fix(rpm): build with the baked toolchain (RUSTUP_TOOLCHAIN), avoid EXDEV

rust-toolchain.toml floats `channel = "stable"` + requests rustfmt/clippy. When a
newer stable lands upstream, that makes rustup try to update the baked, minimal-
profile `stable` toolchain in place during %build, and the builder image's
OverlayFS rejects the staging rename with EXDEV ("Invalid cross-device link"),
failing the RPM build (started the day Rust 1.96.1 shipped). A release build needs
no rustfmt/clippy, so pin RUSTUP_TOOLCHAIN=stable to use the installed toolchain
as-is — no channel re-resolve, no component add, no update. Scoped to the RPM
%build; ci.yml/deb.yml (rust-ci image) are unaffected.

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).
-    # There is NO fixed rsa release (the constant-time rewrite is still unreleased upstream), and rsa
-    # is required for GameStream/Moonlight pairing. Crucially, the host uses rsa ONLY for PKCS#1 v1.5
-    # SIGNING / VERIFYING (gamestream/cert.rs + gamestream/pairing.rs: SigningKey / VerifyingKey /
-    # Signer / Verifier) — it never performs RSA decryption, which is the operation Marvin targets.
-    # So the vulnerable code path is not exercised. Revisit if a fixed rsa ships or we add RSA decrypt.
+    # rsa "Marvin Attack" (RUSTSEC-2023-0071): a timing side-channel in the rsa crate's variable-time
+    # modular exponentiation of the SECRET exponent. IMPORTANT — this affects the RSA private-key op in
+    # general, INCLUDING signing (m^d mod n), which the host DOES perform (gamestream/pairing.rs
+    # `signing_key.sign(&serversecret)`). It is NOT, as an earlier version of this note wrongly claimed,
+    # limited to decryption — so "the vulnerable path isn't exercised" is false; signing exercises it.
+    # 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/android.yml

@@ -80,7 +80,7 @@ jobs:
         run: |
           case "$GITHUB_REF" in
             refs/tags/v*) VN="${GITHUB_REF_NAME#v}"; TRACK="alpha" ;;   # alpha = built-in closed testing
-            *)            VN="0.3.0-ci${GITHUB_RUN_NUMBER}"; TRACK="internal" ;;
+            *)            VN="0.5.0-ci${GITHUB_RUN_NUMBER}"; TRACK="internal" ;;
           esac
           echo "VERSION_NAME=$VN" >> "$GITHUB_ENV"
           echo "PLAY_TRACK=$TRACK" >> "$GITHUB_ENV"

.gitea/workflows/deb.yml

@@ -36,8 +36,8 @@ jobs:
 
       - name: Version + channel
         # vX.Y.Z tag -> X.Y.Z, published to the `stable` apt distribution (a real release).
-        # A main push -> 0.3.0~ciN.g<sha>, published to the `canary` distribution: the '~' sorts
-        # below the eventual 0.3.0 tag, it climbs monotonically by run number, and the canary base
+        # A main push -> 0.5.0~ciN.g<sha>, published to the `canary` distribution: the '~' sorts
+        # below the eventual 0.5.0 tag, it climbs monotonically by run number, and the canary base
         # stays one minor AHEAD of the latest stable so a stable->canary box re-point still moves
         # forward (see channels.md). Computed BEFORE the build so it's stamped into the binary
         # (PUNKTFUNK_BUILD_VERSION -> build.rs -> --version).
@@ -45,7 +45,7 @@ jobs:
           SHORT=$(echo "$GITHUB_SHA" | cut -c1-8)
           case "$GITHUB_REF" in
             refs/tags/v*) V="${GITHUB_REF_NAME#v}"; DIST=stable ;;
-            *)            V="0.3.0~ci${GITHUB_RUN_NUMBER}.g${SHORT}"; DIST=canary ;;
+            *)            V="0.5.0~ci${GITHUB_RUN_NUMBER}.g${SHORT}"; DIST=canary ;;
           esac
           echo "VERSION=$V" >> "$GITHUB_ENV"
           echo "DISTRIBUTION=$DIST" >> "$GITHUB_ENV"
@@ -87,12 +87,13 @@ jobs:
           git config --global --add safe.directory "$PWD"
           cargo build --release -p punktfunk-host -p punktfunk-client-linux --locked
 
-      - name: Build + smoke-boot web console (node-server preset)
-        # Gate the .deb on a real node boot: the punktfunk-web .deb runs `node .output/server`,
-        # so prove the node-server build exists, isn't a bun bundle, and actually serves /login.
+      - name: Build + smoke-boot web console (bun preset)
+        # Gate the .deb on a real bun boot: the punktfunk-web .deb runs the Nitro `bun` preset
+        # (our Bun.serve TLS entry), so prove the build IS a bun bundle and serves /login.
+        # No TLS env here, so the custom entry binds plain HTTP — the smoke curl stays simple.
         run: |
-          # bun builds the console. It's baked into the rust-ci image, but bootstrap it here too so
-          # the job stays green against the PREVIOUS image (docker.yml bootstrap lag).
+          # bun builds AND runs the console. Baked into the rust-ci image; bootstrap here too so the
+          # job stays green against the PREVIOUS image (docker.yml bootstrap lag).
           command -v bun >/dev/null || {
             apt-get install -y --no-install-recommends unzip
             curl -fsSL https://bun.sh/install | bash
@@ -101,21 +102,23 @@ jobs:
           cd web
           bun install --frozen-lockfile
           bun run build
-          if grep -q 'Bun\.serve' .output/server/index.mjs; then
-            echo "ERROR: web build is a bun bundle (Bun.serve) — need the node-server preset"; exit 1
+          if ! grep -q 'Bun\.serve' .output/server/index.mjs; then
+            echo "ERROR: web build is not a bun bundle — need the 'bun' preset + custom entry"; exit 1
           fi
-          PORT=3009 HOST=127.0.0.1 PUNKTFUNK_UI_PASSWORD=ci node .output/server/index.mjs &
+          PORT=3009 HOST=127.0.0.1 PUNKTFUNK_UI_PASSWORD=ci bun .output/server/index.mjs &
           NP=$!; sleep 3
           code=$(curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:3009/login || echo 000)
           kill "$NP" 2>/dev/null || true
           echo "web console smoke: /login -> $code"
-          [ "$code" = 200 ] || { echo "ERROR: web console failed to boot under node"; exit 1; }
+          [ "$code" = 200 ] || { echo "ERROR: web console failed to boot under bun"; exit 1; }
 
       - name: Build .debs
         run: |
+          export PATH="$HOME/.bun/bin:$PATH"
           VERSION="$VERSION" bash packaging/debian/build-deb.sh
           VERSION="$VERSION" bash packaging/debian/build-client-deb.sh
-          VERSION="$VERSION" bash packaging/debian/build-web-deb.sh
+          # Reuse CI's bun for the vendored runtime (matches the amd64 runner) instead of downloading.
+          VERSION="$VERSION" BUN_BIN="$(command -v bun || true)" bash packaging/debian/build-web-deb.sh
 
       - name: Publish to the Gitea apt registry
         env:

.gitea/workflows/flatpak.yml

@@ -73,7 +73,7 @@ jobs:
 
       - name: Version + channel
         # Tag vX.Y.Z -> X.Y.Z on the OSTree `stable` branch (a real release); a main push ->
-        # 0.3.0-ciN.g<sha> on the `canary` branch. The two branches live side-by-side in one repo
+        # 0.5.0-ciN.g<sha> on the `canary` branch. The two branches live side-by-side in one repo
         # (rsync runs without --delete), each tracked by its own .flatpakref, so `flatpak update`
         # on a stable box never jumps to a canary build. The generic-registry version string allows
         # letters/dots/hyphens.
@@ -81,7 +81,7 @@ jobs:
           SHORT=$(echo "$GITHUB_SHA" | cut -c1-8)
           case "$GITHUB_REF" in
             refs/tags/v*) V="${GITHUB_REF_NAME#v}"; BRANCH=stable; ALIAS=latest ;;
-            *)            V="0.3.0-ci${GITHUB_RUN_NUMBER}.g${SHORT}"; BRANCH=canary; ALIAS=canary ;;
+            *)            V="0.5.0-ci${GITHUB_RUN_NUMBER}.g${SHORT}"; BRANCH=canary; ALIAS=canary ;;
           esac
           echo "VERSION=$V" >> "$GITHUB_ENV"
           echo "BUNDLE=punktfunk-client-${V}.flatpak" >> "$GITHUB_ENV"

