: 4,
+    }
+}
+
+/// Assemble the simulated 3-interface USB Deck. The controller handler shares `report` + `feedback`
+/// with the owning [`SteamDeckUsbip`].
+fn build_device(
+    index: u8,
+    report: &Arc<Mutex<[u8; 64]>>,
+    feedback: &Arc<Mutex<SteamFeedback>>,
+) -> UsbDevice {
+    let mut dev = UsbDevice::new(0); // one device per server; bus_id stays the default "0-0-0".
+    dev.vendor_id = STEAM_VENDOR;
+    dev.product_id = STEAMDECK_PRODUCT;
+    dev.usb_version = Version::from(0x0200u16); // bcdUSB 2.00
+    dev.device_bcd = Version::from(0x0300u16); // bcdDevice 3.00 (matches the gadget)
+    dev.set_manufacturer_name("Valve Software");
+    dev.set_product_name("Steam Deck Controller");
+    dev.set_serial_number(&deck_serial(index));
+    dev.with_interface(
+        0x03,
+        0x00,
+        0x02,
+        Some("mouse"),
+        vec![ep(0x81, 8)],
+        boxed(IdleHidHandler {
+            report_desc: RDESC_DECK_MOUSE.to_vec(),
+        }),
+    )
+    .with_interface(
+        0x03,
+        0x01,
+        0x01,
+        Some("keyboard"),
+        vec![ep(0x82, 8)],
+        boxed(IdleHidHandler {
+            report_desc: RDESC_DECK_KBD.to_vec(),
+        }),
+    )
+    .with_interface(
+        0x03,
+        0x00,
+        0x00,
+        Some("controller"),
+        vec![ep(0x83, 64)],
+        boxed(ControllerHandler {
+            report: report.clone(),
+            feedback: feedback.clone(),
+            last_set: vec![],
+            serial: deck_serial(index),
+            unit_id: deck_unit_id(index),
+        }),
+    )
+}
+
+/// Owns the emulation-server thread (a dedicated current-thread tokio runtime) and stops it on drop.
+/// Run on its own thread so `SteamDeckUsbip::open` works whether or not the caller is inside a tokio
+/// runtime (creating a runtime inside one would panic).
+struct ServerThread {
+    stop: Arc<tokio::sync::Notify>,
+    join: Option<JoinHandle<()>>,
+}
+
+impl ServerThread {
+    /// Spawn the server on `listener`, serving exactly the one simulated `dev`.
+    fn spawn(listener: std::net::TcpListener, dev: UsbDevice) -> Result<ServerThread> {
+        let stop = Arc::new(tokio::sync::Notify::new());
+        let stop_t = stop.clone();
+        let join = std::thread::Builder::new()
+            .name("pf-deck-usbip".into())
+            .spawn(move || {
+                let rt = match tokio::runtime::Builder::new_current_thread()
+                    .enable_all()
+                    .build()
+                {
+                    Ok(rt) => rt,
+                    Err(e) => {
+                        tracing::error!(error = %e, "usbip server runtime build failed");
+                        return;
+                    }
+                };
+                rt.block_on(run_server(
+                    listener,
+                    Arc::new(UsbIpServer::new_simulated(vec![dev])),
+                    stop_t,
+                ));
+            })
+            .context("spawn usbip server thread")?;
+        Ok(ServerThread {
+            stop,
+            join: Some(join),
+        })
+    }
+}
+
+impl Drop for ServerThread {
+    fn drop(&mut self) {
+        self.stop.notify_one();
+        if let Some(j) = self.join.take() {
+            let _ = j.join();
+        }
+    }
+}
+
+/// Accept loop: serve each USB/IP connection with the vendored `usbip_sim::handler` until stopped.
+async fn run_server(
+    listener: std::net::TcpListener,
+    server: Arc<UsbIpServer>,
+    stop: Arc<tokio::sync::Notify>,
+) {
+    let listener = match tokio::net::TcpListener::from_std(listener) {
+        Ok(l) => l,
+        Err(e) => {
+            tracing::error!(error = %e, "usbip TcpListener::from_std failed");
+            return;
+        }
+    };
+    loop {
+        tokio::select! {
+            _ = stop.notified() => break,
+            r = listener.accept() => match r {
+                Ok((mut sock, _)) => {
+                    let server = server.clone();
+                    tokio::spawn(async move {
+                        let _ = usbip_sim::handler(&mut sock, server).await;
+                    });
+                }
+                Err(e) => {
+                    tracing::warn!(error = %e, "usbip accept error");
+                    break;
+                }
+            }
+        }
+    }
+}
+
+/// A virtual Steam Deck presented over USB/IP. Dropping it detaches the `vhci_hcd` port (the device
+/// disappears, Steam releases its slot) and stops the emulation server.
+pub struct SteamDeckUsbip {
+    report: Arc<Mutex<[u8; 64]>>,
+    feedback: Arc<Mutex<SteamFeedback>>,
+    /// The `vhci_hcd` port we attached to — written to the sysfs `detach` file on drop.
+    vhci_port: u16,
+    /// Kept alive so the connected socket fd we handed to `vhci_hcd` stays valid (in-process attach
+    /// only; the CLI hands its own fd to the kernel and exits, so this is `None` there).
+    _client_sock: Option<TcpStream>,
+    /// Emulation-server thread; dropped (stopped) after the detach.
+    _server: ServerThread,
+    seq: u32,
+}
+
+impl SteamDeckUsbip {
+    /// Bind a virtual Deck and attach it locally via `vhci_hcd`. `index` varies only the serial.
+    /// Requires `vhci_hcd` loaded and root (the sysfs attach / the CLI both need it). Tries the
+    /// in-process sysfs attach first, then the `usbip` CLI; `PUNKTFUNK_USBIP_ATTACH=inproc|cli`
+    /// pins one path (for debugging).
+    pub fn open(index: u8) -> Result<SteamDeckUsbip> {
+        ensure_modules();
+        if vhci_base().is_none() {
+            bail!(
+                "vhci_hcd unavailable (no /sys/devices/platform/vhci_hcd*/status) — is it loaded?"
+            );
+        }
+        let mode = std::env::var("PUNKTFUNK_USBIP_ATTACH").ok();
+        if mode.as_deref() != Some("cli") {
+            match Self::open_in_process(index) {
+                Ok(d) => return Ok(d),
+                Err(e) if mode.as_deref() == Some("inproc") => return Err(e),
+                Err(e) => {
+                    tracing::warn!(error = %format!("{e:#}"), "in-process vhci attach failed — trying the usbip CLI")
+                }
+            }
+        }
+        Self::open_via_cli(index)
+    }
+
+    /// In-process attach: emulate on a loopback port, do the import handshake ourselves, hand the
+    /// connected socket to `vhci_hcd` via sysfs. No external dependency.
+    fn open_in_process(index: u8) -> Result<SteamDeckUsbip> {
+        let report = Arc::new(Mutex::new(neutral_deck_report()));
+        let feedback = Arc::new(Mutex::new(SteamFeedback::default()));
+        let dev = build_device(index, &report, &feedback);
+
+        // An ephemeral loopback port (avoids contending the usbip default with another pad).
+        let listener =
+            std::net::TcpListener::bind(("127.0.0.1", 0)).context("bind loopback usbip server")?;
+        let port = listener
+            .local_addr()
+            .context("usbip server local_addr")?
+            .port();
+        listener
+            .set_nonblocking(true)
+            .context("usbip listener set_nonblocking")?;
+        let server = ServerThread::spawn(listener, dev)?;
+
+        // Connect to our own server and run the OP_REQ_IMPORT handshake.
+        let mut sock = connect_loopback(port).context("connect to usbip server")?;
+        let (devid, speed) = import_handshake(&mut sock).context("usbip import handshake")?;
+
+        // Hand the connected socket to vhci_hcd. Clear BOTH timeouts first: the kernel's vhci rx/tx
+        // threads honour SO_RCVTIMEO/SO_SNDTIMEO on this socket, so the 3s handshake timeouts would
+        // otherwise tear the device down after 3s idle (rx) or a 3s-blocked send (tx).
+        let vhci_port = vhci_find_free_port(speed).context("find a free vhci port")?;
+        sock.set_read_timeout(None).ok();
+        sock.set_write_timeout(None).ok();
+        vhci_attach(vhci_port, sock.as_raw_fd(), devid, speed).context("write vhci_hcd attach")?;
+
+        tracing::info!(
+            index,
+            vhci_port,
+            "virtual Steam Deck attached via usbip (in-process — Steam Input recognizes it)"
+        );
+        Ok(SteamDeckUsbip {
+            report,
+            feedback,
+            vhci_port,
+            _client_sock: Some(sock),
+            _server: server,
+            seq: 0,
+        })
+    }
+
+    /// Fallback: emulate on the usbip default port and let the `usbip` CLI attach (it picks the vhci
+    /// port itself; we recover it by diffing the sysfs status).
+    fn open_via_cli(index: u8) -> Result<SteamDeckUsbip> {
+        let report = Arc::new(Mutex::new(neutral_deck_report()));
+        let feedback = Arc::new(Mutex::new(SteamFeedback::default()));
+        let dev = build_device(index, &report, &feedback);
+
+        let listener = std::net::TcpListener::bind(("127.0.0.1", USBIP_TCP_PORT))
+            .with_context(|| format!("bind usbip default port {USBIP_TCP_PORT} for CLI attach"))?;
+        listener
+            .set_nonblocking(true)
+            .context("usbip listener set_nonblocking")?;
+        let server = ServerThread::spawn(listener, dev)?;
+
+        let before = vhci_used_ports();
+        usbip_attach_cli().context("usbip CLI attach")?;
+        let vhci_port = wait_for_new_port(&before)
+            .context("could not determine the vhci port the usbip CLI attached to")?;
+
+        tracing::info!(
+            index,
+            vhci_port,
+            "virtual Steam Deck attached via usbip (CLI — Steam Input recognizes it)"
+        );
+        Ok(SteamDeckUsbip {
+            report,
+            feedback,
+            vhci_port,
+            _client_sock: None,
+            _server: server,
+            seq: 0,
+        })
+    }
+
+    /// Serialize `st` into the 64-byte Deck report streamed on the controller interrupt-IN endpoint.
+    pub fn write_state(&mut self, st: &SteamState) {
+        self.seq = self.seq.wrapping_add(1);
+        let mut r = [0u8; 64];
+        super::steam_proto::serialize_deck_state(&mut r, st, self.seq);
+        if let Ok(mut g) = self.report.lock() {
+            *g = r;
+        }
+    }
+
+    /// Drain any rumble feedback the kernel/Steam wrote to the device.
+    pub fn service(&mut self) -> SteamFeedback {
+        self.feedback
+            .lock()
+            .map(|mut f| std::mem::take(&mut *f))
+            .unwrap_or_default()
+    }
+}
+
+impl Drop for SteamDeckUsbip {
+    fn drop(&mut self) {
+        // Detach the vhci port first (the kernel closes its end of the socket + tears down the
+        // device); `_client_sock` + `_server` then drop, closing our side + stopping the server.
+        if let Err(e) = vhci_detach(self.vhci_port) {
+            tracing::debug!(port = self.vhci_port, error = %e, "vhci detach failed (device may already be gone)");
+        }
+    }
+}
+
+// ---- USB/IP import handshake (we act as the usbip *client* before handing the fd to the kernel) ----
+
+const USBIP_VERSION: u16 = 0x0111;
+const OP_REQ_IMPORT: u16 = 0x8003;
+
+/// Connect to our own loopback server, retrying briefly while the server thread comes up.
+fn connect_loopback(port: u16) -> Result<TcpStream> {
+    let addr = ("127.0.0.1", port);
+    let mut last = None;
+    for _ in 0..50 {
+        match TcpStream::connect(addr) {
+            Ok(s) => {
+                s.set_nodelay(true).ok();
+                return Ok(s);
+            }
+            Err(e) => {
+                last = Some(e);
+                std::thread::sleep(Duration::from_millis(10));
+            }
+        }
+    }
+    Err(anyhow::anyhow!(
+        "connect 127.0.0.1:{port}: {}",
+        last.map(|e| e.to_string()).unwrap_or_default()
+    ))
+}
+
+/// Send `OP_REQ_IMPORT` for [`BUS_ID`] and read `OP_REP_IMPORT`, returning `(devid, speed)` parsed
+/// from the device record (the same `devid = bus_num<<16 | dev_num` + speed `vhci_hcd` wants). The
+/// whole 320-byte reply MUST be consumed here so the socket starts clean at the kernel's first
+/// `USBIP_CMD_SUBMIT`.
+fn import_handshake(sock: &mut TcpStream) -> Result<(u32, u32)> {
+    // Bounded so a non-responsive server can't head-block the per-session input thread (this talks
+    // to our own in-process loopback server, so a working handshake completes in well under a ms).
+    sock.set_read_timeout(Some(Duration::from_secs(1))).ok();
+    sock.set_write_timeout(Some(Duration::from_secs(1))).ok();
+
+    let mut req = Vec::with_capacity(40);
+    req.extend_from_slice(&USBIP_VERSION.to_be_bytes());
+    req.extend_from_slice(&OP_REQ_IMPORT.to_be_bytes());
+    req.extend_from_slice(&0u32.to_be_bytes()); // status
+    let mut busid = [0u8; 32];
+    let b = BUS_ID.as_bytes();
+    busid[..b.len()].copy_from_slice(b);
+    req.extend_from_slice(&busid);
+    sock.write_all(&req).context("send OP_REQ_IMPORT")?;
+
+    // Reply: version(2) code(2) status(4), then the 312-byte device record on success.
+    let mut header = [0u8; 8];
+    sock.read_exact(&mut header)
+        .context("read OP_REP_IMPORT header")?;
+    let status = u32::from_be_bytes([header[4], header[5], header[6], header[7]]);
+    if status != 0 {
+        bail!("OP_REP_IMPORT refused (status={status}) — device {BUS_ID} not exported?");
+    }
+    let mut dev = [0u8; 312];
+    sock.read_exact(&mut dev)
+        .context("read OP_REP_IMPORT device record")?;
+    // Device record layout: path[256], bus_id[32], bus_num(4 BE)@288, dev_num(4 BE)@292, speed(4)@296.
+    let be = |o: usize| u32::from_be_bytes([dev[o], dev[o + 1], dev[o + 2], dev[o + 3]]);
+    let bus_num = be(288);
+    let dev_num = be(292);
+    let speed = be(296);
+    Ok(((bus_num << 16) | dev_num, speed))
+}
+
+// ---- vhci_hcd sysfs plumbing ----
+
+/// Best-effort load of `vhci_hcd` (in-tree + signed on SteamOS/Bazzite/most distros).
+pub fn ensure_modules() {
+    let _ = Command::new("modprobe").arg("vhci_hcd").status();
+}
+
+/// Run `usbip attach -r 127.0.0.1 -b 0-0-0`, bounded by a deadline so a hung CLI can't head-block
+/// the per-session input thread indefinitely (the caller runs this inline on that thread).
+fn usbip_attach_cli() -> Result<()> {
+    let mut child = Command::new("usbip")
+        .args(["attach", "-r", "127.0.0.1", "-b", BUS_ID])
+        .spawn()
+        .context("spawn `usbip attach` (is usbip-utils installed?)")?;
+    let deadline = Instant::now() + Duration::from_secs(6);
+    loop {
+        match child.try_wait().context("wait on `usbip attach`")? {
+            Some(st) if st.success() => return Ok(()),
+            Some(st) => bail!("`usbip attach` exited with {st}"),
+            None if Instant::now() >= deadline => {
+                let _ = child.kill();
+                let _ = child.wait();
+                bail!("`usbip attach` timed out (>6s) — killed");
+            }
+            None => std::thread::sleep(Duration::from_millis(20)),
+        }
+    }
+}
+
+/// Whether a usbip attach should be attempted at all. Default on (the universal Steam-promotable
+/// transport on non-SteamOS hosts); `PUNKTFUNK_STEAM_USBIP=0` forces it off, `=1` forces it on.
+/// [`open`](SteamDeckUsbip::open) still degrades gracefully if `vhci_hcd` turns out to be absent.
+pub fn usbip_preferred() -> bool {
+    !matches!(
+        std::env::var("PUNKTFUNK_STEAM_USBIP").ok().as_deref(),
+        Some("0") | Some("false")
+    )
+}
+
+/// The `vhci_hcd.0` (or legacy `vhci_hcd`) platform sysfs directory, if present.
+fn vhci_base() -> Option<PathBuf> {
+    for p in [
+        "/sys/devices/platform/vhci_hcd.0",
+        "/sys/devices/platform/vhci_hcd",
+    ] {
+        let base = Path::new(p);
+        if base.join("status").exists() {
+            return Some(base.to_path_buf());
+        }
+    }
+    None
+}
+
+fn read_status() -> Result<String> {
+    let base = vhci_base().context("vhci_hcd sysfs not present")?;
+    std::fs::read_to_string(base.join("status")).context("read vhci_hcd status")
+}
+
+/// One parsed `status` row: `(port, hub_is_superspeed, sta)`. Handles both the modern
+/// `hub port sta …` and the legacy `port sta …` column layouts; returns `None` for header/blank rows.
+fn parse_status_row(line: &str) -> Option<(u16, bool, u32)> {
+    let t: Vec<&str> = line.split_whitespace().collect();
+    if t.is_empty() {
+        return None;
+    }
+    let (hub_ss, port_str, sta_str) = if t[0] == "hs" || t[0] == "ss" {
+        (Some(t[0] == "ss"), *t.get(1)?, *t.get(2)?)
+    } else if t[0].chars().all(|c| c.is_ascii_digit()) {
+        (None, t[0], *t.get(1)?) // legacy: port sta …
+    } else {
+        return None; // header ("hub"/"prt"/"port" …)
+    };
+    let port = port_str.parse::<u16>().ok()?;
+    let sta = sta_str.parse::<u32>().ok()?;
+    Some((port, hub_ss.unwrap_or(false), sta))
+}
+
+/// `sta == 4` is `VDEV_ST_NULL` (a free port).
+const VDEV_ST_NULL: u32 = 4;
+
+/// Pick a free `vhci_hcd` port matching the device speed (`usbip_speed >= 5` ⇒ SuperSpeed hub).
+fn vhci_find_free_port(usbip_speed: u32) -> Result<u16> {
+    let want_ss = usbip_speed >= 5;
+    let status = read_status()?;
+    for line in status.lines() {
+        if let Some((port, is_ss, sta)) = parse_status_row(line) {
+            if sta == VDEV_ST_NULL && is_ss == want_ss {
+                return Ok(port);
+            }
+        }
+    }
+    // Speed-class match failed (legacy single-hub status): take any free port.
+    for line in status.lines() {
+        if let Some((port, _, sta)) = parse_status_row(line) {
+            if sta == VDEV_ST_NULL {
+                return Ok(port);
+            }
+        }
+    }
+    bail!("no free vhci_hcd port (all ports in use?)")
+}
+
+/// Ports currently in use (`sta != VDEV_ST_NULL`) — snapshotted around a CLI attach to recover its port.
+fn vhci_used_ports() -> HashSet<u16> {
+    read_status()
+        .unwrap_or_default()
+        .lines()
+        .filter_map(parse_status_row)
+        .filter(|&(_, _, sta)| sta != VDEV_ST_NULL)
+        .map(|(port, _, _)| port)
+        .collect()
+}
+
+/// Poll the status file (briefly) for a port that became used since `before` — the one the CLI attached.
+fn wait_for_new_port(before: &HashSet<u16>) -> Result<u16> {
+    let deadline = Instant::now() + Duration::from_secs(2);
+    loop {
+        if let Some(p) = vhci_used_ports().difference(before).copied().min() {
+            return Ok(p);
+        }
+        if Instant::now() >= deadline {
+            bail!("no newly-attached vhci port appeared after `usbip attach`");
+        }
+        std::thread::sleep(Duration::from_millis(50));
+    }
+}
+
+fn vhci_attach(port: u16, sockfd: i32, devid: u32, speed: u32) -> Result<()> {
+    let base = vhci_base().context("vhci_hcd sysfs not present")?;
+    let line = format!("{port} {sockfd} {devid} {speed}");
+    std::fs::write(base.join("attach"), line)
+        .with_context(|| format!("write vhci_hcd attach (port {port}) — root?"))
+}
+
+fn vhci_detach(port: u16) -> Result<()> {
+    let base = vhci_base().context("vhci_hcd sysfs not present")?;
+    std::fs::write(base.join("detach"), format!("{port}")).context("write vhci_hcd detach")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The `status` parser handles the modern `hub port sta …` layout, the legacy `port sta …`
+    /// layout, and skips header/blank lines — a slip here would mean attaching to a busy port.
+    #[test]
+    fn status_parser_handles_both_layouts() {
+        // modern
+        assert_eq!(
+            parse_status_row("hs  0000 004 000 00000000 000000 0-0"),
+            Some((0, false, 4))
+        );
+        assert_eq!(
+            parse_status_row("ss  0008 006 000 00000000 000000 0-0"),
+            Some((8, true, 6))
+        );
+        // legacy (no hub column)
+        assert_eq!(
+            parse_status_row("0001 004 000 00000000 000000 0-0"),
+            Some((1, false, 4))
+        );
+        // header / blank
+        assert_eq!(
+            parse_status_row("hub port sta spd dev      sockfd local_busid"),
+            None
+        );
+        assert_eq!(parse_status_row(""), None);
+    }
+
+    /// A free HS port is preferred for an HS device; a free SS port for an SS device.
+    #[test]
+    fn free_port_selection_matches_speed() {
+        let status = "hub port sta spd dev      sockfd local_busid\n\
+                      hs  0000 006 000 00000000 000000 0-0\n\
+                      hs  0001 004 000 00000000 000000 0-0\n\
+                      ss  0008 004 000 00000000 000000 0-0\n";
+        // Reuse the parser directly (vhci_find_free_port reads sysfs; test the selection logic).
+        let hs = status
+            .lines()
+            .filter_map(parse_status_row)
+            .find(|&(_, is_ss, sta)| sta == VDEV_ST_NULL && !is_ss)
+            .map(|(p, _, _)| p);
+        let ss = status
+            .lines()
+            .filter_map(parse_status_row)
+            .find(|&(_, is_ss, sta)| sta == VDEV_ST_NULL && is_ss)
+            .map(|(p, _, _)| p);
+        assert_eq!(hs, Some(1));
+        assert_eq!(ss, Some(8));
+    }
+
+    /// On-box smoke test (needs root + `vhci_hcd`): attach a virtual Deck, confirm `hid-steam` binds
+    /// it (the `Steam Deck` evdev appears) and that it tears down on drop. `#[ignore]`d in CI.
+    #[test]
+    #[ignore = "attaches a real vhci_hcd device; needs root + vhci_hcd"]
+    fn usbip_deck_binds_and_tears_down() {
+        ensure_modules();
+        let mut pad = SteamDeckUsbip::open(0).expect("open SteamDeckUsbip (root + vhci_hcd?)");
+        let st = SteamState::from_gamepad(punktfunk_core::input::gamepad::BTN_A, 0, 0, 0, 0, 0, 0);
+        let start = Instant::now();
+        while start.elapsed() < Duration::from_millis(800) {
+            pad.write_state(&st);
+            let _ = pad.service();
+            std::thread::sleep(Duration::from_millis(8));
+        }
+        let devs = std::fs::read_to_string("/proc/bus/input/devices").unwrap_or_default();
+        assert!(
+            devs.contains("Steam Deck"),
+            "hid-steam did not bind the usbip Deck"
+        );
+        drop(pad);
+        std::thread::sleep(Duration::from_millis(300));
+        let devs = std::fs::read_to_string("/proc/bus/input/devices").unwrap_or_default();
+        assert!(
+            !devs.contains("Steam Deck Motion Sensors"),
+            "device not torn down on drop"
+        );
+    }
+}

crates/punktfunk-host/src/inject/proto/steam_proto.rs

@@ -0,0 +1,684 @@
+//! Transport-independent Steam Controller / Steam Deck HID contract — the Steam analogue of
+//! [`super::dualsense_proto`]. The report descriptor, the command/feature IDs, the byte-exact
+//! Deck input-report serializer, the `XInput`/rich-input → state mappers, and the rumble-feedback
+//! parser. Pure logic, shared by the Linux UHID backend and (later) a Windows UMDF backend.
+//!
+//! **Layout source of truth:** the kernel `drivers/hid/hid-steam.c` `steam_do_deck_input_event`
+//! (+ `steam_do_deck_sensors_event`) — every offset/bit/sign below is transcribed verbatim from
+//! it and on-box-validated against kernel 7.0 (see `design/steam-controller-deck-support.md`).
+//! M0 proved the device binds + parses; M1 (here) makes the serializer byte-exact.
+//!
+//! Three load-bearing details the DualSense path does NOT have:
+//!   * **report id 0 / unnumbered**: input reports are the raw 64 bytes starting `[0x01,0x00,0x09]`
+//!     (no report-id prefix); FEATURE get/set reports DO carry a leading `0x00` report-id byte
+//!     (`steam_send_report` does `memcpy(buf+1, cmd, …)`, `steam_recv_report` strips `buf[0]`).
+//!   * **`gamepad_mode` gate**: `steam_do_deck_input_event` early-returns when
+//!     `!gamepad_mode && lizard_mode` (the module param, default on). `gamepad_mode` starts false
+//!     and TOGGLES when [`btn::STEAM_MENU_RIGHT`] (`b9.6`, the mode-switch) is held ~450 ms while
+//!     no hidraw client is open. The backend enters gamepad mode at session start (pulse that bit,
+//!     or load `hid_steam lizard_mode=0`) — see the backend, not this module.
+//!   * **the `UHID_SET_REPORT` handshake** must be answered (DualSense omits it).
+#![allow(dead_code)] // Some of the full model is consumed only once the M2 backend + M3 wire land.
+
+use punktfunk_core::input::gamepad as gs;
+use punktfunk_core::quic::RichInput;
+
+/// Valve. `hid-steam` matches purely by VID/PID over `BUS_USB`.
+pub const STEAM_VENDOR: u32 = 0x28DE;
+/// Steam Deck built-in controller (same PID on LCD + OLED).
+pub const STEAMDECK_PRODUCT: u32 = 0x1205;
+/// Classic Steam Controller, wired (report id 1 / `ID_CONTROLLER_STATE`; a later model).
+pub const STEAMCTRL_WIRED_PRODUCT: u32 = 0x1102;
+
+/// The Steam HID state/command report is a fixed 64-byte, **unnumbered** (report-id-0) frame.
+pub const STEAM_REPORT_LEN: usize = 64;
+
+// Command IDs (drivers/hid/hid-steam.c), confirmed against the kernel source.
+pub const ID_CLEAR_DIGITAL_MAPPINGS: u8 = 0x81;
+pub const ID_GET_ATTRIBUTES_VALUES: u8 = 0x83;
+pub const ID_SET_SETTINGS_VALUES: u8 = 0x87;
+pub const ID_LOAD_DEFAULT_SETTINGS: u8 = 0x8E;
+pub const ID_GET_DEVICE_INFO: u8 = 0xA1;
+pub const ID_GET_STRING_ATTRIBUTE: u8 = 0xAE;
+pub const ATTRIB_STR_UNIT_SERIAL: u8 = 0x01;
+/// Host→client feedback: `steam_haptic_rumble` emits report `[0xEB, 9, …]` (FF_RUMBLE → trackpad
+/// actuators / Deck motors). The Deck's rumble path; the classic SC also has `0x8F` pad pulses.
+pub const ID_TRIGGER_RUMBLE_CMD: u8 = 0xEB;
+pub const ID_TRIGGER_HAPTIC_PULSE: u8 = 0x8F;
+/// Input report message types: SC = `ID_CONTROLLER_STATE`, Deck = `ID_CONTROLLER_DECK_STATE`.
+pub const ID_CONTROLLER_STATE: u8 = 0x01;
+pub const ID_CONTROLLER_DECK_STATE: u8 = 0x09;
+
+/// Which Steam device identity to present. M1 implements the Deck fully; the classic Controller
+/// (dual trackpads, report id 1, trackpad-only haptics) is a later identity behind the same path.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum SteamModel {
+    Deck,
+    Controller,
+}
+
+impl SteamModel {
+    pub fn product(self) -> u32 {
+        match self {
+            SteamModel::Deck => STEAMDECK_PRODUCT,
+            SteamModel::Controller => STEAMCTRL_WIRED_PRODUCT,
+        }
+    }
+}
+
+/// Minimal vendor-defined HID report descriptor: one application collection with a 64-byte input
+/// report and a 64-byte feature report, both UNNUMBERED (report id 0). `hid-steam` is a raw-event
+/// driver, so the field layout is cosmetic — but `steam_probe` requires `hid_parse` to succeed AND
+/// a non-empty FEATURE report list (`steam_is_valve_interface`), so the feature item is mandatory.
+#[rustfmt::skip]
+pub const STEAMDECK_RDESC: &[u8] = &[
+    0x06, 0x00, 0xFF, // Usage Page (Vendor-Defined 0xFF00)
+    0x09, 0x01,       // Usage (0x01)
+    0xA1, 0x01,       // Collection (Application)
+    0x15, 0x00,       //   Logical Minimum (0)
+    0x26, 0xFF, 0x00, //   Logical Maximum (255)
+    0x75, 0x08,       //   Report Size (8 bits)
+    0x95, 0x40,       //   Report Count (64)
+    0x09, 0x01,       //   Usage (0x01)
+    0x81, 0x02,       //   Input (Data,Var,Abs)    — the 64-byte state report
+    0x09, 0x01,       //   Usage (0x01)
+    0x95, 0x40,       //   Report Count (64)
+    0xB1, 0x02,       //   Feature (Data,Var,Abs)  — makes steam_is_valve_interface() true
+    0xC0,             // End Collection
+];
+
+/// Deck button bits, indexed in the `u64` packed across report bytes 8..16 — bit `(byte-8)*8 + bit`,
+/// transcribed verbatim from `steam_do_deck_input_event` (bytes 12 + 15 carry no buttons). Naming
+/// follows the physical Deck control; the trailing comment is the kernel `BTN_*` it maps to.
+pub mod btn {
+    // byte 8
+    pub const RT_FULL: u64 = 1 << 0; // BTN_TR2  — right trigger fully pressed
+    pub const LT_FULL: u64 = 1 << 1; // BTN_TL2  — left trigger fully pressed
+    pub const RB: u64 = 1 << 2; //      BTN_TR   — right shoulder
+    pub const LB: u64 = 1 << 3; //      BTN_TL   — left shoulder
+    pub const Y: u64 = 1 << 4;
+    pub const B: u64 = 1 << 5;
+    pub const X: u64 = 1 << 6;
+    pub const A: u64 = 1 << 7;
+    // byte 9
+    pub const DPAD_UP: u64 = 1 << 8;
+    pub const DPAD_RIGHT: u64 = 1 << 9;
+    pub const DPAD_LEFT: u64 = 1 << 10;
+    pub const DPAD_DOWN: u64 = 1 << 11;
+    pub const VIEW: u64 = 1 << 12; //   BTN_SELECT — "menu left" (View / Back)
+    pub const STEAM: u64 = 1 << 13; //  BTN_MODE   — Steam logo button
+    pub const MENU: u64 = 1 << 14; //   BTN_START  — "menu right" (Start / Options)
+    pub const L5: u64 = 1 << 15; //     BTN_GRIPL2 — left BOTTOM back grip
+                                 // byte 10
+    pub const R5: u64 = 1 << 16; //     BTN_GRIPR2 — right BOTTOM back grip
+    pub const LPAD_CLICK: u64 = 1 << 17; // BTN_THUMB  — left pad pressed (click)
+    pub const RPAD_CLICK: u64 = 1 << 18; // BTN_THUMB2 — right pad pressed (click)
+    pub const LPAD_TOUCH: u64 = 1 << 19; // gates ABS_HAT0 (left pad coords)
+    pub const RPAD_TOUCH: u64 = 1 << 20; // gates ABS_HAT1 (right pad coords)
+    pub const L3: u64 = 1 << 22; //     BTN_THUMBL — left joystick click
+                                 // byte 11
+    pub const R3: u64 = 1 << 26; //     BTN_THUMBR — right joystick click
+                                 // byte 13
+    pub const L4: u64 = 1 << 41; //     BTN_GRIPL  — left TOP back grip
+    pub const R4: u64 = 1 << 42; //     BTN_GRIPR  — right TOP back grip
+    pub const LJOY_TOUCH: u64 = 1 << 46;
+    pub const RJOY_TOUCH: u64 = 1 << 47;
+    // byte 14
+    pub const QAM: u64 = 1 << 50; //    BTN_BASE   — quick-access (…) button
+    /// `b9.6` doubles as the mode-switch: held ~450 ms (no hidraw client) it toggles `gamepad_mode`.
+    pub const STEAM_MENU_RIGHT: u64 = MENU;
+}
+
+/// Full virtual Steam Deck controller state. All analog fields are stored as the RAW little-endian
+/// report values the kernel reads (so [`serialize_deck_state`] is a pure memcpy); the kernel applies
+/// its own sign conventions on top (`ABS_Y = -raw`, etc.) — see [`SteamState::from_gamepad`].
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub struct SteamState {
+    /// Packed button bits (see [`btn`]); occupies report bytes 8..16.
+    pub buttons: u64,
+    /// Left / right joystick, raw s16 (report 48/50/52/54). The kernel negates the Y axes.
+    pub lx: i16,
+    pub ly: i16,
+    pub rx: i16,
+    pub ry: i16,
+    /// Left / right analog trigger, raw u16 (report 44/46 → ABS_HAT2Y/X).
+    pub lt: u16,
+    pub rt: u16,
+    /// Left / right trackpad position, raw s16, centred 0 (report 16/18/20/22). Only surfaced by
+    /// the kernel while the matching `*PAD_TOUCH` button bit is set.
+    pub lpad_x: i16,
+    pub lpad_y: i16,
+    pub rpad_x: i16,
+    pub rpad_y: i16,
+    pub lpad_pressure: u16,
+    pub rpad_pressure: u16,
+    /// IMU, raw s16. `accel`/`gyro` are `[X, Y, Z]`; the kernel maps them to ABS_X/Z/Y + ABS_RX/RZ/RY
+    /// (with Z/RZ negated) on the separate sensors evdev.
+    pub accel: [i16; 3],
+    pub gyro: [i16; 3],
+}
+
+impl SteamState {
+    pub fn neutral() -> SteamState {
+        SteamState::default()
+    }
+
+    /// Set/clear a button (or group) by its [`btn`] mask.
+    pub fn press(&mut self, mask: u64, down: bool) {
+        if down {
+            self.buttons |= mask;
+        } else {
+            self.buttons &= !mask;
+        }
+    }
+
+    /// Map an `XInput`/GameStream pad frame (button bitmask + i16 sticks + u8 triggers) into the Deck
+    /// state. Sticks pass through (the kernel negates Y, which yields the conventional direction —
+    /// validated on-box); triggers scale u8 0..255 → u16 0..32640 and set the full-pull bit when
+    /// pressed. Trackpad + motion + the back grips arrive separately ([`apply_rich`], the M3 wire).
+    pub fn from_gamepad(
+        buttons: u32,
+        lx: i16,
+        ly: i16,
+        rx: i16,
+        ry: i16,
+        lt: u8,
+        rt: u8,
+    ) -> SteamState {
+        let on = |bit: u32| buttons & bit != 0;
+        let mut s = SteamState {
+            lx,
+            ly,
+            rx,
+            ry,
+            lt: (lt as u16) * 128,
+            rt: (rt as u16) * 128,
+            ..SteamState::neutral()
+        };
+        let mut b = 0u64;
+        let set = |b: &mut u64, on: bool, m: u64| {
+            if on {
+                *b |= m;
+            }
+        };
+        set(&mut b, on(gs::BTN_A), btn::A);
+        set(&mut b, on(gs::BTN_B), btn::B);
+        set(&mut b, on(gs::BTN_X), btn::X);
+        set(&mut b, on(gs::BTN_Y), btn::Y);
+        set(&mut b, on(gs::BTN_LB), btn::LB);
+        set(&mut b, on(gs::BTN_RB), btn::RB);
+        set(&mut b, lt > 0, btn::LT_FULL);
+        set(&mut b, rt > 0, btn::RT_FULL);
+        set(&mut b, on(gs::BTN_BACK), btn::VIEW);
+        set(&mut b, on(gs::BTN_START), btn::MENU);
+        set(&mut b, on(gs::BTN_GUIDE), btn::STEAM);
+        set(&mut b, on(gs::BTN_LS_CLICK), btn::L3);
+        set(&mut b, on(gs::BTN_RS_CLICK), btn::R3);
+        set(&mut b, on(gs::BTN_DPAD_UP), btn::DPAD_UP);
+        set(&mut b, on(gs::BTN_DPAD_DOWN), btn::DPAD_DOWN);
+        set(&mut b, on(gs::BTN_DPAD_LEFT), btn::DPAD_LEFT);
+        set(&mut b, on(gs::BTN_DPAD_RIGHT), btn::DPAD_RIGHT);
+        // The DualSense touchpad-click wire bit maps to the Deck's RIGHT pad click (the pad that
+        // stands in for the DualSense touchpad — see apply_rich).
+        set(&mut b, on(gs::BTN_TOUCHPAD), btn::RPAD_CLICK);
+        // Back grips (the whole reason for the Deck identity): the wire paddle bits map to the four
+        // Deck grips — PADDLE1/2/3/4 = R4/L4/R5/L5 (see `input::gamepad`); MISC1 = the QAM '…' button.
+        set(&mut b, on(gs::BTN_PADDLE1), btn::R4);
+        set(&mut b, on(gs::BTN_PADDLE2), btn::L4);
+        set(&mut b, on(gs::BTN_PADDLE3), btn::R5);
+        set(&mut b, on(gs::BTN_PADDLE4), btn::L5);
+        set(&mut b, on(gs::BTN_MISC1), btn::QAM);
+        s.buttons = b;
+        s
+    }
+
+    /// Apply one rich client→host event into this state, preserving everything else. The single-pad
+    /// wire [`RichInput::Touchpad`] maps to the **right** trackpad (the Deck pad analogous to the
+    /// DualSense touchpad); the left pad arrives via the M3 `TouchpadEx` surface. [`RichInput::Motion`]
+    /// passes gyro/accel straight through (raw i16; cross-device unit scaling is M3).
+    pub fn apply_rich(&mut self, rich: RichInput) {
+        match rich {
+            RichInput::Touchpad { active, x, y, .. } => {
+                self.press(btn::RPAD_TOUCH, active);
+                // Normalized 0..=65535 (centre 32768) → the pad's centred s16 range.
+                self.rpad_x = ((x as i32) - 32768) as i16;
+                self.rpad_y = ((y as i32) - 32768) as i16;
+            }
+            RichInput::Motion { gyro, accel, .. } => {
+                // The wire carries DualSense-convention units (what every client capture emits); the
+                // Deck's hid-steam report wants 16 LSB/°·s + 16384 LSB/g, so rescale here.
+                let (g, a) = super::steam_remap::motion_wire_to_deck(gyro, accel);
+                self.gyro = g;
+                self.accel = a;
+            }
+            RichInput::TouchpadEx {
+                surface,
+                touch,
+                click,
+                x,
+                y,
+                ..
+            } => {
+                // Steam pads are natively signed (centre 0), so x/y map straight in. surface 1 =
+                // left pad, anything else (0 single / 2 right) = right pad.
+                if surface == 1 {
+                    self.press(btn::LPAD_TOUCH, touch);
+                    self.press(btn::LPAD_CLICK, click);
+                    self.lpad_x = x;
+                    self.lpad_y = y;
+                } else {
+                    self.press(btn::RPAD_TOUCH, touch);
+                    self.press(btn::RPAD_CLICK, click);
+                    self.rpad_x = x;
+                    self.rpad_y = y;
+                }
+            }
+        }
+    }
+}
+
+/// Serialize the full Deck input report (`ID_CONTROLLER_DECK_STATE`) into the 64-byte unnumbered
+/// frame `hid-steam` parses. Pure + byte-exact against `steam_do_deck_input_event`; the report-id
+/// constant is `data[0]=0x01` (NOT a HID report id — this report is unnumbered).
+pub fn serialize_deck_state(r: &mut [u8; STEAM_REPORT_LEN], st: &SteamState, seq: u32) {
+    r.fill(0);
+    r[0] = 0x01;
+    r[1] = 0x00;
+    r[2] = ID_CONTROLLER_DECK_STATE;
+    r[3] = 0x3C; // payload length; the kernel ignores it
+    r[4..8].copy_from_slice(&seq.to_le_bytes());
+    r[8..16].copy_from_slice(&st.buttons.to_le_bytes()); // bytes 8..16 (12+15 stay 0)
+    r[16..18].copy_from_slice(&st.lpad_x.to_le_bytes());
+    r[18..20].copy_from_slice(&st.lpad_y.to_le_bytes());
+    r[20..22].copy_from_slice(&st.rpad_x.to_le_bytes());
+    r[22..24].copy_from_slice(&st.rpad_y.to_le_bytes());
+    r[24..26].copy_from_slice(&st.accel[0].to_le_bytes()); // accel X → IMU ABS_X
+    r[26..28].copy_from_slice(&st.accel[1].to_le_bytes()); // accel Y → IMU ABS_Z (kernel negates)
+    r[28..30].copy_from_slice(&st.accel[2].to_le_bytes()); // accel Z → IMU ABS_Y
+    r[30..32].copy_from_slice(&st.gyro[0].to_le_bytes()); //  gyro X  → IMU ABS_RX
+    r[32..34].copy_from_slice(&st.gyro[1].to_le_bytes()); //  gyro Y  → IMU ABS_RZ (kernel negates)
+    r[34..36].copy_from_slice(&st.gyro[2].to_le_bytes()); //  gyro Z  → IMU ABS_RY
+                                                          // 36..44 quaternion — left 0 (optional; the kernel does not surface it)
+    r[44..46].copy_from_slice(&st.lt.to_le_bytes()); // left trigger  → ABS_HAT2Y
+    r[46..48].copy_from_slice(&st.rt.to_le_bytes()); // right trigger → ABS_HAT2X
+    r[48..50].copy_from_slice(&st.lx.to_le_bytes()); // left joystick X  → ABS_X
+    r[50..52].copy_from_slice(&st.ly.to_le_bytes()); // left joystick Y  → ABS_Y (kernel negates)
+    r[52..54].copy_from_slice(&st.rx.to_le_bytes()); // right joystick X → ABS_RX
+    r[54..56].copy_from_slice(&st.ry.to_le_bytes()); // right joystick Y → ABS_RY (kernel negates)
+    r[56..58].copy_from_slice(&st.lpad_pressure.to_le_bytes());
+    r[58..60].copy_from_slice(&st.rpad_pressure.to_le_bytes());
+}
+
+/// Build the `steam_get_serial` GET_REPORT reply. The Steam feature path is report-id-0 with a
+/// leading report-id byte the kernel strips (`steam_recv_report` does `memcpy(data, buf+1, …)`), so
+/// the wire is `[0x00, 0xAE, len, 0x01, ascii…]`; the kernel then validates `reply[0]==0xAE`,
+/// `1<=reply[1]<=21`, `reply[2]==0x01`. Non-fatal (a bad reply → the `"XXXXXXXXXX"` fallback).
+pub fn serial_reply(serial: &str) -> [u8; STEAM_REPORT_LEN] {
+    let mut buf = [0u8; STEAM_REPORT_LEN];
+    let bytes = serial.as_bytes();
+    let len = bytes.len().clamp(1, 21);
+    buf[0] = 0x00; // report id 0 — stripped by steam_recv_report
+    buf[1] = ID_GET_STRING_ATTRIBUTE;
+    buf[2] = len as u8;
+    buf[3] = ATTRIB_STR_UNIT_SERIAL;