.gitea/workflows/release.yml

@@ -101,7 +101,7 @@ jobs:
         run: |
           case "$GITHUB_REF" in
             refs/tags/v*) V="${GITHUB_REF_NAME#v}"; V="${V%%-*}" ;;  # App Store marketing version is numeric X.Y.Z (drop -rc)
-            *)            V="0.3.0" ;;                                # canary marketing version; the build number disambiguates
+            *)            V="0.5.0" ;;                                # canary marketing version; the build number disambiguates
           esac
           echo "VERSION=$V" >> "$GITHUB_ENV"
           echo "BUILD_NUM=$GITHUB_RUN_NUMBER" >> "$GITHUB_ENV"
@@ -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
-          # archive (manual signing), not the unsigned-then-codesign DMG path. Same App-Manager
-          # ASC-key constraint as iOS/tvOS — MANUAL signing, NOT -allowProvisioningUpdates
-          # (cloud signing the key can't do). Quit Xcode so it can't prune the dropped profile.
+          # Separate archive from the Developer ID one above: App Store needs a signed, entitled
+          # archive that -exportArchive can re-sign for distribution, not the unsigned-then-codesign
+          # DMG path. Archive with AUTOMATIC signing (development). Why not a manually-specified
+          # 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="$BUILD_NUM" \
-            CODE_SIGN_STYLE=Manual \
-            CODE_SIGN_IDENTITY="Apple Distribution" \
-            DEVELOPMENT_TEAM="$TEAM_ID" \
-            PROVISIONING_PROFILE_SPECIFIER="$PROFILE"
+            CODE_SIGN_STYLE=Automatic \
+            DEVELOPMENT_TEAM="$TEAM_ID"
           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
-          # Store provisioning profile. NOT -allowProvisioningUpdates — with an App-Manager-role
-          # ASC key that forces Xcode's CLOUD-managed signing, which the role can't do ("Cloud
-          # signing permission error"). The profile must be installed on the runner under
-          # ~/Library/Developer/Xcode/UserData/Provisioning Profiles/ (install it once with
-          # Xcode.app quit, or it prunes the manually-dropped distribution profile).
-          # A running Xcode.app prunes unrecognized profiles from that dir — quit it so the App
-          # Store profile survives this build; headless xcodebuild doesn't need the GUI app.
+          # Archive with AUTOMATIC signing (development) — see the macOS App Store step for the full
+          # rationale. The SwiftPM resource bundle (PunktfunkKit_PunktfunkKit, added with the in-app
+          # license screens) builds for iphoneos, so even the sdk-scoped PROVISIONING_PROFILE_SPECIFIER
+          # this step used to set matched it and failed the archive ("does not support provisioning
+          # profiles"). Automatic signing profiles only the app and leaves the resource bundle (and
+          # the macOS-host macro plugins) alone. No -allowProvisioningUpdates → OFFLINE, never
+          # cloud-signs (the App-Manager ASC key can't), so the runner needs an iOS *development*
+          # 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">

.gitea/workflows/rpm.yml

@@ -68,8 +68,8 @@ jobs:
           restore-keys: cargo-home-
 
       - name: Version + channel
-        # vX.Y.Z tag -> X.Y.Z-1 in the base group (a real release); main push -> 0.3.0-0.ciN.g<sha>
-        # in the `<base>-canary` group, whose "0." release sorts below the eventual 0.3.0-1 yet
+        # vX.Y.Z tag -> X.Y.Z-1 in the base group (a real release); main push -> 0.5.0-0.ciN.g<sha>
+        # in the `<base>-canary` group, whose "0." release sorts below the eventual 0.5.0-1 yet
         # climbs by run number. The canary base stays one minor ahead of the latest stable so a
         # stable->canary box re-point still moves forward. The spec %build stamps
         # PUNKTFUNK_BUILD_VERSION from these macros into the binary (--version provenance).
@@ -77,7 +77,7 @@ jobs:
           SHORT=$(echo "$GITHUB_SHA" | cut -c1-8)
           case "$GITHUB_REF" in
             refs/tags/v*) V="${GITHUB_REF_NAME#v}"; R="1"; GROUP="${{ matrix.group }}" ;;
-            *)            V="0.3.0"; R="0.ci${GITHUB_RUN_NUMBER}.g${SHORT}"; GROUP="${{ matrix.group }}-canary" ;;
+            *)            V="0.5.0"; R="0.ci${GITHUB_RUN_NUMBER}.g${SHORT}"; GROUP="${{ matrix.group }}-canary" ;;
           esac
           echo "PF_VERSION=$V" >> "$GITHUB_ENV"
           echo "PF_RELEASE=$R" >> "$GITHUB_ENV"

.gitea/workflows/windows-host.yml

@@ -171,8 +171,8 @@ jobs:
           Push-Location web
           & $bun install --frozen-lockfile; if ($LASTEXITCODE) { throw "bun install failed ($LASTEXITCODE)" }
           & $bun run build; if ($LASTEXITCODE) { throw "web build failed ($LASTEXITCODE)" }
-          if (Select-String -Path .output\server\index.mjs -Pattern 'Bun\.serve' -Quiet) {
-            throw "web build is a bun bundle (Bun.serve) - need the node-server preset"
+          if (-not (Select-String -Path .output\server\index.mjs -Pattern 'Bun\.serve' -Quiet)) {
+            throw "web build is not a bun bundle - need the 'bun' preset + custom entry"
           }
           Pop-Location
           # Gate the installer on a real boot under the BUNDLED bun (the runtime it ships), serving /login.

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**
-   (`VTDecompressionSession` + `CAMetalLayer`) is built and live-validated on glass behind the opt-in
-   `punktfunk.presenter` flag (~11 ms p50 capture→present), to become the default after a few
-   resolution/HDR checks. Next: make stage 2 the default, glass-to-glass numbers via
-   `tools/latency-probe`, iOS/iPadOS/tvOS variants.
+   [`clients/apple/README.md`](clients/apple/README.md). **Stage 2 presenter is now the DEFAULT**
+   (stage-1 is the Metal-unavailable / DEBUG fallback): explicit `VTDecompressionSession` decode →
+   `CAMetalLayer`, presented from the hosting view's **main-runloop `CADisplayLink`** (`renderTick` pops
+   the newest ready frame per vsync; macOS `displaySyncEnabled = false` is the real fullscreen-judder fix,
+   ~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

@@ -1995,7 +1995,7 @@ dependencies = [
 
 [[package]]
 name = "latency-probe"
-version = "0.3.0"
+version = "0.4.1"
 
 [[package]]
 name = "lazy_static"
@@ -2127,7 +2127,7 @@ checksum = "0ceec5bc11778974d1bcb055b18002eba7f4b3518b6a0081b3af5f21666da9ad"
 