+    buf[4..4 + len].copy_from_slice(&bytes[..len]);
+    buf
+}
+
+/// One service pass's extracted feedback. Rumble rides the universal 0xCA plane (so any client
+/// feels it); the classic SC's trackpad-pulse haptics (`0x8F`) are a later, model-specific add.
+#[derive(Default, Debug, PartialEq, Eq)]
+pub struct SteamFeedback {
+    /// `(low, high)` motor levels (left/strong, right/weak), if a rumble report carried them.
+    pub rumble: Option<(u16, u16)>,
+}
+
+/// Parse a feature/output report the kernel wrote to our device. The Steam feedback path is a
+/// FEATURE `SET_REPORT` whose wire data is `[0x00 report-id, cmd, len, …]`; `cmd == 0xEB`
+/// (`steam_haptic_rumble`) carries `[…, 0, intensity(2), left_speed(2), right_speed(2), gains(2)]`.
+/// We surface `(left_speed, right_speed)` as `(low, high)` for the 0xCA rumble plane.
+pub fn parse_steam_output(data: &[u8]) -> SteamFeedback {
+    let mut fb = SteamFeedback::default();
+    // data[0] is the stripped report-id byte (0); the command id follows.
+    if data.len() >= 10 && data[1] == ID_TRIGGER_RUMBLE_CMD {
+        let le = |o: usize| u16::from_le_bytes([data[o], data[o + 1]]);
+        let left = le(6); // left_speed  (report[5..7])  → low / strong motor
+        let right = le(8); // right_speed (report[7..9]) → high / weak motor
+        fb.rumble = Some((left, right));
+    }
+    fb
+}
+
+// ===========================================================================================
+// Real-USB Deck device contract (the gadget + usbip transports present a *real* 3-interface USB
+// Deck so Steam Input promotes it; the UHID path above uses the minimal [`STEAMDECK_RDESC`]).
+//
+// These descriptors are captured verbatim from a physical Steam Deck (28DE:1205): mouse =
+// interface 0, keyboard = interface 1, **controller = interface 2** (the interface number Steam's
+// own driver filters on — the reason a UHID Deck, `Interface: -1`, is never promoted). The
+// `0x83`/`0xAE` feature contract is what stops Steam re-probing (the gamepad-evdev churn). Shared
+// by [`super::super::steam_gadget`] (raw_gadget) and [`super::super::steam_usbip`] (usbip/vhci).
+// ===========================================================================================
+
+/// Captured Deck **mouse** report descriptor (interface 0, EP 0x81).
+#[rustfmt::skip]
+pub const RDESC_DECK_MOUSE: &[u8] = &[
+    0x05,0x01,0x09,0x02,0xa1,0x01,0x09,0x01,0xa1,0x00,0x05,0x09,0x19,0x01,0x29,0x02,
+    0x15,0x00,0x25,0x01,0x75,0x01,0x95,0x02,0x81,0x02,0x75,0x06,0x95,0x01,0x81,0x01,
+    0x05,0x01,0x09,0x30,0x09,0x31,0x15,0x81,0x25,0x7f,0x75,0x08,0x95,0x02,0x81,0x06,
+    0x95,0x01,0x09,0x38,0x81,0x06,0x05,0x0c,0x0a,0x38,0x02,0x95,0x01,0x81,0x06,0xc0,0xc0];
+/// Captured Deck **keyboard** (boot) report descriptor (interface 1, EP 0x82).
+#[rustfmt::skip]
+pub const RDESC_DECK_KBD: &[u8] = &[
+    0x05,0x01,0x09,0x06,0xa1,0x01,0x05,0x07,0x19,0xe0,0x29,0xe7,0x15,0x00,0x25,0x01,
+    0x75,0x01,0x95,0x08,0x81,0x02,0x81,0x01,0x19,0x00,0x29,0x65,0x15,0x00,0x25,0x65,
+    0x75,0x08,0x95,0x06,0x81,0x00,0xc0];
+/// Captured Deck **controller** report descriptor (interface 2, EP 0x83; Usage Page `0xFFFF`,
+/// `bCountryCode 33`). The vendor-defined report the `hid-steam` driver binds.
+#[rustfmt::skip]
+pub const RDESC_DECK_CTRL: &[u8] = &[
+    0x06,0xff,0xff,0x09,0x01,0xa1,0x01,0x09,0x02,0x09,0x03,0x15,0x00,0x26,0xff,0x00,
+    0x75,0x08,0x95,0x40,0x81,0x02,0x09,0x06,0x09,0x07,0x15,0x00,0x26,0xff,0x00,0x75,
+    0x08,0x95,0x40,0xb1,0x02,0xc0];
+
+/// Per-instance Deck unit id stamped into the `0x83` GET_ATTRIBUTES device-id attrs (`0x0a`/`0x04`)
+/// so a virtual Deck never collides with a real one or another instance. `"PF"` high word + index.
+pub fn deck_unit_id(index: u8) -> u32 {
+    0x5046_0000 | index as u32
+}
+
+/// A Steam-accepted alphanumeric unit serial (a real Deck's is e.g. `"FVZZ4200469B"`; Steam rejects
+/// a too-short/oddly-formatted one as "Invalid or missing unit serial number" and substitutes its
+/// own — benign, but we present a clean 12-char one). Derived from [`deck_unit_id`] so the `0xAE`
+/// serial reply and the `0x83` unit-id attrs stay consistent.
+pub fn deck_serial(index: u8) -> String {
+    format!("PFDK{:08X}", deck_unit_id(index))
+}
+
+/// The neutral 64-byte Deck input report (header only, all controls released) — the report the
+/// real-USB transports stream until the first [`serialize_deck_state`] call updates it.
+pub fn neutral_deck_report() -> [u8; STEAM_REPORT_LEN] {
+    let mut r = [0u8; STEAM_REPORT_LEN];
+    r[0] = 0x01;
+    r[2] = ID_CONTROLLER_DECK_STATE;
+    r[3] = 0x3C;
+    r
+}
+
+/// Build the HID feature GET_REPORT reply for the host's last SET_REPORT command, for the *real-USB*
+/// Deck (gadget + usbip). Steam's `GetControllerInfo` reads the `0x83` attributes + the `0xAE`
+/// serial; **serving the real `0x83` blob is what stops Steam re-probing** (the gamepad-evdev churn).
+/// The 9-attribute `0x83` layout + the `0xAE` string format were captured from a physical Deck via
+/// hidraw. `unit_id` (see [`deck_unit_id`]) stamps a per-instance value into the device-id attrs.
+///
+/// Note this is the raw 64-byte EP0 feature payload (command id first, no report-id prefix) — the USB
+/// control path, distinct from [`serial_reply`] which carries the UHID report-id byte the kernel
+/// strips.
+pub fn feature_reply(last_set: &[u8], serial: &str, unit_id: u32) -> [u8; STEAM_REPORT_LEN] {
+    let cmd = last_set.first().copied().unwrap_or(ID_GET_STRING_ATTRIBUTE);
+    let mut r = [0u8; STEAM_REPORT_LEN];
+    match cmd {
+        ID_GET_ATTRIBUTES_VALUES => {
+            // GET_ATTRIBUTES_VALUES: [0x83, 0x2d, then 9× (attr-id, value u32-LE)].
+            r[0] = ID_GET_ATTRIBUTES_VALUES;
+            r[1] = 0x2d;
+            let attrs: [(u8, u32); 9] = [
+                (0x01, 0x1205), // product id
+                (0x02, 0),
+                (0x0a, unit_id), // unit serial number (per-instance)
+                (0x04, unit_id ^ 0x5555_5555),
+                (0x09, 0x2e),
+                (0x0b, 0x0fa0),
+                (0x0d, 0),
+                (0x0c, 0),
+                (0x0e, 0),
+            ];
+            let mut o = 2;
+            for (id, val) in attrs {
+                r[o] = id;
+                r[o + 1..o + 5].copy_from_slice(&val.to_le_bytes());
+                o += 5;
+            }
+        }
+        ID_GET_STRING_ATTRIBUTE => {
+            // GET_STRING_ATTRIBUTE: [0xAE, len, attr, ascii…]. The kernel validates the serial (attr
+            // 0x01) wants reply[2]==0x01 and 1<=len<=21; for other attrs we echo the requested id.
+            let attr = last_set.get(2).copied().unwrap_or(ATTRIB_STR_UNIT_SERIAL);
+            let b = serial.as_bytes();
+            let len = b.len().clamp(1, 20);
+            r[0] = ID_GET_STRING_ATTRIBUTE;
+            r[1] = len as u8;
+            r[2] = attr;
+            r[3..3 + len].copy_from_slice(&b[..len]);
+        }
+        _ => {
+            // Settings read-back (e.g. 0x87): echo the host's last command + data.
+            let n = last_set.len().min(STEAM_REPORT_LEN);
+            r[..n].copy_from_slice(&last_set[..n]);
+        }
+    }
+    r
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn descriptor_declares_input_and_feature_reports() {
+        assert!(
+            STEAMDECK_RDESC.contains(&0xB1),
+            "missing Feature main item — steam_is_valve_interface() would fail"
+        );
+        assert!(STEAMDECK_RDESC.contains(&0x81), "missing Input main item");
+        assert_eq!(
+            *STEAMDECK_RDESC.last().unwrap(),
+            0xC0,
+            "unterminated collection"
+        );
+    }
+
+    /// Every analog field lands at the exact offset `steam_do_deck_input_event` reads, the header is
+    /// what `steam_raw_event` requires, and the buttons pack into bytes 8..16 (12+15 zero). A
+    /// one-byte slip here turns the whole controller into noise.
+    #[test]
+    fn serialize_is_byte_exact() {
+        let mut st = SteamState::neutral();
+        st.buttons = btn::A | btn::L4 | btn::R5 | btn::QAM;
+        st.lx = 0x1122;
+        st.ly = 0x3344;
+        st.rx = 0x5566;
+        st.ry = 0x778;
+        st.lt = 0xABCD;
+        st.rt = 0xEF01;
+        st.lpad_x = 0x0A0B;
+        st.lpad_y = 0x0C0D;
+        st.rpad_x = 0x0E0F;
+        st.rpad_y = 0x1011;
+        st.accel = [0x0102, 0x0304, 0x0506];
+        st.gyro = [0x0708, 0x090A, 0x0B0C];
+        st.lpad_pressure = 0x1314;
+        st.rpad_pressure = 0x1516;
+        let mut r = [0u8; STEAM_REPORT_LEN];
+        serialize_deck_state(&mut r, &st, 0xAABB_CCDD);
+        assert_eq!(&r[0..4], &[0x01, 0x00, 0x09, 0x3C]);
+        assert_eq!(&r[4..8], &[0xDD, 0xCC, 0xBB, 0xAA]); // seq LE
+                                                         // buttons: A=bit7 (byte8), L4=bit41 (byte13.1), R5=bit16 (byte10.0), QAM=bit50 (byte14.2).
+        assert_eq!(r[8], 0x80); // A
+        assert_eq!(r[10], 0x01); // R5
+        assert_eq!(r[12], 0x00); // unused button byte
+        assert_eq!(r[13], 0x02); // L4 (bit 1)
+        assert_eq!(r[14], 0x04); // QAM (bit 2)
+        assert_eq!(r[15], 0x00); // unused button byte
+        assert_eq!(&r[16..18], &0x0A0Bi16.to_le_bytes()); // lpad X
+        assert_eq!(&r[20..22], &0x0E0Fi16.to_le_bytes()); // rpad X
+        assert_eq!(&r[24..26], &0x0102i16.to_le_bytes()); // accel X
+        assert_eq!(&r[26..28], &0x0304i16.to_le_bytes()); // accel Y
+        assert_eq!(&r[28..30], &0x0506i16.to_le_bytes()); // accel Z
+        assert_eq!(&r[30..32], &0x0708i16.to_le_bytes()); // gyro X
+        assert_eq!(&r[44..46], &0xABCDu16.to_le_bytes()); // left trigger
+        assert_eq!(&r[46..48], &0xEF01u16.to_le_bytes()); // right trigger
+        assert_eq!(&r[48..50], &0x1122i16.to_le_bytes()); // left joy X
+        assert_eq!(&r[50..52], &0x3344i16.to_le_bytes()); // left joy Y
+        assert_eq!(&r[52..54], &0x5566i16.to_le_bytes()); // right joy X
+        assert_eq!(&r[56..58], &0x1314u16.to_le_bytes()); // left pad pressure
+        assert_eq!(&r[58..60], &0x1516u16.to_le_bytes()); // right pad pressure
+    }
+
+    /// `from_gamepad` sets the right Deck bits + scales triggers, and a touched flag is merged when
+    /// a trackpad contact arrives via `apply_rich`.
+    #[test]
+    fn from_gamepad_and_rich_mapping() {
+        let s = SteamState::from_gamepad(
+            gs::BTN_A | gs::BTN_START | gs::BTN_GUIDE | gs::BTN_LB,
+            1000,
+            -2000,
+            0,
+            0,
+            255,
+            0,
+        );
+        assert_ne!(s.buttons & btn::A, 0);
+        assert_ne!(s.buttons & btn::MENU, 0);
+        assert_ne!(s.buttons & btn::STEAM, 0);
+        assert_ne!(s.buttons & btn::LB, 0);
+        assert_ne!(s.buttons & btn::LT_FULL, 0); // lt=255 → full-pull bit
+        assert_eq!(s.lt, 255 * 128);
+        assert_eq!(s.lx, 1000);
+        assert_eq!(s.ly, -2000);
+
+        let mut s = SteamState::neutral();
+        s.apply_rich(RichInput::Touchpad {
+            pad: 0,
+            finger: 0,
+            active: true,
+            x: 65535,
+            y: 0,
+        });
+        assert_ne!(s.buttons & btn::RPAD_TOUCH, 0);
+        assert_eq!(s.rpad_x, 32767); // 65535-32768
+        assert_eq!(s.rpad_y, -32768); // 0-32768
+                                      // Motion is rescaled from the wire (DualSense) convention into Deck units (gyro ×16/20,
+                                      // accel ×16384/10000) — see steam_remap::motion_wire_to_deck.
+        s.apply_rich(RichInput::Motion {
+            pad: 0,
+            gyro: [1000, -2000, 0],
+            accel: [10000, -5000, 0],
+        });
+        assert_eq!(s.gyro, [800, -1600, 0]);
+        assert_eq!(s.accel, [16384, -8192, 0]);
+    }
+
+    /// M3: the wire back-button bits map to the four Deck grips + QAM, and `TouchpadEx` routes the
+    /// left / right surfaces to the matching pad (signed coords pass straight through).
+    #[test]
+    fn back_buttons_and_dual_trackpad_mapping() {
+        let s = SteamState::from_gamepad(
+            gs::BTN_PADDLE1 | gs::BTN_PADDLE2 | gs::BTN_PADDLE3 | gs::BTN_PADDLE4 | gs::BTN_MISC1,
+            0,
+            0,
+            0,
+            0,
+            0,
+            0,
+        );
+        assert_ne!(s.buttons & btn::R4, 0); // PADDLE1 = R4
+        assert_ne!(s.buttons & btn::L4, 0); // PADDLE2 = L4
+        assert_ne!(s.buttons & btn::R5, 0); // PADDLE3 = R5
+        assert_ne!(s.buttons & btn::L5, 0); // PADDLE4 = L5
+        assert_ne!(s.buttons & btn::QAM, 0); // MISC1 = QAM
+
+        let mut s = SteamState::neutral();
+        s.apply_rich(RichInput::TouchpadEx {
+            pad: 0,
+            surface: 1,
+            finger: 0,
+            touch: true,
+            click: true,
+            x: -5000,
+            y: 6000,
+            pressure: 100,
+        });
+        assert_ne!(s.buttons & btn::LPAD_TOUCH, 0);
+        assert_ne!(s.buttons & btn::LPAD_CLICK, 0);
+        assert_eq!((s.lpad_x, s.lpad_y), (-5000, 6000));
+        s.apply_rich(RichInput::TouchpadEx {
+            pad: 0,
+            surface: 2,
+            finger: 0,
+            touch: true,
+            click: false,
+            x: 7000,
+            y: -8000,
+            pressure: 0,
+        });
+        assert_ne!(s.buttons & btn::RPAD_TOUCH, 0);
+        assert_eq!((s.rpad_x, s.rpad_y), (7000, -8000));
+    }
+
+    /// The serial reply carries the leading report-id byte the kernel strips, so the *stripped*
+    /// view (`reply[1..]`) is what `steam_get_serial` validates: `[0xAE, len, 0x01, ascii…]`.
+    #[test]
+    fn serial_reply_has_stripped_prefix() {
+        let r = serial_reply("PUNKTFUNK01");
+        assert_eq!(r[0], 0x00); // report id, stripped by steam_recv_report
+        assert_eq!(r[1], ID_GET_STRING_ATTRIBUTE); // becomes reply[0] after strip
+        assert!((1..=21).contains(&r[2]));
+        assert_eq!(r[3], ATTRIB_STR_UNIT_SERIAL);
+        assert_eq!(&r[4..4 + r[2] as usize], b"PUNKTFUNK01");
+    }
+
+    /// A `0xEB` rumble feature report parses to `(left_speed, right_speed)`; other commands don't.
+    #[test]
+    fn parse_rumble_feedback() {
+        // [report-id 0, 0xEB, len 9, 0, intensity(2), left(2), right(2), gains(2)]
+        let mut d = vec![0u8; 12];
+        d[1] = ID_TRIGGER_RUMBLE_CMD;
+        d[2] = 9;
+        d[6..8].copy_from_slice(&0x8000u16.to_le_bytes()); // left_speed
+        d[8..10].copy_from_slice(&0x4000u16.to_le_bytes()); // right_speed
+        assert_eq!(parse_steam_output(&d).rumble, Some((0x8000, 0x4000)));
+
+        let mut d = vec![0u8; 12];
+        d[1] = ID_SET_SETTINGS_VALUES; // a settings write — no rumble
+        assert_eq!(parse_steam_output(&d).rumble, None);
+    }
+
+    /// The shared real-USB Deck feature contract (gadget + usbip): the `0x83` GET_ATTRIBUTES reply
+    /// carries the 9-attribute blob with the per-instance unit id, and the `0xAE` reply carries the
+    /// Steam-accepted serial — both keyed off the host's last SET_REPORT command. A slip here is the
+    /// gamepad-evdev churn (Steam re-probing).
+    #[test]
+    fn deck_feature_reply_contract() {
+        let serial = deck_serial(0);
+        let unit_id = deck_unit_id(0);
+        assert_eq!(serial, "PFDK50460000"); // 12-char alphanumeric, derived from the unit id
+        assert_eq!(serial.len(), 12);
+
+        // 0x83 GET_ATTRIBUTES_VALUES: header + (0x0a, unit_id) at the 3rd attribute slot.
+        let r = feature_reply(&[ID_GET_ATTRIBUTES_VALUES], &serial, unit_id);
+        assert_eq!(r[0], ID_GET_ATTRIBUTES_VALUES);
+        assert_eq!(r[1], 0x2d);
+        assert_eq!(r[12], 0x0a); // 3rd attr id (slots at 2,7,12,…)
+        assert_eq!(
+            u32::from_le_bytes([r[13], r[14], r[15], r[16]]),
+            unit_id,
+            "unit serial attribute must carry the per-instance unit id"
+        );
+
+        // 0xAE GET_STRING_ATTRIBUTE: [0xAE, len, attr(0x01), ascii serial…].
+        let r = feature_reply(
+            &[ID_GET_STRING_ATTRIBUTE, 0, ATTRIB_STR_UNIT_SERIAL],
+            &serial,
+            unit_id,
+        );
+        assert_eq!(r[0], ID_GET_STRING_ATTRIBUTE);
+        assert_eq!(r[1] as usize, serial.len());
+        assert_eq!(r[2], ATTRIB_STR_UNIT_SERIAL);
+        assert_eq!(&r[3..3 + serial.len()], serial.as_bytes());
+
+        // Distinct pad indices get distinct unit ids + serials (no collision between virtual Decks).
+        assert_ne!(deck_unit_id(0), deck_unit_id(1));
+        assert_ne!(deck_serial(0), deck_serial(1));
+    }
+}

crates/punktfunk-host/src/inject/proto/steam_remap.rs

@@ -0,0 +1,149 @@
+//! Pure fallback-remap policy for the Steam Controller / Steam Deck rich inputs when the resolved
+//! host backend is **not** the virtual `hid-steam` device (DualSense / DualShock 4 / Xbox), so a
+//! client's Steam-only inputs aren't silently dropped — plus the cross-device motion rescale the
+//! Deck backend itself needs.
+//!
+//! Driven by the host's `PUNKTFUNK_STEAM_REMAP` env (`key=value`, `,`/`;`-separated, e.g.
+//! `paddles=stickclicks`). No I/O beyond [`RemapConfig::from_env`]; everything else is pure +
+//! unit-testable. The uinput Xbox pad already exposes the back grips as Elite paddles
+//! (`BTN_TRIGGER_HAPPY5-8`), so only the slot-less DualSense / DS4 backends fold them.
+
+use punktfunk_core::input::gamepad as gs;
+
+/// Where the four Steam back grips go on a backend with no native back-button HID slot.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Default)]
+pub enum PaddleFallback {
+    /// Drop them — the back buttons are simply absent on this pad. The honest default: don't fire
+    /// buttons the user didn't ask for. Set the env to map them instead.
+    #[default]
+    Drop,
+    /// L4/L5 → left-stick click, R4/R5 → right-stick click.
+    StickClicks,
+    /// L4/L5 → left bumper, R4/R5 → right bumper.
+    Shoulders,
+}
+
+/// Fallback-remap knobs parsed from `PUNKTFUNK_STEAM_REMAP`.
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub struct RemapConfig {
+    pub paddles: PaddleFallback,
+}
+
+impl RemapConfig {
+    /// Parse the host's `PUNKTFUNK_STEAM_REMAP` env (absent / unrecognized → defaults).
+    pub fn from_env() -> RemapConfig {
+        std::env::var("PUNKTFUNK_STEAM_REMAP")
+            .map(|s| RemapConfig::parse(&s))
+            .unwrap_or_default()
+    }
+
+    /// Pure parse of the `key=value[,key=value…]` string (the testable core of [`from_env`]).
+    pub fn parse(s: &str) -> RemapConfig {
+        let mut cfg = RemapConfig::default();
+        for kv in s.split([',', ';']) {
+            let mut it = kv.splitn(2, '=');
+            if let (Some(k), Some(v)) = (it.next(), it.next()) {
+                if k.trim().eq_ignore_ascii_case("paddles") {
+                    cfg.paddles = match v.trim().to_ascii_lowercase().as_str() {
+                        "stickclicks" | "l3r3" | "sticks" => PaddleFallback::StickClicks,
+                        "shoulders" | "lbrb" | "bumpers" => PaddleFallback::Shoulders,
+                        _ => PaddleFallback::Drop,
+                    };
+                }
+            }
+        }
+        cfg
+    }
+}
+
+/// Fold the wire back-grip bits (`BTN_PADDLE1..4`) into standard buttons per `policy` for a pad with
+/// no native back-button slot, clearing the paddle bits. Pure. PADDLE1/2/3/4 = R4/L4/R5/L5.
+pub fn fold_paddles(mut buttons: u32, policy: PaddleFallback) -> u32 {
+    let left = buttons & (gs::BTN_PADDLE2 | gs::BTN_PADDLE4) != 0; // L4 | L5
+    let right = buttons & (gs::BTN_PADDLE1 | gs::BTN_PADDLE3) != 0; // R4 | R5
+    buttons &= !(gs::BTN_PADDLE1 | gs::BTN_PADDLE2 | gs::BTN_PADDLE3 | gs::BTN_PADDLE4);
+    let (lbit, rbit) = match policy {
+        PaddleFallback::Drop => return buttons,
+        PaddleFallback::StickClicks => (gs::BTN_LS_CLICK, gs::BTN_RS_CLICK),
+        PaddleFallback::Shoulders => (gs::BTN_LB, gs::BTN_RB),
+    };
+    if left {
+        buttons |= lbit;
+    }
+    if right {
+        buttons |= rbit;
+    }
+    buttons
+}
+
+// Motion rescale. The wire uses the DualSense convention (20 LSB/°·s gyro, 10000 LSB/g accel — the
+// scale every client capture applies). The Steam Deck's `hid-steam` report wants 16 LSB/°·s and
+// 16384 LSB/g, so the Deck backend rescales; the DualSense / DS4 backends consume the wire 1:1.
+const GYRO_NUM: i32 = 16;
+const GYRO_DEN: i32 = 20;
+const ACCEL_NUM: i32 = 16384;
+const ACCEL_DEN: i32 = 10000;
+
+fn scale(v: i16, num: i32, den: i32) -> i16 {
+    ((v as i32 * num) / den).clamp(i16::MIN as i32, i16::MAX as i32) as i16
+}
+
+/// Rescale a wire (DualSense-convention) motion sample into the Steam Deck's `hid-steam` units.
+pub fn motion_wire_to_deck(gyro: [i16; 3], accel: [i16; 3]) -> ([i16; 3], [i16; 3]) {
+    (
+        gyro.map(|g| scale(g, GYRO_NUM, GYRO_DEN)),
+        accel.map(|a| scale(a, ACCEL_NUM, ACCEL_DEN)),
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn parse_paddle_policy() {
+        assert_eq!(RemapConfig::parse("").paddles, PaddleFallback::Drop);
+        assert_eq!(
+            RemapConfig::parse("paddles=stickclicks").paddles,
+            PaddleFallback::StickClicks
+        );
+        assert_eq!(
+            RemapConfig::parse("foo=bar; paddles = Shoulders").paddles,
+            PaddleFallback::Shoulders
+        );
+        assert_eq!(
+            RemapConfig::parse("paddles=nonsense").paddles,
+            PaddleFallback::Drop
+        );
+    }
+
+    #[test]
+    fn fold_paddles_maps_and_clears() {
+        // All four grips set + a real A button.
+        let b = gs::BTN_A | gs::BTN_PADDLE1 | gs::BTN_PADDLE2 | gs::BTN_PADDLE3 | gs::BTN_PADDLE4;
+        // Drop: paddle bits cleared, A preserved, nothing added.
+        assert_eq!(fold_paddles(b, PaddleFallback::Drop), gs::BTN_A);
+        // StickClicks: left grips → L3, right grips → R3.
+        assert_eq!(
+            fold_paddles(b, PaddleFallback::StickClicks),
+            gs::BTN_A | gs::BTN_LS_CLICK | gs::BTN_RS_CLICK
+        );
+        // Only a left grip (L4 = PADDLE2) → only the left bumper under Shoulders.
+        assert_eq!(
+            fold_paddles(gs::BTN_PADDLE2, PaddleFallback::Shoulders),
+            gs::BTN_LB
+        );
+    }
+
+    #[test]
+    fn motion_rescale_to_deck_units() {
+        // gyro × 16/20 = 0.8; accel × 16384/10000 = 1.6384.
+        let (g, a) = motion_wire_to_deck([1000, -2000, 0], [10000, -5000, 0]);
+        assert_eq!(g, [800, -1600, 0]);
+        assert_eq!(a, [16384, -8192, 0]);
+        // Saturates rather than wraps.
+        let (_, a) = motion_wire_to_deck([0; 3], [32767, i16::MIN, 0]);
+        assert_eq!(a[0], i16::MAX);
+        assert_eq!(a[1], i16::MIN);
+    }
+}

crates/punktfunk-host/src/inject/windows/dualsense_windows.rs

@@ -385,7 +385,9 @@ impl DualSenseWindowsManager {
     /// Apply one rich client→host event (touchpad contact / motion sample) to an existing pad.
     pub fn apply_rich(&mut self, rich: RichInput) {
         let idx = match rich {
-            RichInput::Touchpad { pad, .. } | RichInput::Motion { pad, .. } => pad as usize,
+            RichInput::Touchpad { pad, .. }
+            | RichInput::Motion { pad, .. }
+            | RichInput::TouchpadEx { pad, .. } => pad as usize,
         };
         if idx >= MAX_PADS || self.pads[idx].is_none() {
             return;
@@ -409,6 +411,26 @@ impl DualSenseWindowsManager {
                 self.state[idx].gyro = gyro;
                 self.state[idx].accel = accel;
             }
+            RichInput::TouchpadEx {
+                surface,
+                finger,
+                touch,
+                x,
+                y,
+                ..
+            } => {
+                // A Steam right/single pad maps onto the one DualSense touchpad (signed centre-0 →
+                // 0..=65535); surface 1 (the Steam left pad) has no DualSense equivalent.
+                if surface != 1 {
+                    let slot = (finger as usize).min(1);
+                    let n = |v: i16| ((v as i32) + 32768) as u32;
+                    let t = &mut self.state[idx].touch[slot];
+                    t.active = touch;
+                    t.id = slot as u8;
+                    t.x = (n(x) * (DS_TOUCH_W - 1) as u32 / u16::MAX as u32) as u16;
+                    t.y = (n(y) * (DS_TOUCH_H - 1) as u32 / u16::MAX as u32) as u16;
+                }
+            }
         }
         self.write(idx);
     }

crates/punktfunk-host/src/inject/windows/dualshock4_windows.rs

@@ -186,7 +186,9 @@ impl DualShock4WindowsManager {
     /// Apply one rich client→host event (touchpad contact / motion sample) to an existing pad.
     pub fn apply_rich(&mut self, rich: RichInput) {
         let idx = match rich {
-            RichInput::Touchpad { pad, .. } | RichInput::Motion { pad, .. } => pad as usize,
+            RichInput::Touchpad { pad, .. }
+            | RichInput::Motion { pad, .. }
+            | RichInput::TouchpadEx { pad, .. } => pad as usize,
         };
         if idx >= MAX_PADS || self.pads[idx].is_none() {
             return;
@@ -210,6 +212,26 @@ impl DualShock4WindowsManager {
                 self.state[idx].gyro = gyro;
                 self.state[idx].accel = accel;
             }
+            RichInput::TouchpadEx {
+                surface,
+                finger,
+                touch,
+                x,
+                y,
+                ..
+            } => {
+                // A Steam right/single pad maps onto the one DS4 touchpad (signed centre-0 →
+                // 0..=65535); surface 1 (the Steam left pad) has no DS4 equivalent.
+                if surface != 1 {
+                    let slot = (finger as usize).min(1);
+                    let n = |v: i16| ((v as i32) + 32768) as u32;
+                    let t = &mut self.state[idx].touch[slot];
+                    t.active = touch;
+                    t.id = slot as u8;
+                    t.x = (n(x) * (DS4_TOUCH_W - 1) as u32 / u16::MAX as u32) as u16;
+                    t.y = (n(y) * (DS4_TOUCH_H - 1) as u32 / u16::MAX as u32) as u16;
+                }
+            }
         }
         self.write(idx);
     }

crates/punktfunk-host/src/inject/windows/gamepad_raii.rs

@@ -21,10 +21,18 @@ use windows::Win32::System::Memory::{
     MEMORY_MAPPED_VIEW_ADDRESS, PAGE_READWRITE,
 };
 