 [[package]]
 name = "loss-harness"
-version = "0.3.0"
+version = "0.4.1"
 dependencies = [
  "punktfunk-core",
 ]
@@ -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"
@@ -2709,7 +2720,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-client-android"
-version = "0.3.0"
+version = "0.4.1"
 dependencies = [
  "android_logger",
  "jni",
@@ -2723,7 +2734,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-client-linux"
-version = "0.3.0"
+version = "0.4.1"
 dependencies = [
  "anyhow",
  "async-channel",
@@ -2743,7 +2754,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-client-windows"
-version = "0.3.0"
+version = "0.4.1"
 dependencies = [
  "anyhow",
  "async-channel",
@@ -2763,7 +2774,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-core"
-version = "0.3.0"
+version = "0.4.1"
 dependencies = [
  "aes-gcm",
  "bytes",
@@ -2793,7 +2804,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-host"
-version = "0.3.0"
+version = "0.4.1"
 dependencies = [
  "aes",
  "aes-gcm",
@@ -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",
@@ -2857,7 +2870,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-probe"
-version = "0.3.0"
+version = "0.4.1"
 dependencies = [
  "anyhow",
  "mdns-sd",
@@ -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",
@@ -11,9 +12,11 @@ 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"
+version = "0.4.1"
 edition = "2021"
 rust-version = "1.82"
 license = "MIT OR Apache-2.0"

api/openapi.json

@@ -10,7 +10,7 @@
       "name": "MIT OR Apache-2.0",
       "identifier": "MIT OR Apache-2.0"
     },
-    "version": "0.0.1"
+    "version": "0.4.1"
   },
   "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",

ci/fedora-rpm.Dockerfile

@@ -16,8 +16,9 @@ RUN dnf -y install \
       "https://mirrors.rpmfusion.org/free/fedora/rpmfusion-free-release-$(rpm -E %fedora).noarch.rpm" \
       "https://mirrors.rpmfusion.org/nonfree/fedora/rpmfusion-nonfree-release-$(rpm -E %fedora).noarch.rpm" \
   && dnf -y install \
-      # rpmbuild + source-tarball tooling; nodejs runs the Gitea Actions JS (checkout/cache)
-      # AND the punktfunk-web .output at runtime; unzip is for the bun installer below.
+      # rpmbuild + source-tarball tooling; nodejs runs the Gitea Actions JS (checkout/cache) only
+      # — the punktfunk-web console builds AND runs on bun (installed below); unzip is for the bun
+      # installer.
       rpm-build rpmdevtools systemd-rpm-macros git tar gzip nodejs unzip \
       # build toolchain + bindgen
       gcc gcc-c++ clang clang-devel cmake nasm pkgconf-pkg-config curl ca-certificates \
@@ -28,9 +29,10 @@ RUN dnf -y install \
       gtk4-devel libadwaita-devel SDL3-devel \
   && dnf clean all
 
-# bun — the build tool for the punktfunk-web console (`bun run build` -> the node-server .output
-# the punktfunk-web RPM ships and runs with plain node). Not in Fedora repos; install the official
-# standalone binary to a system PATH dir so the rpmbuild `%build` (run as any uid) finds it.
+# bun — both the BUILD tool and the RUNTIME for the punktfunk-web console (`bun run build` -> the
+# Nitro `bun`-preset .output, served by `Bun.serve` with TLS — HTTP/1.1 over TLS). The
+# RPM vendors THIS bun binary. Not in Fedora repos; install the official standalone binary to a
+# system PATH dir so the rpmbuild `%build`/`%install` (run as any uid) find it.
 RUN curl -fsSL https://bun.sh/install | bash \
     && install -m0755 /root/.bun/bin/bun /usr/local/bin/bun \
     && bun --version

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
-            // proper SDR stream rather than PQ the panel would mis-tone-map).
-            val hdrEnabled = displaySupportsHdr(context)
+            // Advertise HDR only when the user enabled it AND this device's display can present it
+            // (else the host sends a proper SDR stream rather than PQ the panel would mis-tone-map).
+            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:
-     * `[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.
+     * Returns 14 doubles:
+     * `[fps, mbps, latP50Ms, latP95Ms, latValid, skewCorrected, width, height, refreshHz, framesDropped,
+     * 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
-/// `[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;
-/// 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).
+/// Returns 14 doubles
+/// `[fps, mbps, latP50Ms, latP95Ms, latValid, skewCorrected, width, height, refreshHz, framesDropped,
+/// bitDepth, colorPrimaries, colorTransfer, chromaFormatIdc]`
+/// (the two flags are 1.0/0.0; the trailing four describe the negotiated video feed — see below), or
+/// `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

@@ -10,32 +10,59 @@ struct AcknowledgementsView: View {
 
     var body: some View {
         ScrollView {
-            VStack(alignment: .leading, spacing: 18) {
-                Text("punktfunk")
-                    .font(.title2).bold()
-                if let version {
-                    Text("Version \(version)")
-                        .font(.caption)
+            // 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(.geist(22, .bold, relativeTo: .title2))
+                    if let version {
+                        Text("Version \(version)")
+                            .font(.geist(12, relativeTo: .caption))
+                            .foregroundStyle(.secondary)
+                    }
+                    Text(Licenses.appLicense)
+                        .font(.caption.monospaced())
+                        .modifier(SelectableText())
+
+                    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(.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(.geist(12, relativeTo: .caption))
+                    .foregroundStyle(.secondary)
                 }
-                Text(Licenses.appLicense)
-                    .font(.caption.monospaced())
-                    .modifier(SelectableText())
+                .frame(maxWidth: .infinity, alignment: .leading)
+                .padding(.bottom, 18)
 
-                Divider()
-
-                Text("Third-party software")
-                    .font(.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)
-                .foregroundStyle(.secondary)
-                Text(Licenses.thirdPartyNotices)
-                    .font(.caption2.monospaced())
-                    .modifier(SelectableText())
+                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)

clients/apple/Sources/PunktfunkClient/AddHostSheet.swift

@@ -81,6 +81,11 @@ 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
             // window's default/cancel keyboard actions. The 380-wide .fixedSize panel below
@@ -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
-        // scrolls inside the detent — nothing is clipped. (.height(_:) is iOS 16+, safe at iOS 17.)
+        // centered formSheet card). The Form itself is .scrollDisabled (above) so it can't
+        // 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 {
-                    if let fp = model.connection?.hostFingerprint {
-                        store.pin(host.id, fingerprint: fp)
-                    }
-                    awaitingApproval = nil
+                let approvedFingerprint = awaitingApproval?.host.id == host.id
+                    ? model.connection?.hostFingerprint : nil
+                if awaitingApproval?.host.id == host.id { awaitingApproval = nil }
+                // Persist on the next runloop tick: HostStore is an ObservableObject, and mutating
+                // 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.productCategory).font(.caption).foregroundStyle(.secondary)
+                Text(c.name).font(.geist(17, .semibold, relativeTo: .headline))
+                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")
-                    .toolbar {
-                        Button("Done") { showSettings = false }
-                    }
-            }
+            SettingsView()
+                .settingsSheetSizing()
         }
         #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 iconBox: CGFloat
-    let cardPadding: CGFloat
-    let nameFont: Font
+    let tile: CGFloat       // monogram tile side
+    let monogram: CGFloat   // monogram letter point size
+    let name: CGFloat       // host-name point size
+    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) {
-                ZStack {
-                    Image(systemName: "play.display")
-                        .font(.system(size: m.iconSize, weight: .light))
-                        .foregroundStyle(.tint)
-                        .opacity(isConnecting ? 0.3 : 1)
-                    if isConnecting {
-                        ProgressView()
-                    }
+            HStack(spacing: m.spacing) {
+                monogramTile(monogram(host.displayName), m: m, connecting: isConnecting, filled: true)
+                VStack(alignment: .leading, spacing: 4) {
+                    Text(host.displayName)
+                        .font(.geist(m.name, .bold, relativeTo: .title3))
+                        .foregroundStyle(.primary)
+                        .lineLimit(1)
+                    Text("\(host.address):\(String(host.port))")
+                        .font(.geist(m.meta, relativeTo: .caption))
+                        .foregroundStyle(.secondary)
+                        .lineLimit(1)
+                    statusRow(m)
                 }
-                .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)
-                            .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)
-                            .foregroundStyle(.secondary)
-                            .lineLimit(1)
-                    }
-                    if let last = host.lastConnected {
-                        Text("Connected \(last, format: .relative(presentation: .named))")
-                            .font(.caption2)
-                            .foregroundStyle(.tertiary)
-                            .lineLimit(1)
-                    }
+                Spacer(minLength: 0)
+            }
+            .padding(m.padding)
+            .frame(maxWidth: .infinity, alignment: .leading)
+            #if !os(tvOS)
+            // tvOS: the .card button style owns platter + focus motion; extra chrome mutes it.
+            // Elsewhere: a flat material panel with a hairline border (industrial, not a soft blob),
+            // and a brand accent bar down the leading edge for the most-recent host.
+            .background(.regularMaterial)
+            .overlay(alignment: .leading) {
+                if isMostRecent {
+                    Rectangle().fill(Color.brand).frame(width: 3)
                 }
             }
-            .frame(maxWidth: .infinity)
-            .padding(.vertical, m.cardPadding)
-            .padding(.horizontal, 12)
-            #if !os(tvOS)
-            // tvOS: the .card button style owns platter + focus motion — extra chrome
-            // inside it mutes the grow/tilt. Material + accent ring are for pointer UIs.
-            // Deliberately .regularMaterial, not Liquid Glass: HIG keeps glass off content
-            // tiles (it flattens hierarchy over an opaque grid) — see GlassStyle.swift.
-            .background(.regularMaterial, in: RoundedRectangle(cornerRadius: 14))
+            .clipShape(RoundedRectangle(cornerRadius: m.radius, style: .continuous))
             .overlay {
-                if isMostRecent {
-                    RoundedRectangle(cornerRadius: 14)
-                        .strokeBorder(Color.accentColor.opacity(0.35), lineWidth: 1.5)
-                }
+                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;
-/// tapping saves it and connects (or pairs, if the host requires it).
+/// A host found on the LAN but not yet saved. A tinted-outline monogram + dashed panel border
+/// 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) {
-                Image(systemName: "play.display")
-                    .font(.system(size: m.iconSize, weight: .light))
-                    .foregroundStyle(.tint)
-                    .frame(height: m.iconBox)
-                VStack(spacing: 2) {
+            HStack(spacing: m.spacing) {
+                monogramTile(monogram(discovered.name), m: m, connecting: false, filled: false)
+                VStack(alignment: .leading, spacing: 4) {
                     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)
-                            .foregroundStyle(.secondary)
-                            .lineLimit(1)
+                    Text("\(discovered.host):\(String(discovered.port))")
+                        .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")
                     }
-                    Text(discovered.requiresPairing ? "Pairing required" : "Discovered")
-                        .font(.caption2)
-                        .foregroundStyle(.tertiary)
+                    .font(.geist(m.status, .medium, relativeTo: .caption2))
+                    .tracking(0.8)
+                    .foregroundStyle(.secondary)
+                    .lineLimit(1)
                 }
+                Spacer(minLength: 0)
             }
-            .frame(maxWidth: .infinity)
-            .padding(.vertical, m.cardPadding)
-            .padding(.horizontal, 12)
+            .padding(m.padding)
+            .frame(maxWidth: .infinity, alignment: .leading)
             #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

@@ -12,20 +12,36 @@ 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") {
-            #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;
-            // the whole path is absent from Release builds.
-            if let scene = ScreenshotMode.requestedScene {
-                ScreenshotHostView(scene: scene)
-            } else {
+            // 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;
+                // the whole path is absent from Release builds.
+                if let scene = ScreenshotMode.requestedScene {
+                    ScreenshotHostView(scene: scene)
+                } else {
+                    ContentView()
+                }
+                #else
                 ContentView()
+                #endif
             }
-            #else
-            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.
@@ -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()
-                .navigationTitle("Settings")
-                .navigationBarTitleDisplayMode(.inline)
-        }
+        // SettingsView owns its NavigationSplitView (sidebar + detail) and Done button, so it is
+        // rendered directly — a wrapping NavigationStack would nest a split view in a stack. Open
+        // on General so the shot lands on real controls (iPad: sidebar + General detail; iPhone:
+        // 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
-// outgrown one scrolling pane); iOS uses a single grouped Form; 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.
+// Navigation differs per platform, but all three group the same categories (General, Display,
+// Audio, Controllers, Advanced, About): macOS uses a tabbed preferences window; iOS/iPadOS uses
+// an adaptive NavigationSplitView — a category sidebar + detail pane on iPad, auto-collapsing to
+// 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 {
-            streamModeSection
-            audioSection
-            compositorSection
-            presenterSection
-            statisticsSection
-            experimentalSection
-            controllersSection
-            Section {
-                NavigationLink("Acknowledgements") { AcknowledgementsView() }
+        NavigationSplitView(columnVisibility: $columnVisibility) {
+            List(selection: $settingsSelection) {
+                ForEach(SettingsCategory.allCases) { category in
+                    // On iPhone the split view collapses to a push list, but a selection List
+                    // draws no disclosure indicator of its own — add one in compact width for the
+                    // expected drill-in affordance. On iPad the selected row highlights instead, so
+                    // the chevron is omitted there.
+                    HStack {
+                        Label(category.title, systemImage: category.symbol)
+                        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",
-                    options: [("Stage 1 (default)", "stage1"), ("Stage 2 (experimental)", "stage2")],
+                    title: "Presenter (debug)",
+                    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 (experimental)").tag("stage2")
+                Text("Stage 2 (default)").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). "
-                + "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 shortens the present tail. Applies from the next session.")
-                .font(.caption)
+            Text("Stage 2 (default) decodes explicitly and presents through Metal with a display "
+                + "link — it adds a capture→present (glass-to-glass) latency line in the HUD and "
+                + "self-recovers from decode stalls. Stage 1 feeds compressed video straight to the "
+                + "system display layer; it freezes on a lost HEVC reference frame, so it's a debug "
+                + "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)
         }
     }
@@ -405,9 +711,9 @@ struct SettingsView: View {
         } footer: {
             Text("Adds a “Browse Library…” action to each host that lists its games "
                 + "(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)
+                + "Works once you've paired with the host — the library is authorized by this "
+                + "device's certificate, with no extra host setup.")
+                .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.
-            if !ok { self.teardown() }
+            // still holds an exclusive reference to. Back off so a broken XPC isn't re-hit every
+            // 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
-            // error). Do NOT latch broken — the next nonzero amplitude retries setup().
-            log.warning("rumble: haptics present but engine setup failed — will retry on next rumble")
+            // Haptics present but no engine could be built right now (server busy / XPC broken). Do
+            // NOT latch broken — back off and the next nonzero amplitude past the cooldown retries.
+            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,
-                    // and the cap + stop check keep sustained 0xCD traffic (a game writing
-                    // per-frame LED/trigger reports) from starving the rumble poll above
-                    // or blocking stop() past one cycle.
+                    // Drain a BOUNDED burst of hidout events so sustained 0xCD traffic (a game writing
+                    // per-frame LED/trigger reports) can't spin here or block stop() past one cycle.
                     var burst = 0
                     while burst < 64, !flag.isStopped,
-                          let ev = try connection.nextHidOutput(
-                            timeoutMs: burst == 0 ? hidTimeout : 0) {
+                          let ev = try connection.nextHidOutput(timeoutMs: 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/LibraryClient.swift

@@ -3,10 +3,11 @@
 // /library page renders. Read-only on the client for now; launching a chosen title is a later
 // step. Gated behind `DefaultsKey.libraryEnabled` in the UI.
 //
-// The management API is HTTP on a port distinct from the punktfunk/1 data plane (default 47990),
-// binds loopback unless started with a token, and REQUIRES a bearer token for any non-loopback
-// bind. So to browse a host's library remotely the host must expose the mgmt API on the LAN with
-// `--mgmt-token`; the client carries that token per host. This mirrors the GameEntry/Artwork/
+// The management API serves HTTPS on a port distinct from the punktfunk/1 data plane (default
+// 47990, also advertised in the host's mDNS `mgmt` TXT). A paired client is authorized for the
+// read-only library route by its **mTLS certificate** — no bearer token. The host binds this read
+// surface to the LAN by DEFAULT (the bearer-gated admin surface stays loopback-only), so a paired
+// client browses a host's library with no operator step. This mirrors the GameEntry/Artwork/
 // LaunchSpec schema in `crates/punktfunk-host/src/library.rs`.
 