-/// A named, anonymous (pagefile-backed) shared section + its mapped read/write view, created with the
+/// A named, anonymous (pagefile-backed) shared section + its mapped read/write view. RAII: drop unmaps
-/// permissive `D:(A;;GA;;;WD)` SDDL the restricted-token driver needs to open it. RAII: drop unmaps the
+/// the view, then the [`OwnedHandle`] closes the section handle (in that order). Replaces the three
-/// view, then the [`OwnedHandle`] closes the section handle (in that order). Replaces the three backends'
+/// backends' hand-duplicated `CreateFileMappingW` + `MapViewOfFile` + manual `Drop`.
-/// hand-duplicated `CreateFileMappingW` + `MapViewOfFile` + manual `Drop`.
+///
+/// SDDL `D:(A;;GA;;;SY)(A;;GA;;;LS)`: GENERIC_ALL to **SYSTEM** (the host creates the section and
+/// writes the live HID input report into it) and **LocalService** (the account the UMDF driver's
+/// WUDFHost runs under, which reads it). The old SDDL granted **Everyone** (`WD`) — on the (mistaken)
+/// assumption the driver needed a restricted token's broad access — letting any local user
+/// `OpenFileMapping` the section to inject controller input or tamper the trusted channel
+/// (security-review 2026-06-28 #5). Verified on the RTX box (2026-06-29): the WUDFHost token is
+/// `S-1-5-19` (LocalService), SYSTEM integrity, with **zero restricted SIDs** — so scoping to SY+LS is
+/// sufficient for the driver and excludes normal (medium-IL, non-service) user processes.
 pub(super) struct Shm {
     /// Owns the section handle (closed on drop). Held only for ownership — never read after construction.
     _handle: OwnedHandle,
@@ -40,7 +48,7 @@ impl Shm {
         // exit — acceptable for a host-lifetime object).
         unsafe {
             ConvertStringSecurityDescriptorToSecurityDescriptorW(
-                w!("D:(A;;GA;;;WD)"),
+                w!("D:(A;;GA;;;SY)(A;;GA;;;LS)"),
                 SDDL_REVISION_1,
                 &mut psd,
                 None,

crates/punktfunk-host/src/main.rs

@@ -56,9 +56,6 @@ mod spike;
 mod stats_recorder;
 mod vdisplay;
 #[cfg(target_os = "windows")]
-#[path = "windows/wgc_helper.rs"]
-mod wgc_helper;
-#[cfg(target_os = "windows")]
 #[path = "windows/win_adapter.rs"]
 mod win_adapter;
 #[cfg(target_os = "windows")]
@@ -392,35 +389,6 @@ fn real_main() -> Result<()> {
                 paired_store: None,
             })
         }
-        // USER-session WGC helper (Windows two-process secure-desktop design): capture the EXISTING
-        // SudoVDA via WGC + NVENC, stream AUs on stdout to the SYSTEM host. Spawned by the host
-        // (CreateProcessAsUser), not run by hand. See design/archive/windows-secure-desktop.md.
-        #[cfg(target_os = "windows")]
-        Some("wgc-helper") => {
-            let get = |flag: &str| {
-                args.iter()
-                    .skip_while(|a| *a != flag)
-                    .nth(1)
-                    .map(String::as_str)
-            };
-            let (width, height, fps) = get("--mode")
-                .and_then(|m| {
-                    let p: Vec<u32> = m.split('x').filter_map(|s| s.parse().ok()).collect();
-                    (p.len() == 3).then(|| (p[0], p[1], p[2]))
-                })
-                .unwrap_or((1920, 1080, 60));
-            wgc_helper::run(wgc_helper::HelperOptions {
-                target_id: get("--target-id").and_then(|s| s.parse().ok()).unwrap_or(0),
-                gdi_name: get("--gdi").unwrap_or("").to_string(),
-                width,
-                height,
-                fps,
-                bitrate_kbps: get("--bitrate")
-                    .and_then(|s| s.parse().ok())
-                    .unwrap_or(20000),
-                bit_depth: get("--bit-depth").and_then(|s| s.parse().ok()).unwrap_or(8),
-            })
-        }
         // Windows service control: install/uninstall/start/stop/status + the SCM `run` entry point.
         // Replaces the ad-hoc launch chain — `service install` registers an auto-start SYSTEM service
         // that launches the host into the active interactive session.

crates/punktfunk-host/src/mgmt.rs

@@ -394,6 +394,12 @@ struct ArmNativePairing {
     /// Window length in seconds (default 120; clamped to 15–600).
     #[schema(example = 120)]
     ttl_secs: Option<u32>,
+    /// Optional: bind the window to ONE device fingerprint (hex SHA-256, e.g. from a pending knock).
+    /// When set, only a pairing attempt from that fingerprint consumes the window — so an unpaired
+    /// LAN peer can neither pair nor burn a window armed for a specific device (security-review #9).
+    /// Omit for an unbound window (any device may use the PIN — trusted-LAN only).
+    #[schema(example = "9f86d081884c7d659a2feaa0c55ad015a3bf4f1b2b0b822cd15d6c15b0f00a08")]
+    fingerprint: Option<String>,
 }
 
 /// A paired native (punktfunk/1) client.
@@ -879,8 +885,21 @@ async fn arm_native_pairing(
         );
     };
     let ttl = req.ttl_secs.unwrap_or(120).clamp(15, 600);
-    let _pin = np.arm(std::time::Duration::from_secs(ttl as u64));
+    // A bound window (operator selected a specific device) is DoS-proof: only that fingerprint can
-    tracing::info!(ttl_secs = ttl, "management API: native pairing armed");
+    // consume it (#9). An unbound window (no fingerprint) keeps the legacy any-device behavior.
+    let bound = req
+        .fingerprint
+        .as_deref()
+        .map(str::trim)
+        .filter(|s| !s.is_empty())
+        .map(|s| s.to_ascii_lowercase());
+    let bound_to_device = bound.is_some();
+    let _pin = np.arm_for(std::time::Duration::from_secs(ttl as u64), bound);
+    tracing::info!(
+        ttl_secs = ttl,
+        bound_to_device,
+        "management API: native pairing armed"
+    );
     Json(native_status(&st)).into_response()
 }
 
@@ -1975,8 +1994,8 @@ mod tests {
         assert_eq!(b.as_array().unwrap().len(), 0);
 
         // Two devices knock (what the QUIC gate records); they appear in the list.
-        np.note_pending("Enrico's MacBook", "aa11");
+        np.note_pending("Enrico's MacBook", "aa11", None);
-        np.note_pending("device bb22cc33", "bb22");
+        np.note_pending("device bb22cc33", "bb22", None);
         let (_, b) = send(&app, get_req("/api/v1/native/pending")).await;
         assert_eq!(b.as_array().unwrap().len(), 2);
         assert_eq!(b[0]["name"], "Enrico's MacBook");

crates/punktfunk-host/src/native_pairing.rs

@@ -8,6 +8,7 @@
 //! armed on demand for a short window — rather than accepting one.
 
 use anyhow::Result;
+use std::net::IpAddr;
 use std::path::PathBuf;
 use std::sync::Mutex;
 use std::time::{Duration, Instant};
@@ -42,10 +43,29 @@ struct PairedState {
 
 /// The current arming window. `pin == None` ⇒ disarmed. `expires_at == None` ⇒ armed with no
 /// expiry (the CLI `--allow-pairing` flag); `Some(t)` ⇒ a web-armed window that auto-disarms.
+///
+/// `bound_fp == Some(fp)` ⇒ the window is **bound to one operator-selected device fingerprint**:
+/// only a pairing attempt from that fingerprint may consume it (security-review 2026-06-28 #9). This
+/// closes the window-burn DoS — an unpaired LAN peer cannot consume a window armed for a specific
+/// device, because the QUIC client-auth proves cert possession (it can't forge the bound fingerprint).
+/// `None` ⇒ unbound (the CLI flag / a console "arm open"): any well-formed attempt consumes it (the
+/// legacy behavior, retaining the window-burn DoS — acceptable only on a trusted LAN).
 #[derive(Default)]
 struct Armed {
     pin: Option<String>,
     expires_at: Option<Instant>,
+    bound_fp: Option<String>,
+}
+
+/// The result of resolving the armed PIN for a specific client fingerprint ([`NativePairing::pin_for_attempt`]).
+pub enum PinAttempt {
+    /// No window is armed (disarmed/expired) — reject; do not run the ceremony.
+    Disarmed,
+    /// A window IS armed but **bound to a different fingerprint** — reject WITHOUT consuming it, so
+    /// an unrelated (attacker) fingerprint can't burn the operator's armed window (#9).
+    BoundToOther,
+    /// Proceed: the PIN to run the ceremony with (the window is unbound, or bound to this fingerprint).
+    Pin(String),
 }
 
 /// An unpaired (but identified) device that knocked on a pairing-required host — held for
@@ -57,6 +77,13 @@ struct Pending {
     name: String,
     fp_hex: String,
     requested_at: Instant,
+    /// QUIC-validated source address of the knock — used for the per-source cap (#13), so one host
+    /// can't fill the queue. `None` if unknown (e.g. tests / a caller that doesn't supply it).
+    src_ip: Option<IpAddr>,
+    /// True while a connection is held open in [`NativePairing::wait_for_decision`] for this knock.
+    /// A live parked knock is a genuine device waiting for the operator — eviction skips it unless
+    /// every entry is parked, so a cert-rotating flood can't evict the device being onboarded (#13).
+    parked: bool,
 }
 
 #[derive(Default)]
@@ -94,6 +121,10 @@ pub enum PairingDecision {
 const PENDING_TTL: Duration = Duration::from_secs(10 * 60);
 /// Cap on the pending list — a LAN scanner must not grow it unboundedly. Oldest entries drop.
 const PENDING_CAP: usize = 32;
+/// Max pending knocks one source IP may occupy, so a single host can't fill the whole queue and hide
+/// / evict a genuine device's knock (security-review 2026-06-28 #13). The QUIC path is address-
+/// validated, so the source IP isn't off-path spoofable; an attacker would need that many real hosts.
+const MAX_PENDING_PER_IP: usize = 4;
 
 /// Shared native-pairing state: the arming PIN window + the persistent trust store + the
 /// pending-approval queue.
@@ -209,6 +240,7 @@ impl NativePairing {
             Armed {
                 pin: Some(fixed_pin.unwrap_or_else(random_pin)),
                 expires_at: None,
+                bound_fp: None,
             }
         } else {
             Armed::default()
@@ -221,16 +253,43 @@ impl NativePairing {
         })
     }
 
-    /// Arm pairing with a fresh random PIN, valid for `ttl`. Returns the PIN to display.
+    /// Arm pairing with a fresh random PIN, valid for `ttl`, **unbound** (any well-formed attempt
+    /// consumes it). Returns the PIN to display. Prefer [`Self::arm_for`] with a specific device
+    /// fingerprint on untrusted LANs — an unbound window is burnable by any peer (#9).
     pub fn arm(&self, ttl: Duration) -> String {
+        self.arm_for(ttl, None)
+    }
+
+    /// Arm pairing with a fresh random PIN, valid for `ttl`. If `bound_fp` is `Some`, the window is
+    /// bound to that device fingerprint: only a pairing attempt from it consumes the window, so an
+    /// unrelated (attacker) fingerprint can neither pair nor burn the window (#9). Returns the PIN.
+    pub fn arm_for(&self, ttl: Duration, bound_fp: Option<String>) -> String {
         let pin = random_pin();
         *self.arm.lock().unwrap() = Armed {
             pin: Some(pin.clone()),
             expires_at: Some(Instant::now() + ttl),
+            bound_fp,
         };
         pin
     }
 
+    /// Resolve the PIN for an attempt from `client_fp_hex`, honoring fingerprint binding (#9):
+    /// `Disarmed` if no window is armed; `BoundToOther` if a window is armed but bound to a different
+    /// fingerprint (the caller MUST reject without consuming it); else `Pin` to run the ceremony.
+    pub fn pin_for_attempt(&self, client_fp_hex: &str) -> PinAttempt {
+        let mut arm = self.arm.lock().unwrap();
+        Self::expire(&mut arm);
+        match &arm.pin {
+            None => PinAttempt::Disarmed,
+            Some(pin) => match &arm.bound_fp {
+                Some(bound) if !bound.eq_ignore_ascii_case(client_fp_hex) => {
+                    PinAttempt::BoundToOther
+                }
+                _ => PinAttempt::Pin(pin.clone()),
+            },
+        }
+    }
+
     /// Disarm pairing (no new ceremonies accepted).
     pub fn disarm(&self) {
         *self.arm.lock().unwrap() = Armed::default();
@@ -342,11 +401,30 @@ impl NativePairing {
             .retain(|p| p.requested_at.elapsed() < PENDING_TTL);
     }
 
-    /// Record an unpaired device's knock for delegated approval. Re-knocks from the same
+    /// Pick the entry to evict, optionally restricted to a single source IP: the least-recently-active
-    /// fingerprint refresh the existing entry in place (same id; a connect-retry loop must not spam
+    /// **non-parked** entry (a live parked knock is a genuine device awaiting the operator — never
-    /// the list); a fresh fingerprint gets a new id, evicting the **least-recently-active** entry
+    /// evict it under load); only if every candidate is parked does it fall back to the oldest of
-    /// past [`PENDING_CAP`]. The name is sanitized (untrusted; see [`sanitize_device_name`]).
+    /// those (#13). Returns the index, or `None` if there's nothing to evict.
-    pub fn note_pending(&self, name: &str, fp_hex: &str) {
+    fn evict_index(items: &[Pending], only_ip: Option<IpAddr>) -> Option<usize> {
+        let pick = |allow_parked: bool| {
+            items
+                .iter()
+                .enumerate()
+                .filter(|(_, p)| only_ip.is_none_or(|ip| p.src_ip == Some(ip)))
+                .filter(|(_, p)| allow_parked || !p.parked)
+                .min_by_key(|(_, p)| p.requested_at)
+                .map(|(i, _)| i)
+        };
+        pick(false).or_else(|| pick(true))
+    }
+
+    /// Record an unpaired device's knock for delegated approval. Re-knocks from the same fingerprint
+    /// refresh the existing entry in place (same id; a connect-retry loop must not spam the list). A
+    /// fresh fingerprint gets a new id; the queue is bounded two ways so a flood can't crowd out a
+    /// genuine knock (#13): a **per-source-IP cap** ([`MAX_PENDING_PER_IP`]) means one host can hold at
+    /// most a few slots, and the global [`PENDING_CAP`] evicts the least-recently-active **non-parked**
+    /// entry (never a live, held-open parked knock). The name is sanitized (untrusted).
+    pub fn note_pending(&self, name: &str, fp_hex: &str, src_ip: Option<IpAddr>) {
         let name = sanitize_device_name(name, fp_hex);
         let mut pending = self.pending.lock().unwrap();
         Self::expire_pending(&mut pending);
@@ -357,19 +435,31 @@ impl NativePairing {
         {
             p.requested_at = Instant::now();
             p.name = name;
+            if p.src_ip.is_none() {
+                p.src_ip = src_ip;
+            }
             return;
         }
-        if pending.items.len() >= PENDING_CAP {
+        // Per-source-IP cap: a single host can't occupy more than MAX_PENDING_PER_IP slots — evict its
-            // Evict the least-recently-active entry. NOT index 0: the in-place refresh above means
+        // own oldest entry first so it can't crowd out other devices' knocks (#13).
-            // Vec order no longer tracks recency, so pick the minimum `requested_at` explicitly.
+        if let Some(ip) = src_ip {
-            if let Some(at) = pending
+            if pending
                 .items
                 .iter()
-                .enumerate()
+                .filter(|p| p.src_ip == Some(ip))
-                .min_by_key(|(_, p)| p.requested_at)
+                .count()
-                .map(|(i, _)| i)
+                >= MAX_PENDING_PER_IP
             {
-                pending.items.remove(at);
+                if let Some(i) = Self::evict_index(&pending.items, Some(ip)) {
+                    pending.items.remove(i);
+                }
+            }
+        }
+        // Global cap: evict the least-recently-active non-parked entry (Vec order no longer tracks
+        // recency after in-place refreshes, so pick explicitly).
+        if pending.items.len() >= PENDING_CAP {
+            if let Some(i) = Self::evict_index(&pending.items, None) {
+                pending.items.remove(i);
             }
         }
         let id = pending.next_id;
@@ -379,9 +469,24 @@ impl NativePairing {
             name,
             fp_hex: fp_hex.to_string(),
             requested_at: Instant::now(),
+            src_ip,
+            parked: false,
         });
     }
 