 import Foundation
@@ -56,8 +57,9 @@ public enum LibraryError: LocalizedError {
         case .http(let code):
             return "The management API returned HTTP \(code)."
         case .unreachable(let why):
-            return "Couldn't reach the host's management API: \(why). The host must expose it on "
-                + "the LAN (serve --mgmt-bind 0.0.0.0)."
+            return "Couldn't reach the host's management API: \(why). It binds the LAN by default, "
+                + "so check the host is updated and reachable (a host pinned to "
+                + "`--mgmt-bind 127.0.0.1` is loopback-only and can't be browsed remotely)."
         }
     }
 }

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
-// drawable with a BT.709 YUV→RGB shader. The display link (owned by the hosting view) drives
-// `render` once per vsync with the target present time, so a present can finally be stamped and
-// the present tail hand-paced. See docs apple-stage2-presenter.md.
+// Stage-2 presenter, present half: draw a decoded NV12 / P010 / 4:4:4 CVPixelBuffer into a CAMetalLayer
+// drawable with a Y′CbCr→RGB shader. The hosting view's CADisplayLink drives `render` once per vsync
+// (via Stage2Pipeline.renderTick) with the target present time, so a present can be stamped and the
+// present tail hand-paced. See docs apple-stage2-presenter.md.
 //
-// Main-thread only: created during view setup, `render` called from the view's CADisplayLink
-// (which fires on the main runloop). The Metal objects + texture cache are touched only here.
+// 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. 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
-/// 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
-/// for now — matches the host; 10-bit/HDR + other matrices are a later tie-in.)
+private let presenterLog = Logger(subsystem: "io.unom.punktfunk", category: "presenter")
+
+/// HDR reference white (BT.2408 "HDR Reference White"): the absolute luminance, in nits, that the
+/// 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
-// matrix to get PQ-encoded R'G'B' and output it as-is — the CAMetalLayer's itur_2100_PQ colour
-// space + EDR tells the compositor the samples are PQ, so it does the PQ→display mapping. No EOTF
-// here (matching the host, which emitted BT.2020 PQ). P010 stores the 10-bit code in the high bits
-// of each 16-bit sample, so an .r16Unorm sample reads ~code/1023 (the /1024 vs /1023 error is < 0.1%).
+// 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 space
+// + edrMetadata tell the compositor the samples are PQ, so it does the PQ→display tone-map. No EOTF
+// here. P010/x444 store the 10-bit code in the high bits of each 16-bit sample, so an .r16Unorm sample
+// 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.
-    /// Selected per frame by `render`; the layer is reconfigured when the mode flips (HDR toggle).
+    /// SDR (BT.709 8-bit → bgra8) and HDR (BT.2020 PQ 10-bit → rgba16Float) pipelines. Selected per
+    /// 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.
-    public init?() {
+    /// Current layer configuration — switched in `configure(hdr:)` when a frame's HDR-ness differs.
+    /// 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
-        self.queue = queue
+        let pipelineSDR: MTLRenderPipelineState
+        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)
-        guard textureCache != nil else { return nil }
+        var cache: CVMetalTextureCache?
+        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
-        // own vsync wait on means `commandBuffer.present` ALSO blocks for the hardware vsync,
-        // so `nextDrawable()` stalls the MAIN thread until a drawable frees — windowed, the
-        // WindowServer's looser compositing hides it; FULLSCREEN's tighter, more-direct path
-        // serializes the main thread to the display and the stall surfaces as bad judder.
-        // Disabling the layer-level sync lets present return promptly (the display link is the
-        // pacing source), which is what fixes the fullscreen stutter. macOS-only property.
+        // The display link already paces exactly one present per vsync. Leaving the layer's own vsync
+        // wait on means `commandBuffer.present` ALSO blocks for the hardware vsync, so `nextDrawable()`
+        // stalls the MAIN thread until a drawable frees — windowed, the WindowServer's looser
+        // compositing hides it; FULLSCREEN's tighter path serializes the main thread to the display and
+        // the stall surfaces as bad judder. Disabling the layer-level sync lets present return promptly
+        // (the display link is the pacing source) — the fix for the fullscreen stutter. macOS-only.
         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.
-    public func setDrawableSize(_ size: CGSize) {
-        guard size.width > 0, size.height > 0 else { return }
-        if layer.drawableSize != size { layer.drawableSize = size }
-    }
-
-    /// 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) {
+    /// Configure the layer + active pipeline for an SDR or HDR session. MAIN THREAD ONLY. Called once at
+    /// session start and again per-frame from `render` (idempotent — the guard makes a same-state call a
+    /// no-op), so a mid-session HDR toggle (the host re-inits its encoder; the decoded `frame.isHDR`
+    /// 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) {
         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
-    /// BT.2020 PQ path (P010 input) vs the 8-bit BT.709 path (NV12 input). 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 for this frame.
+    #if !os(tvOS)
+    private func makeEDR(_ meta: PunktfunkConnection.HdrMeta?) -> CAEDRMetadata {
+        CAEDRMetadata.hdr10(
+            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
-        let chromaFmt: MTLPixelFormat = isHDR ? .rg16Unorm : .rg8Unorm
+
+        // P010/x444 store 10-bit luma/chroma in 16-bit samples → R16/RG16; NV12/444v is 8-bit → R8/RG8.
+        // 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 chroma = makeTexture(pixelBuffer, plane: 1, format: chromaFmt, cache: textureCache)
+              let luma = makeTexture(
+                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
-        // out. The fullscreen triangle scales the decoded texture to fill the drawable.
-        guard layer.drawableSize.width > 0, layer.drawableSize.height > 0,
-              let drawable = layer.nextDrawable(),
+        // Size the drawable to the decoded frame so the fullscreen triangle samples 1:1 (pixel-exact);
+        // the layer's contentsGravity then scales it to the on-screen bounds via the system compositor
+        // (matching stage-1). drawableSize does NOT track bounds (defaults to 0), so set it BEFORE
+        // 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
-        // finishes sampling — releasing them at scope exit could free the backing mid-read.
+        // Hold the CVMetalTextures + source pixel buffer (its IOSurface) alive until the GPU finishes
+        // 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
-    /// the draw — the MTLTexture is only valid while its CVMetalTexture is retained.
+    /// Returns the CVMetalTexture (not just its MTLTexture) so the caller can keep it alive past the
+    /// draw — the MTLTexture is only valid while its CVMetalTexture is retained.
     private func makeTexture(
-        _ pixelBuffer: CVPixelBuffer, plane: Int, format: MTLPixelFormat,
-        cache: CVMetalTextureCache
+        _ pixelBuffer: CVPixelBuffer, plane: Int, format: MTLPixelFormat, 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
-    /// default device; on iOS the UIDs are ignored entirely (routes are
-    /// AVAudioSession-managed). Main thread (engine setup); returns after the engines
-    /// start — the mic may start slightly later if the permission prompt is pending.
+    /// Start playback (and, if enabled+authorized, the mic uplink). Empty UIDs = system default
+    /// device; on iOS the UIDs are ignored entirely (routes are AVAudioSession-managed). On macOS
+    /// the engines start synchronously on the caller's (main) thread. On iOS/tvOS start() is
+    /// 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)
-        // Route + policy live in the session, not per-engine: stereo playback, mic
-        // capture when enabled, Bluetooth allowed. Failure is non-fatal (defaults).
+        #if os(macOS)
+        // No AVAudioSession on macOS — start the engines directly (caller's thread, as before).
+        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
-                // EARPIECE; only affects the built-in route (headphones/BT still win).
+                // .defaultToSpeaker: .playAndRecord otherwise routes to the iPhone EARPIECE; only
+                // 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
+    }
+    #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 !os(macOS)
+        // Release the session so audio we interrupted (Music, podcasts) gets its resume cue. Like
+        // 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))
         }
-        #if !os(macOS)
-        // Release the session so audio we interrupted (Music, podcasts) gets its
-        // resume cue.
-        do {
-            try AVAudioSession.sharedInstance().setActive(
-                false, options: .notifyOthersOnDeactivation)
-        } catch {
-            log.warning("AVAudioSession deactivation failed: \(error.localizedDescription)")
-        }
-        #endif
     }
 
     // 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
-// output drops the newest decoded frame into a 1-slot ring; the hosting view's display link
-// calls `renderTick` once per vsync to draw + present the newest ready frame and stamp
-// capture→present. Mirrors StreamPump's lifecycle (one per start; cancel is permanent).
+// Stage-2 presenter orchestrator: a pump thread pulls AUs → VideoDecoder; the decoder's async output
+// drops the newest decoded frame into a 1-slot ring; the hosting view's display link calls `renderTick`
+// once per vsync to draw + present the newest ready frame and stamp capture→present. Mirrors
+// 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`
-// + `setDrawableSize` + `start`/`stop` on the MAIN thread (the view's CADisplayLink fires there).
-// Only the ring + decoder cross threads and both are internally locked.
+// Threading: the pump runs on its own thread; the decoder callback on a VT thread; `renderTick` +
+// `start`/`stop` on the MAIN thread (the view's CADisplayLink fires there). Only the ring (lock-guarded)
+// 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
-/// directly makes a `view → link → view` cycle that only `invalidate()` breaks — if a teardown
-/// is ever missed the view leaks and keeps ticking. This proxy holds the handler weakly, so the
-/// view can deallocate and its `deinit` invalidate the link.
+/// Weak-target wrapper for CADisplayLink. The link retains its target, so targeting a view directly
+/// makes a `view → link → view` cycle that only `invalidate()` breaks — if a teardown is ever missed
+/// the view leaks and keeps ticking. This proxy holds the handler weakly, so the view can deallocate
+/// 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
-/// (a VT thread) and the pump thread (a submit failure) both signal a wedge; this coalesces
-/// them so the control stream isn't flooded while the decode stays stalled for several frames
-/// until the requested IDR lands. Bound to the live connection in `start`, unbound in `stop`.
+/// Throttled host keyframe requests for decode recovery. The decoder's async error callback (a VT
+/// thread) and the pump thread (a submit failure) both signal a wedge; this coalesces them so the
+/// control stream isn't flooded while the decode stays stalled for several frames until the requested
+/// 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
-    /// unavailable so the caller can fall back to stage-1.
+    /// The Metal layer the hosting view installs + sizes.
     public var layer: CAMetalLayer { presenter.layer }
 
-    /// `presentMeter` records capture→present (the glass-to-glass term). Returns nil if Metal
-    /// can't be set up (headless / no GPU) — caller falls back to the stage-1 presenter.
+    /// `presentMeter` records capture→present (the glass-to-glass term). Returns nil if Metal can't be
+    /// 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
-            // resets to re-gate on the next IDR, and we ask the host to send one now (infinite
-            // GOP — it wouldn't otherwise come soon). Throttled in KeyframeRecovery.
+            // Async decode failure (a bad P-frame referencing a lost/corrupt IDR): the pump resets to
+            // re-gate on the next IDR, and we ask the host to send one now (infinite GOP — it wouldn't
+            // otherwise come soon). Throttled in KeyframeRecovery.
             onDecodeError: { _ in recovery.request() })
     }
 