+    /// Mark/unmark the pending entry for `fp_hex` as having a live parked waiter (no-op if it's gone).
+    /// A parked entry is protected from eviction under load (#13).
+    fn set_parked(&self, fp_hex: &str, parked: bool) {
+        let mut pending = self.pending.lock().unwrap();
+        if let Some(p) = pending
+            .items
+            .iter_mut()
+            .find(|p| p.fp_hex.eq_ignore_ascii_case(fp_hex))
+        {
+            p.parked = parked;
+        }
+    }
+
     /// The devices currently awaiting approval (for the management API).
     pub fn pending(&self) -> Vec<PendingRequest> {
         let mut pending = self.pending.lock().unwrap();
@@ -462,6 +567,23 @@ impl NativePairing {
     /// to keep the knocking connection open until a human clicks Approve — so the device pairs and
     /// streams with no reconnect (delegated approval, roadmap §8b-1).
     pub async fn wait_for_decision(&self, fp_hex: &str, timeout: Duration) -> PairingDecision {
+        // Mark this knock parked so a cert-rotating flood can't evict the genuine, held-open
+        // connection out of the pending queue while the operator decides (#13). Cleared on every
+        // exit path by the guard's Drop.
+        self.set_parked(fp_hex, true);
+        struct ParkGuard<'a> {
+            np: &'a NativePairing,
+            fp: &'a str,
+        }
+        impl Drop for ParkGuard<'_> {
+            fn drop(&mut self) {
+                self.np.set_parked(self.fp, false);
+            }
+        }
+        let _park = ParkGuard {
+            np: self,
+            fp: fp_hex,
+        };
         let deadline = tokio::time::Instant::now() + timeout;
         loop {
             // Arm the wakeup BEFORE re-reading state, and `enable()` it, so an approve/deny that
@@ -548,8 +670,8 @@ mod tests {
 
         // A knock appears; a re-knock from the same fingerprint refreshes (same id, new name)
         // instead of duplicating.
-        np.note_pending("device aa11", "AA11");
+        np.note_pending("device aa11", "AA11", None);
-        np.note_pending("Bedroom TV", "aa11");
+        np.note_pending("Bedroom TV", "aa11", None);
         let pend = np.pending();
         assert_eq!(pend.len(), 1, "re-knock dedups by fingerprint");
         assert_eq!(pend[0].name, "Bedroom TV");
@@ -562,7 +684,7 @@ mod tests {
         assert!(!np.is_paired("aa11"));
 
         // Approve pairs the fingerprint (operator label wins) and clears the entry.
-        np.note_pending("device bb22", "BB22");
+        np.note_pending("device bb22", "BB22", None);
         let id = np.pending()[0].id;
         assert!(
             np.approve_pending(9999, None).unwrap().is_none(),
@@ -578,8 +700,11 @@ mod tests {
         assert_eq!(np.list()[0].name, "Living Room");
 
         // The cap evicts the oldest knock.
+        // Flood from many DISTINCT source IPs (so the per-IP cap doesn't kick in) → the global cap
+        // holds at PENDING_CAP, evicting the oldest non-parked entries first.
         for i in 0..(PENDING_CAP + 3) {
-            np.note_pending("flood", &format!("f{i:03}"));
+            let ip = IpAddr::from([10, 0, (i / 256) as u8, (i % 256) as u8]);
+            np.note_pending("flood", &format!("f{i:03}"), Some(ip));
         }
         let pend = np.pending();
         assert_eq!(pend.len(), PENDING_CAP);
@@ -610,7 +735,7 @@ mod tests {
         let p = temp();
         let _ = std::fs::remove_file(&p);
         let np = NativePairing::load_with(Some(p.clone()), None, false).unwrap();
-        np.note_pending("Knocker", "cc44");
+        np.note_pending("Knocker", "cc44", None);
         assert_eq!(np.pending().len(), 1);
         // Pairing the same fingerprint (e.g. via the PIN ceremony) drops the stale pending entry.
         np.add("Knocker", "CC44").unwrap();
@@ -656,7 +781,7 @@ mod tests {
         let np = Arc::new(NativePairing::load_with(Some(p.clone()), None, false).unwrap());
 
         // TimedOut: a parked knock with no decision returns TimedOut; the entry survives.
-        np.note_pending("Knocker", "ab01");
+        np.note_pending("Knocker", "ab01", None);
         let d = np
             .wait_for_decision("ab01", Duration::from_millis(80))
             .await;
@@ -681,7 +806,7 @@ mod tests {
         assert!(np.is_paired("ab01"));
 
         // Denied: denying WHILE parked wakes the waiter with Denied (not held until timeout).
-        np.note_pending("Knock2", "cd02");
+        np.note_pending("Knock2", "cd02", None);
         let np3 = np.clone();
         let waiter =
             tokio::spawn(
@@ -703,4 +828,62 @@ mod tests {
         assert_eq!(d, PairingDecision::Approved);
         let _ = std::fs::remove_file(&p);
     }
+
+    /// #9: a window can be bound to one operator-selected fingerprint, so an unrelated (attacker)
+    /// fingerprint can neither pair nor BURN the window (it's rejected without a PIN).
+    #[test]
+    fn armed_pin_is_fingerprint_bindable() {
+        let p = temp();
+        let _ = std::fs::remove_file(&p);
+        let np = NativePairing::load_with(Some(p.clone()), None, false).unwrap();
+        // Unbound: any fingerprint resolves to the PIN (legacy behavior).
+        let pin = np.arm(Duration::from_secs(60));
+        assert!(matches!(np.pin_for_attempt("aa11"), PinAttempt::Pin(x) if x == pin));
+        assert!(matches!(np.pin_for_attempt("bb22"), PinAttempt::Pin(_)));
+        // Bound to AA11: only that fp (case-insensitive) gets the PIN; another fp is BoundToOther —
+        // the caller rejects it WITHOUT consuming the window.
+        let pin = np.arm_for(Duration::from_secs(60), Some("AA11".into()));
+        assert!(matches!(np.pin_for_attempt("aa11"), PinAttempt::Pin(x) if x == pin));
+        assert!(matches!(
+            np.pin_for_attempt("bb22"),
+            PinAttempt::BoundToOther
+        ));
+        np.disarm();
+        assert!(matches!(np.pin_for_attempt("aa11"), PinAttempt::Disarmed));
+        let _ = std::fs::remove_file(&p);
+    }
+
+    /// #13: one source IP can't exceed the per-IP cap, and a parked (held-open) genuine knock is
+    /// never evicted by a flood — even one that fills the global cap from many distinct IPs.
+    #[test]
+    fn pending_per_ip_cap_and_parked_protection() {
+        let p = temp();
+        let _ = std::fs::remove_file(&p);
+        let np = NativePairing::load_with(Some(p.clone()), None, false).unwrap();
+        // Per-IP cap: one source flooding distinct fingerprints holds at most MAX_PENDING_PER_IP.
+        let attacker = IpAddr::from([192, 168, 1, 66]);
+        for i in 0..20 {
+            np.note_pending("flood", &format!("atk{i:03}"), Some(attacker));
+        }
+        assert_eq!(
+            np.pending().len(),
+            MAX_PENDING_PER_IP,
+            "one IP can't exceed the per-IP cap"
+        );
+        // A genuine knock from a different IP, parked (a live held-open connection), survives a flood
+        // from many distinct IPs that fills the global cap.
+        let legit = IpAddr::from([192, 168, 1, 50]);
+        np.note_pending("Living Room", "legit01", Some(legit));
+        np.set_parked("legit01", true);
+        for i in 0..(PENDING_CAP * 2) {
+            let ip = IpAddr::from([10, 0, (i / 256) as u8, (i % 256) as u8]);
+            np.note_pending("flood2", &format!("g{i:04}"), Some(ip));
+        }
+        assert!(
+            np.pending_contains("legit01"),
+            "a parked, held-open knock is never evicted by a flood"
+        );
+        assert!(np.pending().len() <= PENDING_CAP, "global cap still holds");
+        let _ = std::fs::remove_file(&p);
+    }
 }

crates/punktfunk-host/src/punktfunk1.rs

@@ -532,10 +532,25 @@ async fn serve_session(
         .await
         .map_err(|_| anyhow!("first message timeout"))??;
     if let Ok(req) = PairRequest::decode(&first) {
-        // Read the live arming PIN per attempt, so a window that lapsed no longer pairs.
+        // The client fingerprint (cert possession is proven by the QUIC handshake) is needed to honor
-        let pin = np
+        // a fingerprint-bound PIN window (#9): a window the operator armed for a SPECIFIC device must
-            .current_pin()
+        // not be consumable — or burnable — by any other fingerprint.
-            .context("pairing not armed (arm it in the console, or start with --allow-pairing)")?;
+        let client_fp = endpoint::peer_fingerprint(&conn)
+            .ok_or_else(|| anyhow!("pairing requires the client to present a certificate"))?;
+        let client_fp_hex = fingerprint_hex(&client_fp);
+        // Resolve the live arming PIN per attempt (so a lapsed window no longer pairs), honoring any
+        // fingerprint binding.
+        let pin = match np.pin_for_attempt(&client_fp_hex) {
+            crate::native_pairing::PinAttempt::Pin(pin) => pin,
+            crate::native_pairing::PinAttempt::Disarmed => anyhow::bail!(
+                "pairing not armed (arm it in the console, or start with --allow-pairing)"
+            ),
+            // Armed for a DIFFERENT device — reject without running the ceremony, so this attempt does
+            // NOT consume (burn) the operator's window for the device they actually selected (#9).
+            crate::native_pairing::PinAttempt::BoundToOther => anyhow::bail!(
+                "pairing is armed for a different device — this attempt does not consume the window"
+            ),
+        };
         {
             let mut last = last_pairing.lock().unwrap();
             if let Some(t) = *last {
@@ -589,7 +604,9 @@ async fn serve_session(
             );
             tracing::info!(name = %label, fingerprint = %fp_hex,
                 "unpaired device knocked — parking connection for delegated approval in the console");
-            np.note_pending(&label, &fp_hex);
+            // Record the QUIC-validated source IP so the pending queue's per-source cap can stop one
+            // host from flooding/evicting genuine knocks (#13).
+            np.note_pending(&label, &fp_hex, Some(peer.ip()));
             // Free the session slot while a human decides — a parked knock must not hold an NVENC
             // permit (a handful of parked knocks would otherwise block every real session).
             drop(permit);
@@ -738,14 +755,18 @@ async fn serve_session(
         // opens a tiny encoder; it runs only when both opt-ins are set and is cached after the first.
         let host_wants_444 = crate::config::config().four_four_four;
         let client_supports_444 = hello.video_caps & punktfunk_core::quic::VIDEO_CAP_444 != 0;
-        let single_process = crate::session_plan::resolve_topology()
+        // The active capturer must be able to deliver a full-chroma (RGB) source — the honest-downgrade
-            == crate::session_plan::SessionTopology::SingleProcess;
+        // gate. Linux's portal capturer can; the Windows IDD-push path delivers subsampled NV12/P010
+        // today (full-chroma IDD-push capture is a follow-up), so it returns false there and the host
+        // negotiates 4:2:0. (Replaces the old `single_process` gate — single-process is now the only
+        // topology, and 4:4:4 routed to DDA, which was removed.)
+        let capture_supports_444 = crate::capture::capturer_supports_444();
         // The GPU probe opens a real (tiny) encoder on first use, so run it off the reactor like the
         // compositor probe above (blocking probes → spawn_blocking). Short-circuit so it only runs when
         // the cheap gates already pass. The result is cached process-wide (a negative latches until
         // restart — acceptable: a GPU either supports HEVC 4:4:4 or it doesn't, and a transient open
         // failure here is rare since the session's own encoder isn't open yet).
-        let gpu_supports_444 = if host_wants_444 && client_supports_444 && single_process {
+        let gpu_supports_444 = if host_wants_444 && client_supports_444 && capture_supports_444 {
             tokio::task::spawn_blocking(|| {
                 crate::encode::can_encode_444(crate::encode::Codec::H265)
             })
@@ -763,7 +784,7 @@ async fn serve_session(
             chroma = ?chroma,
             host_wants_444,
             client_supports_444,
-            single_process,
+            capture_supports_444,
             "encode chroma"
         );
 
@@ -1382,6 +1403,8 @@ enum PadBackend {
     DualSense(crate::inject::dualsense::DualSenseManager),
     #[cfg(target_os = "linux")]
     DualShock4(crate::inject::dualshock4::DualShock4Manager),
+    #[cfg(target_os = "linux")]
+    SteamDeck(crate::inject::steam_controller::SteamControllerManager),
     #[cfg(target_os = "windows")]
     DualSenseWindows(crate::inject::dualsense_windows::DualSenseWindowsManager),
     #[cfg(target_os = "windows")]
@@ -1403,6 +1426,12 @@ impl PadBackend {
                 tracing::info!("gamepad backend: virtual DualShock 4 (UHID hid-playstation)");
                 return PadBackend::DualShock4(crate::inject::dualshock4::DualShock4Manager::new());
             }
+            GamepadPref::SteamDeck => {
+                tracing::info!("gamepad backend: virtual Steam Deck (UHID hid-steam)");
+                return PadBackend::SteamDeck(
+                    crate::inject::steam_controller::SteamControllerManager::new(),
+                );
+            }
             GamepadPref::XboxOne => {
                 tracing::info!("gamepad backend: uinput X-Box One/Series pad");
                 return PadBackend::Xbox360(crate::inject::gamepad::GamepadManager::with_identity(
@@ -1438,6 +1467,8 @@ impl PadBackend {
             PadBackend::DualSense(m) => m.handle(ev),
             #[cfg(target_os = "linux")]
             PadBackend::DualShock4(m) => m.handle(ev),
+            #[cfg(target_os = "linux")]
+            PadBackend::SteamDeck(m) => m.handle(ev),
             #[cfg(target_os = "windows")]
             PadBackend::DualSenseWindows(m) => m.handle(ev),
             #[cfg(target_os = "windows")]
@@ -1454,6 +1485,8 @@ impl PadBackend {
             PadBackend::DualSense(m) => m.apply_rich(_rich),
             #[cfg(target_os = "linux")]
             PadBackend::DualShock4(m) => m.apply_rich(_rich),
+            #[cfg(target_os = "linux")]
+            PadBackend::SteamDeck(m) => m.apply_rich(_rich),
             #[cfg(target_os = "windows")]
             PadBackend::DualSenseWindows(m) => m.apply_rich(_rich),
             #[cfg(target_os = "windows")]
@@ -1479,6 +1512,8 @@ impl PadBackend {
             PadBackend::DualSense(m) => m.pump(rumble, hidout),
             #[cfg(target_os = "linux")]
             PadBackend::DualShock4(m) => m.pump(rumble, hidout),
+            #[cfg(target_os = "linux")]
+            PadBackend::SteamDeck(m) => m.pump(rumble, hidout),
             #[cfg(target_os = "windows")]
             PadBackend::DualSenseWindows(m) => m.pump(rumble, hidout),
             #[cfg(target_os = "windows")]
@@ -1498,6 +1533,8 @@ impl PadBackend {
             PadBackend::DualSense(m) => m.heartbeat(std::time::Duration::from_millis(8)),
             #[cfg(target_os = "linux")]
             PadBackend::DualShock4(m) => m.heartbeat(std::time::Duration::from_millis(8)),
+            #[cfg(target_os = "linux")]
+            PadBackend::SteamDeck(m) => m.heartbeat(std::time::Duration::from_millis(8)),
             #[cfg(target_os = "windows")]
             PadBackend::DualSenseWindows(m) => m.heartbeat(std::time::Duration::from_millis(8)),
             #[cfg(target_os = "windows")]
@@ -1877,10 +1914,94 @@ fn pick_gamepad(pref: GamepadPref, env: Option<&str>, linux: bool, windows: bool
         // One/Series: a real, distinct uinput identity on Linux; folded into the 360 backend on
         // Windows (XInput can't tell them apart anyway).
         GamepadPref::XboxOne if linux => GamepadPref::XboxOne,
+        // Steam Deck: Linux UHID hid-steam. The classic Steam Controller's backend isn't built yet,
+        // so it folds to Xbox360 for now (Windows Steam devices are M7).
+        GamepadPref::SteamDeck if linux => GamepadPref::SteamDeck,
         _ => GamepadPref::Xbox360,
     }
 }
 
+/// Runtime degrade for the Linux UHID backends (DualSense / DualShock 4 / Steam Deck): if
+/// `/dev/uhid` can't be opened for write *now*, fall back to the uinput X-Box 360 pad rather than a
+/// dead controller (the UHID device-create would just fail). Cheap — opens + drops the char device,
+/// no `UHID_CREATE2`, so no device is created. A no-op on non-Linux (those backends are UMDF/uinput).
+#[cfg(target_os = "linux")]
+fn degrade_if_no_uhid(chosen: GamepadPref) -> GamepadPref {
+    let needs_uhid = matches!(
+        chosen,
+        GamepadPref::DualSense | GamepadPref::DualShock4 | GamepadPref::SteamDeck
+    );
+    if needs_uhid
+        && std::fs::OpenOptions::new()
+            .write(true)
+            .open("/dev/uhid")
+            .is_err()
+    {
+        tracing::warn!(
+            wanted = chosen.as_str(),
+            "/dev/uhid not writable — falling back to the X-Box 360 pad"
+        );
+        return GamepadPref::Xbox360;
+    }
+    chosen
+}
+
+#[cfg(not(target_os = "linux"))]
+fn degrade_if_no_uhid(chosen: GamepadPref) -> GamepadPref {
+    chosen
+}
+
+/// True if a **physical** Valve Steam controller (`28DE`) is already attached. The host's own Steam
+/// Input is then managing a `28DE` device, and presenting a second (virtual) one makes Steam juggle
+/// two Decks — confirmed conflict-prone on a Deck-as-host (the physical `28DE:1205` + Steam's
+/// `28DE:11FF` XInput output pad are both live). HID device dirs are named `BUS:VID:PID.INST`
+/// (uppercase); a UHID virtual device resolves through `/devices/virtual/…`, a real one does not.
+#[cfg(target_os = "linux")]
+fn physical_steam_controller_present() -> bool {
+    let Ok(entries) = std::fs::read_dir("/sys/bus/hid/devices") else {
+        return false;
+    };
+    entries.flatten().any(|e| {
+        if !e.file_name().to_string_lossy().contains(":28DE:") {
+            return false;
+        }
+        match std::fs::read_link(e.path()) {
+            Ok(target) => !target.to_string_lossy().contains("/virtual/"),
+            Err(_) => true,
+        }
+    })
+}
+
+/// Gate a virtual Steam pad off when a physical Steam controller is attached (§ conflict). Degrade to
+/// DualSense (then the uhid ladder), which Steam treats as an ordinary, distinct pad. Override with
+/// `PUNKTFUNK_STEAM_FORCE=1` when the host has no competing Steam Input (e.g. a remote-only box).
+#[cfg(target_os = "linux")]
+fn degrade_steam_on_conflict(chosen: GamepadPref) -> GamepadPref {
+    if !matches!(
+        chosen,
+        GamepadPref::SteamDeck | GamepadPref::SteamController
+    ) {
+        return chosen;
+    }
+    let forced = std::env::var("PUNKTFUNK_STEAM_FORCE")
+        .map(|v| v == "1" || v.eq_ignore_ascii_case("true"))
+        .unwrap_or(false);
+    if !forced && physical_steam_controller_present() {
+        tracing::warn!(
+            wanted = chosen.as_str(),
+            "a physical Steam controller is attached — the host's Steam Input would manage two 28DE \
+             devices; falling back to DualSense (set PUNKTFUNK_STEAM_FORCE=1 to override)"
+        );
+        return degrade_if_no_uhid(GamepadPref::DualSense);
+    }
+    chosen
+}
+
+#[cfg(not(target_os = "linux"))]
+fn degrade_steam_on_conflict(chosen: GamepadPref) -> GamepadPref {
+    chosen
+}
+
 /// Resolve the client's gamepad-backend preference (the env/logging shell around
 /// [`pick_gamepad`]). Always concrete — the `Welcome` reports what the session will drive.
 fn resolve_gamepad(pref: GamepadPref) -> GamepadPref {
@@ -1891,6 +2012,14 @@ fn resolve_gamepad(pref: GamepadPref) -> GamepadPref {
         cfg!(target_os = "linux"),
         cfg!(target_os = "windows"),
     );
+    // Runtime degrade (separate from the compile-time platform check above): the Linux UHID
+    // backends need `/dev/uhid` usable *now*, else creating the device just fails and the controller
+    // goes dead — fall back to the always-available uinput X-Box 360 pad instead.
+    let chosen = degrade_if_no_uhid(chosen);
+    // Conflict gate: don't present a virtual Steam (28DE) pad when the host already has a physical
+    // Steam controller — its own Steam Input would then manage two Decks (confirmed conflict-prone on
+    // a Deck-as-host). `PUNKTFUNK_STEAM_FORCE=1` overrides.
+    let chosen = degrade_steam_on_conflict(chosen);
     match pref {
         GamepadPref::Auto => {
             // The operator's env knob deserves a diagnostic when it didn't drive the
@@ -2571,7 +2700,7 @@ fn session_watcher_loop(tx: std::sync::mpsc::Sender<SessionSwitch>, stop: Arc<At
     }
 }
 
-/// All per-session inputs for [`virtual_stream`] / [`virtual_stream_relay`], bundled so the session entry
+/// All per-session inputs for [`virtual_stream`], bundled so the session entry
 /// is one moved value instead of a 13-positional-argument `#[allow(too_many_arguments)]` signature
 /// (Goal-1 stage 4, plan §2.4). Everything is **owned** — the receivers move in (`virtual_stream` is their
 /// only consumer) — so the whole context moves into the stream thread and the borrow plumbing disappears.
@@ -2619,8 +2748,9 @@ struct SessionContext {
 }
 
 fn virtual_stream(ctx: SessionContext) -> Result<()> {
-    // This thread runs the capture+encode loop (single-process: Linux / synthetic / NO_WGC DDA) — or
+    // This thread runs the capture+encode loop (single-process — the only topology: Linux portal /
-    // tail-calls the relay below. Elevate it so a CPU-heavy game can't deschedule our GPU submission.
+    // synthetic, Windows in-process IDD-push). Elevate it so a CPU-heavy game can't deschedule our GPU
+    // submission.
     boost_thread_priority(true);
     // Resolve the per-session capture / topology / encoder decision ONCE (Goal-1 stage 3): the deployed
     // path now reads this typed `SessionPlan` instead of re-deriving from config at each dispatch site
@@ -2628,14 +2758,6 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
     // only per-session input — capture/topology/encoder are otherwise pure functions of `HostConfig`.
     let plan = crate::session_plan::SessionPlan::resolve(ctx.bit_depth, ctx.chroma);
     tracing::info!(?plan, "resolved session plan");
-    // Windows two-process secure-desktop path: when the host runs as SYSTEM (required for the secure
-    // desktop + SendInput), WGC can't activate in-process, so we capture the normal desktop via a
-    // helper spawned in the user session and relay its AUs. (Single-process WGC/DDA is used as the
-    // user, and stays the path on Linux.) See design/archive/windows-secure-desktop.md.
-    #[cfg(target_os = "windows")]
-    if plan.topology == crate::session_plan::SessionTopology::TwoProcessRelay {
-        return virtual_stream_relay(ctx);
-    }
     // Single-process path: unpack the context into the locals the loop below uses (names unchanged, so the
     // body is byte-for-byte the same; the receivers are now owned but `try_recv()` is identical).
     let SessionContext {
@@ -2670,6 +2792,11 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
     // host-lifetime VirtualDisplayManager (§2.5). It does NO monitor work, so it must precede the IDD-push
     // preempt below (which reaches the manager) — otherwise `vdm()` is called before init and panics.
     let mut vd = crate::vdisplay::open(compositor)?;
+    // Per-client STABLE monitor identity (Phase 2): hand the backend the connecting client's cert
+    // fingerprint so a freshly CREATED virtual monitor gets this client's persistent id — Windows then
+    // reapplies the client's saved per-monitor config (DPI scaling) on reconnect. No-op on Linux backends
+    // and for anonymous/GameStream clients (no fingerprint → the driver auto-allocates).
+    vd.set_client_identity(endpoint::peer_fingerprint(&conn));
     // IDD-push reconnect preempt (the dance now lives in the manager, Goal-1 §2.5): serialize setup so a
     // reconnect FLOOD can't run concurrent monitor create/teardown, STOP the prior session + WAIT for it
     // to release its monitor (instead of tearing a monitor out from under a still-live session), and
@@ -2685,20 +2812,7 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
     #[cfg(target_os = "windows")]
     drop(_idd_setup_guard);
 
-    // Windows single-process DDA path (PUNKTFUNK_NO_WGC=1): the SudoVDA virtual display, isolated as the
+    // Windows: capture is live — launch the requested library title into the
-    // SOLE active output, goes into fullscreen independent-flip (one plane on one display) which Desktop
-    // Duplication cannot capture → the born-lost ACCESS_LOST storm we measured on the RTX4090+iGPU box
-    // (hook verified-firing, DPI=2, yet 100% DuplicateOutput1 E_ACCESSDENIED + born-lost). A tiny topmost
-    // layered overlay disqualifies independent-flip and forces DWM composition, which DDA CAN capture.
-    // (Apollo never hits this because it runs WITH a physical monitor attached — multi-display is already
-    // DWM-composited; we isolate to sole-display, so we must force composition ourselves.) Unlike the WGC
-    // relay path — where WGC owns the normal desktop and the overlay is secure-only — here DDA owns the
-    // normal desktop too, so it must run unconditionally. Held for the session; Drop tears it down.
-    // Best-effort; disable with PUNKTFUNK_FORCE_COMPOSED=0.
-    #[cfg(target_os = "windows")]
-    let _composed_flip = crate::capture::composed_flip::ForceComposedFlip::start();
-
-    // Windows: capture is live (and composition forced) — launch the requested library title into the
     // interactive user session so it renders onto the captured desktop and grabs foreground. Linux
     // nests its launch in gamescope instead (the handshake `PUNKTFUNK_GAMESCOPE_APP` path). Best-effort:
     // a launch failure (no recipe for the kind, no interactive user) leaves the user on the desktop.
@@ -3170,480 +3284,6 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
     Ok(())
 }
 