-    /// Start pulling AUs into the decoder. `onFrame` fires per AU at receipt (capture→client
-    /// meter, exactly as stage-1); `onSessionEnd` on close. `clockOffsetNs` (host minus client)
-    /// makes the present stamp cross-machine valid.
+    /// 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) makes the
+    /// 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
-                    // AUs (framesDropped) and the decoder then conceals the reference-missing delta
-                    // frames that follow — often rendering them WITHOUT an error callback — so the
-                    // onDecodeError trigger rarely fires after a real network blip. Ask the host for
-                    // 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.
+                    // Loss recovery (the primary path). The reassembler drops unrecoverable AUs and the
+                    // decoder conceals the reference-missing deltas — often WITHOUT an error callback —
+                    // so key off the drop count climbing, then keep asking (awaitingIDR) until a fresh
+                    // IDR re-anchors decode.
                     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
-                    // attaches it to subsequent HDR frames. Non-blocking; only HDR sessions emit these.
-                    if connection.isHDR, let meta = try? connection.nextHdrMeta(timeoutMs: 0) {
-                        decoder.setHdrMeta(meta)
+                    if awaitingIDR { recovery.request() }
+                    // Drain HDR mastering metadata (0xCE) and hand it to the PRESENTER (→ CAEDRMetadata).
+                    // Polled UNCONDITIONALLY (not gated on connection.isHDR, the fixed Welcome flag): the
+                    // 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)
+                        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) {
-                        // Submit/decoder error: drop the session and re-gate on the next IDR's
-                        // in-band parameter sets (a delta frame can't recover) — stage-1's policy
-                        // — and ask the host for that IDR now (infinite GOP; throttled).
+                    if decoder.decode(au: au, format: f) {
+                        decodeFailRun = 0
+                    } else {
+                        // 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
-    /// capture→present at `targetPresentNs` — the display link's target present instant, already
-    /// converted to `CLOCK_REALTIME` (see `realtimeNs(forDisplayLinkTimestamp:)`).
+    /// MAIN thread, once per vsync. Present the newest ready frame (if any) and stamp capture→present at
+    /// `targetPresentNs` — the display link's target present instant, already converted to
+    /// `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).
-    public func setDrawableSize(_ size: CGSize) {
-        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).
+    /// Stop the pump (≤ one poll timeout) and drop the decode session. MAIN THREAD; idempotent. 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
-    /// target present time (when the frame is actually on glass), not the moment we drew.
+    /// nanosecond instant — the present clock the AU pts + skew offset live in. Projects to the target
+    /// 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
-            // the IDR lands, so requesting on every frame would flood the control stream.
+            // Recovery is a persistent WANT, not a one-shot edge: set it on detected loss (or a
+            // 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
-                    // only recovery keyframe is one we request. The reassembler drops unrecoverable
-                    // AUs (framesDropped); the decoder then *conceals* the reference-missing delta
-                    // frames that follow — a frozen / garbage picture, WITHOUT flipping the layer to
-                    // .failed — so the .failed check below rarely fires after a real network blip.
-                    // Ask the host for a fresh IDR whenever the drop count climbs. Polled every
-                    // iteration (not just per AU) so a total-loss drought still recovers the moment
-                    // packets resume and the reassembler counts the gap.
+                    // Loss recovery (the primary path). Under the host's infinite GOP the only
+                    // recovery keyframe is one we request. The reassembler drops unrecoverable AUs
+                    // (framesDropped); the decoder then *conceals* the reference-missing deltas — a
+                    // frozen / garbage picture that never flips the layer to .failed — so key off the
+                    // drop count climbing, then keep asking (awaitingIDR) until an IDR lands. Polled
+                    // every iteration so a total-loss drought still recovers when packets resume.
                     let dropped = connection.framesDropped()
                     if dropped > lastFramesDropped {
+                        // Log only on the false→true transition (once per recovery cycle), not per
+                        // 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
-                        requestKeyframeThrottled()
+                        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) {
-                        format = f // refreshed on every IDR (mode changes included)
+                    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")
+                        }
+                        awaitingIDR = false // a fresh IDR re-anchored decode — recovery complete
                     }
-                    if layer.status == .failed {
+                    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
-                        // a delta frame can't recover), AND ask the host for a fresh IDR. Throttled:
-                        // the layer stays .failed across several polls until the IDR lands.
+                        // IDR): flush and, unless THIS AU is the recovering IDR (re-anchored above),
+                        // re-gate on the next in-band parameter sets and keep asking — enqueuing a
+                        // 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)
-                        requestKeyframeThrottled()
+                        if idrFormat == nil {
+                            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
-    /// display link instead of the StreamPump → displayLayer path. nil = stage-1 (default).
+    /// Stage-2 presenter (default): a CAMetalLayer sublayer driven by a display link instead of the
+    /// 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
-        // (`punktfunk.presenter == "stage2"`) takes explicit VTDecompressionSession decode + a
-        // CAMetalLayer/display-link present; it falls back here if Metal can't be set up.
-        if UserDefaults.standard.string(forKey: DefaultsKey.presenter) == "stage2",
+        // Presenter choice — stage-2 is the DEFAULT (explicit VTDecompressionSession decode + a
+        // CAMetalLayer/display-link present): it can detect + recover a wedged decoder where
+        // stage-1's AVSampleBufferDisplayLayer freezes hard on a lost HEVC reference. 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)
@@ -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,
-    /// so this is usually the full bounds; it letterboxes a resized window). drawableSize is the
-    /// layer's pixel size — the fullscreen-triangle shader scales the decoded texture to fill it.
+    /// 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; it letterboxes a resized window). 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, 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
-// the cursor — the gaming-grade path. When it CAN'T lock (Stage Manager, not frontmost,
-// iPhone) the system shows its own cursor 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.
-// GCMouse is gated off whenever the lock isn't held so the two paths can't double-send.
+// (full-screen + frontmost iPad, and the user hasn't disabled pointer capture in Settings —
+// see PointerLockChain, which steers the lock request through SwiftUI's hosting controllers)
+// GCMouse delivers raw relative deltas and the system hides the cursor — the gaming-grade path.
+// InputCapture handles EVERY connected mouse (GCMouse.mice), not just the current one, so a
+// trackpad + a second pointer (e.g. a Universal Control mouse) both drive. When the scene CAN'T
+// lock (Stage Manager, not frontmost, iPhone, capture disabled) the system shows its own cursor
+// 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
-    /// CADisplayLink instead of the StreamPump → displayLayer path. nil = stage-1 (default).
+    /// Stage-2 presenter (default): a CAMetalLayer sublayer driven by a CADisplayLink instead of the
+    /// 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
-    // only when capture is engaged. The system additionally requires full-screen + frontmost
-    // and may drop it (Slide Over/Stage Manager/backgrounding) — verified in setCaptured().
-    public override var prefersPointerLocked: Bool {
-        captured && UIDevice.current.userInterfaceIdiom == .pad
+    /// Whether the user wants the mouse/trackpad pointer CAPTURED (pointer lock → relative
+    /// movement, the gaming default) rather than forwarded as an absolute position (desktop
+    /// use). Read live from UserDefaults so it tracks the Settings toggle; defaults to on when
+    /// unset. iPad-only — gated again in `prefersPointerLocked`.
+    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
-    // its pointer-lock decision to its children will then reach this VC. (UIHostingController
-    // itself does not consult children, which is why GCMouse deltas can never arrive there —
-    // the touch path, always forwarded, is the unconditional fallback.)
-    public override var childViewControllerForPointerLock: UIViewController? { self }
+    // NOTE: we deliberately do NOT override `childViewControllerForPointerLock`. The default
+    // returns nil, which tells the system to use THIS controller's own `prefersPointerLocked` —
+    // exactly what we want, since `PointerLockChain` forces our SwiftUI ancestors to forward the
+    // downward walk to us and we are the terminal anchor. Returning `self` here would make the
+    // 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
-        // through InputCapture so the forwarding gate and release-on-blur apply uniformly.
+        // Indirect pointer (mouse/trackpad) WITHOUT a lock → absolute cursor + buttons + scroll.
+        // 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
-        // (`punktfunk.presenter == "stage2"`) takes VTDecompressionSession decode + a
-        // CAMetalLayer/display-link present; falls back here if Metal can't be set up.
-        if UserDefaults.standard.string(forKey: DefaultsKey.presenter) == "stage2",
+        // Presenter choice — stage-2 is the DEFAULT (VTDecompressionSession decode + a
+        // CAMetalLayer/display-link present): it can detect + recover a wedged decoder, where
+        // stage-1's AVSampleBufferDisplayLayer freezes hard on a lost HEVC reference frame with no
+        // 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,
-    /// so this is usually the full bounds). drawableSize is the layer's pixel size; the shader's
-    /// fullscreen triangle scales the decoded texture to fill it.
+    /// The display scale to render the metal drawable at. `traitCollection.displayScale` is the
+    /// canonical render scale and is reliable once the controller is in the hierarchy;
+    /// `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
-    /// each decoded HDR pixel buffer so the compositor tone-maps from the real grade. Guarded by its
-    /// own lock — written by the pump thread, read on the VT decode callback.
-    private let metaLock = NSLock()
-    private var hdrMeta: PunktfunkConnection.HdrMeta?
+    /// Whether the negotiated stream is full-chroma 4:4:4 (`connection.isChroma444`), set once at
+    /// session start before any decode. Selects the 4:4:4 decode pixel format (orthogonal to bit
+    /// depth / HDR). Read inside `createSessionLocked` under `lock`.
+    private var chroma444 = false
 
     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
-    /// attached to subsequent decoded HDR pixel buffers. Thread-safe; cheap to call on each update.
-    public func setHdrMeta(_ meta: PunktfunkConnection.HdrMeta) {
-        metaLock.lock()
-        hdrMeta = meta
-        metaLock.unlock()
+    /// Select the chroma subsampling of the decode output (4:2:0 vs full-chroma 4:4:4). Call once at
+    /// session start, before decoding, from `connection.isChroma444`. Takes effect on the next
+    /// session (re)build. Thread-safe.
+    public func setChroma444(_ on: Bool) {
+        lock.lock()
+        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
-    /// (the host re-emits parameter sets with the new VUI → a new format desc → session rebuild).
+    /// HEVC VUI, so this picks the decode bit depth (10-bit P010/x444 vs 8-bit NV12/444v) from the
+    /// 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 =
-            hdr
-            ? kCVPixelFormatType_420YpCbCr10BiPlanarVideoRange // P010 (10-bit)
-            : kCVPixelFormatType_420YpCbCr8BiPlanarVideoRange // NV12 (8-bit)
+        let pixelFormat: OSType = {
+            switch (chroma444, hdr) {
+            case (false, false): return kCVPixelFormatType_420YpCbCr8BiPlanarVideoRange // NV12
+            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)
-            == kCVPixelFormatType_420YpCbCr10BiPlanarVideoRange
-        // Attach the source's mastering display + content light level (ST.2086 / CEA-861.3) so the
-        // compositor tone-maps from the real grade rather than inferring from the PQ colourspace
-        // 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)
-            }
-        }
+            fmt == kCVPixelFormatType_420YpCbCr10BiPlanarVideoRange
+            || fmt == kCVPixelFormatType_420YpCbCr10BiPlanarFullRange
+            || fmt == kCVPixelFormatType_444YpCbCr10BiPlanarVideoRange
+            || fmt == kCVPixelFormatType_444YpCbCr10BiPlanarFullRange
         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) => ({
-            data: g,
-            label: g === "auto" ? "Automatic" : g === "xbox360" ? "Xbox 360" : "DualSense",
-          }))}
+          rgOptions={GAMEPADS.map((g) => ({ data: g, label: GAMEPAD_LABELS[g] ?? g }))}
           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(
-                self.subsystem
-                    .type_for_id(sdl3::sys::joystick::SDL_JoystickID(id)),
-            ),
+            pref,
         })
     }
 