-/// Windows two-process video stream: the SYSTEM host creates the SudoVDA virtual output (and holds
-/// its keepalive = the sole topology/isolation owner), spawns the WGC helper in the user session to
-/// capture+encode the NORMAL desktop, and relays the helper's AUs onto the QUIC data plane via the
-/// same send thread as the single-process path. A [`DesktopWatcher`](crate::capture::desktop_watch)
-/// muxes the source: while the input desktop is Winlogon (UAC / lock / login — which WGC can't
-/// capture), the host captures it with its OWN DDA encoder; back on Default it resumes the relay.
-/// Every source switch latches a "wait for IDR" so the client's decoder resumes on a keyframe (the
-/// two encoders keep independent infinite-GOP state). Reconfigure rebuilds the output + re-spawns the
-/// helper at the new mode (and drops the stale-target DDA); keyframe requests forward to the active
-/// source.
-#[cfg(target_os = "windows")]
-fn virtual_stream_relay(ctx: SessionContext) -> Result<()> {
-    use crate::capture::dxgi::WinCaptureTarget;
-    use crate::capture::wgc_relay::HelperRelay;
-    use crate::capture::Capturer; // trait methods (set_active/next_frame) on the concrete DuplCapturer
-
-    // Unpack the context (names unchanged so the body is identical). The relay doesn't yet send the
-    // source's 0xCE HDR metadata — the helper's in-band SEI carries it (a Windows follow-up) — so `conn`
-    // is held unused.
-    let SessionContext {
-        session,
-        mode,
-        seconds,
-        stop,
-        reconfig,
-        keyframe,
-        compositor,
-        bitrate_kbps,
-        bit_depth,
-        // The two-process WGC relay encodes 4:2:0 in v1 — the handshake's `single_process` gate already
-        // forced `chroma` to Yuv420 for this topology, so the helper + secure-desktop DDA stay 4:2:0.
-        chroma: _,
-        probe_rx,
-        probe_result_tx,
-        fec_target,
-        conn: _conn,
-        stats,
-        client_label,
-        launch,
-    } = ctx;
-    tracing::info!(
-        ?mode,
-        bitrate_kbps,
-        bit_depth,
-        "punktfunk/1 two-process stream (SYSTEM host + user-session WGC helper)"
-    );
-
-    let mut vd = crate::vdisplay::open(compositor)?;
-
-    // Create the SudoVDA output + spawn a helper capturing it by GDI name. Returns the keepalive
-    // (held for the output's life — the sole isolation owner), the running relay, the capture target
-    // (so the host can also open DDA on it for the secure desktop), and the achieved refresh.
-    type Built = (Box<dyn Send>, HelperRelay, WinCaptureTarget, u32);
-    let build = |vd: &mut Box<dyn crate::vdisplay::VirtualDisplay>,
-                 mode: punktfunk_core::Mode|
-     -> Result<Built> {
-        let vout = vd.create(mode).context("create virtual output")?;
-        let effective_hz = vout
-            .preferred_mode
-            .map(|(_, _, hz)| hz)
-            .filter(|&hz| hz > 0)
-            .unwrap_or(mode.refresh_hz);
-        let target = vout.win_capture.clone().ok_or_else(|| {
-            anyhow!("SudoVDA target not yet an active display (needs a WDDM GPU to activate it)")
-        })?;
-        // HDR is driven by the SudoVDA monitor's ACTUAL advanced-color state, not the handshake bit
-        // depth: the whole pipeline follows the monitor (WGC captures FP16 when HDR is on; NVENC forces
-        // Main10 + BT.2020 PQ from the 10-bit capture format regardless of the negotiated depth; the
-        // client auto-detects PQ from the HEVC VUI). So:
-        //   - a negotiated 10-bit session PROACTIVELY enables HDR on the monitor (below), but
-        //   - we must NEVER force HDR *off* here — that would wipe out a user's deliberate Windows HDR
-        //     toggle on the virtual display on every build (the "HDR doesn't persist" bug). Leaving the
-        //     monitor's state alone lets a user-enabled HDR session flow through end-to-end.
-        // The secure-desktop HDR drop (for the DDA leg) keys off the monitor's real state in the mux loop.
-        #[cfg(target_os = "windows")]
-        if bit_depth >= 10 {
-            // SAFETY: `set_advanced_color` is marked `unsafe` only because it drives the Win32 CCD API
-            // internally; it takes `target_id` by value (Copy `u32` — this session's live SudoVDA
-            // monitor's CCD target id) and sizes + owns every buffer it hands the OS on its own stack.
-            // We pass no pointers, so nothing must outlive the call and there is no aliasing; an
-            // unknown/absent target id simply returns false.
-            unsafe {
-                if crate::win_display::set_advanced_color(target.target_id, true) {
-                    // Let the colorspace change settle before WGC creates its capture item / detects HDR.
-                    std::thread::sleep(std::time::Duration::from_millis(250));
-                }
-            }
-        }
-        let relay = HelperRelay::spawn(
-            &target,
-            (mode.width, mode.height, effective_hz),
-            bitrate_kbps,
-            bit_depth,
-        )
-        .context("spawn WGC helper")?;
-        Ok((vout.keepalive, relay, target, effective_hz))
-    };
-
-    let (mut _keepalive, mut relay, mut target, mut effective_hz) = build(&mut vd, mode)?;
-    let mut cur_mode = mode;
-
-    // Capture is live (the WGC helper is relaying) — launch the requested library title into the
-    // interactive user session so it renders onto the captured desktop and grabs foreground.
-    // Best-effort: a failure (no recipe for the kind, no interactive user) leaves the user on the desktop.
-    if let Some(id) = launch.as_deref() {
-        if let Err(e) = crate::library::launch_title(id) {
-            tracing::warn!(launch_id = id, error = %e, "could not launch requested library title");
-        }
-    }
-
-    // O3.1: optionally observe the IDD-push ring alongside WGC (WGC = the presentation trigger) to
-    // confirm the 0257 driver pushes frames into a HOST-created ring. Diagnostic only; gated.
-    if std::env::var_os("PUNKTFUNK_IDD_PUSH_OBSERVE").is_some() {
-        crate::capture::idd_push::spawn_observer(
-            target.clone(),
-            Some((cur_mode.width, cur_mode.height, effective_hz)),
-        );
-    }
-
-    // The host's own DDA capturer+encoder for the SECURE (Winlogon) desktop, which WGC — and thus the
-    // helper — cannot capture. Opened lazily on the first secure transition (so a session that never
-    // hits a UAC/lock screen never pays for a second NVENC session), then kept for fast re-switch.
-    struct DdaPipe {
-        cap: Box<dyn crate::capture::Capturer>,
-        enc: Box<dyn crate::encode::Encoder>,
-        frame: crate::capture::CapturedFrame,
-    }
-    // Note: takes the dimensions as args rather than capturing `cur_mode` — `cur_mode` is reassigned
-    // on reconfig, and a closure holding a shared borrow of it for the whole fn would forbid that.
-    let open_dda =
-        |target: &WinCaptureTarget, w: u32, h: u32, hz: u32, hdr: bool| -> Result<DdaPipe> {
-            // The host already holds the real keepalive (sole isolation owner), so DDA gets a no-op one.
-            // `hdr` requests an FP16 DuplicateOutput1 so the secure desktop is captured in HDR (→ BT.2020
-            // PQ Main10) instead of black — legacy DuplicateOutput can't capture an HDR/FP16 desktop.
-            let mut cap = crate::capture::dxgi::DuplCapturer::open(
-                target.clone(),
-                Some((w, h, hz)),
-                Box::new(()),
-                // The relay's host encoder is GPU (NVENC/AMF/QSV unless software) — pass `gpu` in (Goal-1
-                // stage 5) so the DDA capturer doesn't re-derive it.
-                crate::capture::gpu_encode(),
-                hdr,
-                false, // the two-process relay path is 4:2:0 in v1
-            )
-            .context("open DDA for secure desktop")?;
-            cap.set_active(true);
-            let frame = cap.next_frame().context("DDA first frame")?;
-            let enc = crate::encode::open_video(
-                crate::encode::Codec::H265,
-                frame.format,
-                frame.width,
-                frame.height,
-                hz,
-                bitrate_kbps as u64 * 1000,
-                frame.is_cuda(),
-                bit_depth,
-                // Secure-desktop DDA on the two-process relay path: 4:2:0 in v1 (matches the helper).
-                crate::encode::ChromaFormat::Yuv420,
-            )
-            .context("open video encoder for DDA")?;
-            Ok(DdaPipe {
-                cap: Box::new(cap),
-                enc,
-                frame,
-            })
-        };
-
-    let perf = crate::config::config().perf;
-    let burst_cap = std::env::var("PUNKTFUNK_PACE_BURST_KB")
-        .ok()
-        .and_then(|s| s.parse::<usize>().ok())
-        .unwrap_or(128)
-        * 1024;
-
-    // Same encode|send split as the single-process path: this thread relays AUs, a dedicated send
-    // thread owns the Session and does FEC+seal+paced-send. The relay encodes in the helper process,
-    // so this path's FrameMsgs carry no cap/submit/encode split (those stages stay 0 in the sample);
-    // the send thread still emits fps/goodput/pacing/loss from `session.stats()`.
-    let send_stats = SendStats {
-        rec: stats,
-        width: mode.width,
-        height: mode.height,
-        fps: effective_hz,
-        codec: "hevc",
-        client: client_label,
-        bitrate_kbps,
-    };
-    let (frame_tx, frame_rx) = std::sync::mpsc::sync_channel::<FrameMsg>(3);
-    let send_thread = std::thread::Builder::new()
-        .name("punktfunk-send".into())
-        .spawn({
-            let stop = stop.clone();
-            move || {
-                send_loop(
-                    session,
-                    frame_rx,
-                    probe_rx,
-                    probe_result_tx,
-                    stop,
-                    perf,
-                    burst_cap,
-                    fec_target,
-                    send_stats,
-                )
-            }
-        })
-        .context("spawn send thread")?;
-
-    // Test hook: PUNKTFUNK_SECURE_TEST_PERIOD_MS=N drives a square-wave secure/normal toggle every N ms
-    // instead of the real watcher — exercises the mid-session helper↔DDA mux without a live UAC/lock.
-    let secure_test_ms: Option<u128> = std::env::var("PUNKTFUNK_SECURE_TEST_PERIOD_MS")
-        .ok()
-        .and_then(|s| s.parse().ok())
-        .filter(|&n| n > 0);
-    // Switching to the host DDA on the secure (Winlogon) desktop is OPT-IN: DDA can't reliably capture
-    // the secure desktop's HDR independent-flip (it storms ACCESS_LOST → black), whereas the WGC helper
-    // STAYS LIVE through a lock/UAC. So by default the mux keeps WGC the whole time (no DesktopWatcher
-    // switch, no overlay). Enable the experimental DDA-on-secure path with PUNKTFUNK_SECURE_DDA=1.
-    let dda_secure = crate::config::config().secure_dda || secure_test_ms.is_some();
-    // The authoritative Default↔Winlogon signal (requires SYSTEM to read the Winlogon desktop name);
-    // only needed when the DDA-on-secure path is enabled.
-    let watcher = dda_secure.then(crate::capture::desktop_watch::DesktopWatcher::start);
-    // Force-composed-flip overlay (only with DDA-on-secure): keeps the secure desktop out of fullscreen
-    // independent-flip so DDA can duplicate it. Off by default to avoid touching the normal desktop.
-    let _composed_flip = dda_secure
-        .then(crate::capture::composed_flip::ForceComposedFlip::start)
-        .flatten();
-    let start = std::time::Instant::now();
-
-    let mut interval = std::time::Duration::from_secs_f64(1.0 / effective_hz.max(1) as f64);
-    let deadline = std::time::Instant::now() + std::time::Duration::from_secs(seconds as u64);
-    let mut sent: u64 = 0;
-    // Mux state: which source is live, the lazily-opened DDA pipe, a DDA pacing clock, and a
-    // "wait for the next IDR before forwarding" latch set on every source switch (the client's
-    // decoder must resume on a keyframe — the two encoders keep independent infinite-GOP state).
-    let mut dda: Option<DdaPipe> = None;
-    let mut on_secure = false;
-    let mut next = std::time::Instant::now();
-    let mut await_idr = false;
-    // Step 6 relaunch watchdog: how many times in a row the helper has died without producing a frame.
-    // A console disconnect/reconnect or a helper crash kills it; we respawn (the new helper picks up
-    // the now-active session via WTSGetActiveConsoleSessionId). Reset on the first relayed frame; only
-    // give up (end the stream) after a run of failures spanning a few seconds.
-    let mut helper_fails = 0u32;
-    const MAX_HELPER_FAILS: u32 = 20;
-
-    // Build a FrameMsg + hand it to the send thread; returns false if the send thread is gone (caller
-    // breaks the loop). Kept as a macro (not a closure) so each use borrows `frame_tx`/`sent`/`interval`
-    // at its own site without a long-lived capture, and `break 'outer` stays a literal at the call site
-    // (a `break 'outer` inside the macro body risks label-hygiene resolution failures).
-    macro_rules! forward {
-        ($data:expr, $capture_ns:expr, $keyframe:expr) => {{
-            let flags = if $keyframe {
-                (FLAG_PIC | FLAG_SOF) as u32
-            } else {
-                FLAG_PIC as u32
-            };
-            let capture_ns = $capture_ns;
-            let encode_us = (now_ns().saturating_sub(capture_ns) / 1000) as u32;
-            let msg = FrameMsg {
-                data: $data,
-                capture_ns,
-                flags,
-                deadline: std::time::Instant::now() + interval,
-                encode_us,
-                cap_us: 0,
-                submit_us: 0,
-                wait_us: 0,
-                repeat: false,
-                was_measured: false,
-            };
-            let ok = frame_tx.send(msg).is_ok();
-            if ok {
-                sent += 1;
-            }
-            ok
-        }};
-    }
-
-    'outer: while !stop.load(Ordering::SeqCst) && std::time::Instant::now() < deadline {
-        // Mode switch: rebuild the output + re-spawn the helper at the new mode (drop the old relay +
-        // keepalive only after the new pair is up, so a failed rebuild keeps the current stream). The
-        // DDA pipe (on the old target) is dropped — it reopens on the next secure transition.
-        let mut want = None;
-        while let Ok(m) = reconfig.try_recv() {
-            want = Some(m);
-        }
-        if let Some(new_mode) = want {
-            tracing::info!(?new_mode, "two-process: rebuilding for mode switch");
-            match build(&mut vd, new_mode) {
-                Ok((ka, rl, tg, hz)) => {
-                    relay = rl; // drops the old relay (kills old helper) ...
-                    _keepalive = ka; // ... then releases the old output
-                    target = tg;
-                    effective_hz = hz;
-                    cur_mode = new_mode;
-                    dda = None; // old-target DDA is stale; reopen on next secure
-                    interval = std::time::Duration::from_secs_f64(1.0 / hz.max(1) as f64);
-                }
-                Err(e) => {
-                    tracing::error!(error = %format!("{e:#}"), ?new_mode,
-                        "two-process mode-switch rebuild failed — staying on the current mode");
-                }
-            }
-        }
-        // Coalesce client decode-recovery keyframe requests and forward to the active source.
-        let mut want_kf = false;
-        while keyframe.try_recv().is_ok() {
-            want_kf = true;
-        }
-
-        // Source mux: capture the secure (Winlogon) desktop via the host's DDA, the normal desktop via
-        // the helper relay. On a switch, latch await_idr + force the now-active source to emit an IDR
-        // so the client resumes cleanly.
-        let secure = dda_secure
-            && match secure_test_ms {
-                Some(p) => (start.elapsed().as_millis() / p) % 2 == 1,
-                None => watcher.as_ref().is_some_and(|w| w.is_secure()),
-            };
-        if secure != on_secure {
-            on_secure = secure;
-            await_idr = true;
-            tracing::info!(
-                to = if secure {
-                    "secure(DDA)"
-                } else {
-                    "normal(WGC relay)"
-                },
-                "two-process: source switch"
-            );
-            if secure {
-                // Capture the secure (Winlogon) desktop in its NATIVE colorspace. Don't try to drop the
-                // SudoVDA out of HDR for the DDA leg — display-config changes are denied on the secure
-                // desktop (the drop just churned + still went black). Instead, if the monitor is in HDR,
-                // open DDA in HDR (FP16 DuplicateOutput1 → BT.2020 PQ Main10); the normal-desktop DDA
-                // overlay/flip issues that drove us to WGC don't apply to the composed Winlogon UI.
-                // SAFETY: `advanced_color_enabled` is `unsafe` only because it queries the Win32 CCD
-                // API; it takes `target_id` by value (the live SudoVDA monitor's CCD target id) and
-                // allocates + owns every buffer it passes the OS internally. No caller pointer is
-                // involved, so nothing must outlive the call and there is no aliasing; a missing
-                // target id just yields false.
-                let hdr = unsafe { crate::win_display::advanced_color_enabled(target.target_id) };
-                dda = None; // reopen to capture the secure desktop
-                match open_dda(&target, cur_mode.width, cur_mode.height, effective_hz, hdr) {
-                    Ok(mut p) => {
-                        tracing::info!(hdr, "two-process: opened DDA for the secure desktop");
-                        p.enc.request_keyframe();
-                        dda = Some(p);
-                    }
-                    Err(e) => {
-                        tracing::error!(error = %format!("{e:#}"),
-                            "two-process: DDA open failed — secure desktop will freeze on last frame");
-                    }
-                }
-                next = std::time::Instant::now();
-            } else {
-                // Returning to the normal desktop: RESUME from the still-alive WGC helper. Do NOT
-                // recreate the SudoVDA monitor or respawn the helper — build()'s vd.create() is an
-                // IOCTL_REMOVE+ADD of the monitor (the audible disconnect/connect chime + the
-                // teardown/recreate kernel stress that broke DDA, now applied to the mux). The monitor +
-                // helper persist for the WHOLE session; only the host-DDA leg opens (secure) and closes
-                // (normal). Apply the DDA learning here: reuse, don't tear down.
-                dda = None; // free the secure DDA encoder; the relay (helper) is the source again
-                while relay.try_recv().is_ok() {} // drop secure-dwell backlog
-                relay.request_keyframe(); // client decoder resumes on the helper's next IDR
-                                          // Nothing to restore: we no longer toggle the SudoVDA's HDR state for the DDA leg, so the
-                                          // monitor's colorspace is unchanged and the still-alive WGC helper just resumes.
-                next = std::time::Instant::now();
-            }
-        }
-        if want_kf {
-            if secure {
-                if let Some(d) = dda.as_mut() {
-                    d.enc.request_keyframe();
-                }
-            } else {
-                relay.request_keyframe();
-            }
-            await_idr = true;
-        }
-
-        if secure {
-            // DDA capture+encode for the secure desktop, paced to the frame interval.
-            let Some(d) = dda.as_mut() else {
-                std::thread::sleep(interval);
-                continue;
-            };
-            if let Some(f) = d.cap.try_latest().context("DDA capture")? {
-                d.frame = f;
-            }
-            let capture_ns = now_ns();
-            d.enc.submit(&d.frame).context("DDA encoder submit")?;
-            next += interval;
-            while let Some(au) = d.enc.poll().context("DDA encoder poll")? {
-                if await_idr && !au.keyframe {
-                    continue;
-                }
-                await_idr = false;
-                if !forward!(au.data, capture_ns, au.keyframe) {
-                    break 'outer; // send thread gone
-                }
-            }
-            match next.checked_duration_since(std::time::Instant::now()) {
-                Some(dur) => std::thread::sleep(dur),
-                None => next = std::time::Instant::now(),
-            }
-        } else {
-            // Relay the helper's AUs for the normal desktop. Timeout → keep servicing the loop;
-            // Disconnected → the helper exited (step 6 adds the relaunch watchdog).
-            let au = match relay.recv_timeout(std::time::Duration::from_millis(500)) {
-                Ok(au) => au,
-                Err(std::sync::mpsc::RecvTimeoutError::Timeout) => {
-                    if stop.load(Ordering::SeqCst) {
-                        break;
-                    }
-                    tracing::warn!("two-process: no AU from helper within 500ms");
-                    continue;
-                }
-                Err(std::sync::mpsc::RecvTimeoutError::Disconnected) => {
-                    // The helper exited (crash, or a console disconnect killed its session). REBUILD
-                    // the whole output + helper (not just respawn on the old target): an abruptly-killed
-                    // helper leaves the SudoVDA's DXGI output briefly unresolvable ("no DXGI output for
-                    // target N yet"), and a console reconnect needs a fresh output in the new session —
-                    // `build` recreates both. Back off so a hard-failing rebuild (e.g. no active session
-                    // yet) doesn't spin; give up only after a sustained run of failures.
-                    helper_fails += 1;
-                    if helper_fails > MAX_HELPER_FAILS {
-                        tracing::error!(
-                            fails = helper_fails,
-                            "two-process: WGC helper keeps dying — ending stream"
-                        );
-                        break;
-                    }
-                    std::thread::sleep(std::time::Duration::from_millis(500));
-                    match build(&mut vd, cur_mode) {
-                        Ok((ka, rl, tg, hz)) => {
-                            tracing::warn!(
-                                fails = helper_fails,
-                                "two-process: WGC helper exited — rebuilt output + helper"
-                            );
-                            relay = rl;
-                            _keepalive = ka;
-                            target = tg;
-                            effective_hz = hz;
-                            dda = None; // old-target DDA is stale
-                            interval = std::time::Duration::from_secs_f64(1.0 / hz.max(1) as f64);
-                            await_idr = true; // resume on the new helper's opening IDR
-                        }
-                        Err(e) => {
-                            tracing::warn!(error = %format!("{e:#}"), fails = helper_fails,
-                                "two-process: helper rebuild failed — will retry");
-                        }
-                    }
-                    continue;
-                }
-            };
-            if await_idr && !au.keyframe {
-                continue; // skip stale deltas until the post-switch IDR
-            }
-            await_idr = false;
-            helper_fails = 0; // a frame flowed → the helper is healthy again
-                              // The helper's pts_ns is on this machine's monotonic clock (same `now_ns()` source).
-            if !forward!(au.data, au.pts_ns, au.keyframe) {
-                break 'outer; // send thread gone
-            }
-        }
-    }
-    drop(frame_tx);
-    let _ = send_thread.join();
-    drop(watcher);
-    tracing::info!(sent, "punktfunk/1 two-process stream complete");
-    Ok(())
-}
-
 /// One mode's capture/encode pipeline: (capturer, encoder, first frame, frame interval).
 /// Dropping the capturer tears down the PipeWire stream and the virtual output with it.
 type Pipeline = (
@@ -3675,6 +3315,23 @@ fn build_pipeline_with_retry(
     // 30-60s to produce its first frame, and a first-connect timeout would tear down the warm
     // session (forcing another cold start on reconnect). A genuinely permanent failure still fails
     // fast via `is_permanent_build_error`; only transient "no frame yet" retries consume the budget.
+    // IDD-push only: HOLD one monitor lease across all build attempts. A failed attempt's capturer
+    // drop releases ITS lease, but this held lease keeps the shared monitor Active (refs >= 1), so the
+    // next attempt's `vd.create` JOINS it (refcount++) instead of finding it Lingering and tripping the
+    // IDD-push reconnect PREEMPT (teardown + recreate). That preempt-per-retry was the REMOVE→ADD churn
+    // that exhausts the IddCx monitor-slot pool and wedges ADD at 0x80070490 — one ADD per cold start
+    // now, not one per attempt. Non-IDD-push backends (Linux portal, WGC) don't use the refcount manager
+    // and aren't churn-wedge-prone, so they keep create-per-attempt (a held lease there would allocate a
+    // second virtual output). Dropped when this fn returns — on success the Pipeline's own lease keeps
+    // the monitor Active; on failure refs falls to 0 → Lingering → linger-timeout teardown.
+    let _retry_hold = if matches!(plan.capture, crate::session_plan::CaptureBackend::IddPush) {
+        Some(
+            vd.create(mode)
+                .context("acquire virtual output for the session (retry-hold lease)")?,
+        )
+    } else {
+        None
+    };
     const MAX_ATTEMPTS: u32 = 8;
     let mut backoff = std::time::Duration::from_millis(500);
     for attempt in 1..=MAX_ATTEMPTS {

crates/punktfunk-host/src/session_plan.rs

@@ -26,12 +26,9 @@ pub enum CaptureBackend {
     /// Linux: the xdg ScreenCast portal → PipeWire (the only Linux capture path).
     Portal,
     /// Windows: IDD direct-push — frames pulled straight from the pf-vdisplay driver's shared ring
-    /// (in-process, Session 0; no Desktop Duplication, no WGC helper).
+    /// (in-process, Session 0; captures the secure desktop too). The sole Windows capture path —
+    /// DXGI Desktop Duplication (DDA) and the WGC two-process relay were removed.
     IddPush,
-    /// Windows: DXGI Desktop Duplication (`PUNKTFUNK_CAPTURE=dda|dxgi` or `PUNKTFUNK_NO_WGC`).
-    Dda,
-    /// Windows: Windows.Graphics.Capture (the composed-desktop default), with a DDA watchdog fallback.
-    Wgc,
 }
 
 impl CaptureBackend {
@@ -42,20 +39,10 @@ impl CaptureBackend {
         CaptureBackend::Portal
     }
 
-    /// Windows precedence (identical to the pre-stage-3 `capture_virtual_output` branch order):
+    /// Windows: IDD direct-push is the sole capture path (DDA + the WGC two-process relay were removed).
-    /// IDD-push wins; else an explicit `dda`/`dxgi` request or `PUNKTFUNK_NO_WGC` selects DDA; else WGC.
     #[cfg(target_os = "windows")]
     pub fn resolve() -> Self {
-        let cfg = crate::config::config();
-        if cfg.idd_push {
         CaptureBackend::IddPush
-        } else if matches!(cfg.capture_backend.as_str(), "dda" | "dxgi")
-            || crate::capture::wgc_disabled()
-        {
-            CaptureBackend::Dda
-        } else {
-            CaptureBackend::Wgc
-        }
     }
 
     #[cfg(not(any(target_os = "linux", target_os = "windows")))]
@@ -67,11 +54,9 @@ impl CaptureBackend {
 /// How a session is structured across processes.
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 pub enum SessionTopology {
-    /// One process captures + encodes (Linux; Windows non-SYSTEM / IDD-push / `NO_WGC`).
+    /// One process captures + encodes. The only topology: Linux (portal) and Windows (in-process
+    /// IDD-push in Session 0). The SYSTEM-host + user-session WGC relay was removed with DDA/WGC.
     SingleProcess,
-    /// SYSTEM host + a user-session WGC helper relay (the Windows normal-desktop path under SYSTEM,
-    /// where in-process WGC can't activate). See `virtual_stream_relay`.
-    TwoProcessRelay,
 }
 
 /// The resolved encode backend (recorded for logging / stages 4–5; the per-session encoder open still
@@ -103,8 +88,8 @@ pub struct SessionPlan {
     pub encoder: EncoderBackend,
     /// Handshake-negotiated encode bit depth (8, or 10 = HEVC Main10).
     pub bit_depth: u8,
-    /// The IDD-push HDR hint (`bit_depth >= 10`) — the want-HDR flag the capturer was passed before.
+    /// The IDD-push HDR hint (`bit_depth >= 10`) — the want-HDR flag handed to the capturer so it
-    /// Non-IDD-push Windows backends ignore it and auto-detect HDR from the monitor; Linux is 8-bit.
+    /// proactively enables advanced color on the virtual display. Linux is 8-bit (HDR blocked upstream).
     pub hdr: bool,
     /// Handshake-negotiated chroma subsampling (4:2:0, or full-chroma 4:4:4 when the client + host +
     /// GPU all support it). Resolved before the Welcome; `Yuv420` on every backend that declined it.
@@ -151,26 +136,8 @@ impl SessionPlan {
     }
 }
 
-/// Process topology. On Windows this is the former `punktfunk1::should_use_helper` logic verbatim; on
+/// Process topology. Single-process is the only topology now: Linux (portal) and Windows (in-process
-/// every other platform the session is always single-process.
+/// IDD-push in Session 0). The Windows SYSTEM-host + user-session WGC relay was removed with DDA/WGC.
-#[cfg(target_os = "windows")]
-pub(crate) fn resolve_topology() -> SessionTopology {
-    let cfg = crate::config::config();
-    // `NO_HELPER`/`NO_WGC` force single-process; IDD-push captures in-process in Session 0 (no helper);
-    // otherwise the helper runs when forced or when we're SYSTEM (in-process WGC can't activate there).
-    let helper = if cfg.no_helper || crate::capture::wgc_disabled() || cfg.idd_push {
-        false
-    } else {
-        cfg.force_helper || crate::capture::wgc_relay::running_as_system()
-    };
-    if helper {
-        SessionTopology::TwoProcessRelay
-    } else {
-        SessionTopology::SingleProcess
-    }
-}
-
-#[cfg(not(target_os = "windows"))]
 pub(crate) fn resolve_topology() -> SessionTopology {
     SessionTopology::SingleProcess
 }