@@ -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
-                        .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,
-                        });
+                    w.forward_touch(which, touchpad as u32, finger as u8, x, y, true);
                 }
                 Event::ControllerTouchpadUp {
                     which,
+                    touchpad,
                     finger,
                     x,
                     y,
                     ..
                 } if active == Some(which) && w.attached.is_some() => {
-                    let _ = w
-                        .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,
-                        });
+                    w.forward_touch(which, touchpad as u32, finger as u8, x, y, false);
                 }
                 // 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(
-                self.subsystem
-                    .type_for_id(sdl3::sys::joystick::SDL_JoystickID(id)),
-            ),
+            pref,
         })
     }
 
@@ -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
-                        .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,
-                        });
+                    w.forward_touch(which, touchpad as u32, finger as u8, x, y, true);
                 }
                 Event::ControllerTouchpadUp {
                     which,
+                    touchpad,
                     finger,
                     x,
                     y,
                     ..
                 } if active == Some(which) && w.attached.is_some() => {
-                    let _ = w
-                        .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,
-                        });
+                    w.forward_touch(which, touchpad as u32, finger as u8, x, y, false);
                 }
                 // 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
-/// `Hello`/`Welcome` — older peers simply omit/ignore it (an unknown byte degrades to `Auto`).
+/// byte (`0 = Auto`, `1 = Xbox360`, `2 = DualSense`, `3 = XboxOne`, `4 = DualShock4`,
+/// `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
-//! "CABLE Input", VoiceMeeter, or anything with "virtual" in the name. If none 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.
+//! VB-Audio "CABLE Input" (bundled by the installer — the preferred, dedicated mic target), the
+//! "Steam Streaming Microphone", VoiceMeeter, or anything with "virtual" in the name.
+//! [`super::audio_control`] sets the default playback to a DIFFERENT loopback-capable device so the
+//! 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
-    // NV12/P010 by default, which can't reconstruct 4:4:4; route a 4:4:4 session to DDA, which delivers
-    // RGB (Bgra) when its `chroma_444` flag is set. (IDD-push/WGC 4:4:4 capture is a follow-up.)
-    if want.chroma_444 && capture != CaptureBackend::Dda {
-        tracing::info!("4:4:4 session — using DDA capture (RGB source) instead of {capture:?}");
-        return dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
-            .map(|c| Box::new(c) as Box<dyn Capturer>);
-    }
-    // P2 direct frame push (kill DDA): consume frames straight from the pf-vdisplay driver's shared
-    // 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>)
-        }
-        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)
-                .map(|c| Box::new(c) as Box<dyn Capturer>)
-        }
-    }
+    // IDD direct-push is the sole Windows capture path: consume frames straight from the pf-vdisplay
+    // driver's shared ring (in-process, Session 0 — it captures the secure desktop too; no Desktop
+    // Duplication, no WGC helper). A FRESH monitor + ring is created per session: a REUSED monitor's
+    // swap-chain dies after ~2 sessions and can't be revived. The ring is always FP16 when the display
+    // is HDR (the driver composes the IDD in FP16); `want.hdr` proactively enables advanced color and
+    // selects the per-frame conversion (FP16 → P010 vs BGRA → NV12). `IddPushCapturer` takes the
+    // 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.
+    idd_push::IddPushCapturer::open(target, pref, want.hdr, keep)
+        .map(|c| Box::new(c) as Box<dyn Capturer>)
+        .map_err(|(e, _keep)| e.context("IDD-push capture open (no fallback)"))
+}
+
+/// Whether the active capturer can deliver a full-chroma (RGB) source for a 4:4:4 HEVC encode. The
+/// 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/`
-// (`#[path]` keeps the module names flat, so every `crate::capture::*` path is unchanged).
-#[cfg(target_os = "windows")]
-#[path = "capture/windows/composed_flip.rs"]
-pub mod composed_flip;
-#[cfg(target_os = "windows")]
-#[path = "capture/windows/desktop_watch.rs"]
-pub mod desktop_watch;
+// 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). Windows capture
+// is IDD direct-push only — DXGI Desktop Duplication (DDA) and the WGC two-process relay were removed.
 #[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
-//! token that canNOT create named kernel objects, so — exactly like the gamepad UMDF drivers
-//! (`inject/dualsense_windows.rs`) — the HOST (privileged) CREATES the shared header + frame-ready
-//! event + ring of keyed-mutex textures (`Global\` names, permissive `D:(A;;GA;;;WD)` SDDL) on the
-//! discrete render GPU, and the driver only OPENS them and copies frames in. We then consume the ring
+//! P2 direct frame push (kill DDA) — HOST side. The pf-vdisplay driver's WUDFHost canNOT create named
+//! kernel objects, so — exactly like the gamepad UMDF drivers (`inject/dualsense_windows.rs`) — the
+//! HOST (privileged) CREATES the shared header + frame-ready event + ring of keyed-mutex textures
+//! (`Global\` names, scoped `D:(A;;GA;;;SY)(A;;GA;;;LS)` to SYSTEM + the driver's LocalService host —
+//! 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
-/// can OPEN the host-created objects — the same `D:(A;;GA;;;WD)` SDDL the gamepad shared section uses.
-/// The returned `psd` backing must outlive `sa`; both are dropped when the process exits.
-unsafe fn permissive_sa() -> Result<(SECURITY_ATTRIBUTES, PSECURITY_DESCRIPTOR)> {
+/// Build a `SECURITY_ATTRIBUTES` granting GENERIC_ALL to **SYSTEM** (the host creates+publishes the
+/// shared event + texture ring) and **LocalService** (the account the pf_vdisplay WUDFHost runs under,
+/// which consumes them) — `D:(A;;GA;;;SY)(A;;GA;;;LS)`, the same scoping as the gamepad section. The
+/// 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
-//! plan-named `secure_dda`/`idd_depth`/`zerocopy`/`ten_bit`/`four_four_four` and the multi-site `perf`/`compositor`/
+//! `encoder_pref`, `render_adapter`, the vdisplay backend select — plus the plan-named
+//! `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
-    /// capture; no WGC helper). **Value-aware** (`0`/`false`/`no`/`off`/empty ⇒ off, else on); unset ⇒ off.
-    /// The installer's default `host.env` sets it on, so a fresh install runs the validated IDD-push path
-    /// (it falls back to DDA if the driver can't attach — see [`crate::capture`]). NOT a bare presence flag
-    /// (so an operator can turn it OFF in `host.env` with `=0`, which a `var_os` presence check can't).
+    /// `PUNKTFUNK_IDD_PUSH` — IDD direct-push monitor mode (the per-session monitor + ring recreate and
+    /// the discrete-render-GPU pin in [`crate::vdisplay::manager`]). IDD-push is the sole Windows capture
+    /// path (DXGI Desktop Duplication and the WGC relay were removed), so this should stay on — the
+    /// installer's `host.env` sets it. **Value-aware** (`0`/`false`/`no`/`off`/empty ⇒ off, else on);
+    /// 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/discovery.rs

@@ -11,7 +11,10 @@
 //!   lets a picker show the fingerprint and pre-pin a chosen host;
 //! - `pair` — `required` or `optional`, so a client can tell up front whether it must run the PIN
 //!   pairing ceremony before it can stream;
-//! - `id` — the stable host uniqueid (dedup across IPs / re-advertises).
+//! - `id` — the stable host uniqueid (dedup across IPs / re-advertises);
+//! - `mgmt` — the management API's TCP port (when it serves one), so a client can fetch the host's
+//!   game library (`GET /api/v1/library`, mTLS) on the SAME IP without assuming the default port.
+//!   Omitted by a host with no mgmt API (the standalone `punktfunk1-host`).
 
 use anyhow::{Context, Result};
 use mdns_sd::{ServiceDaemon, ServiceInfo};
@@ -30,7 +33,9 @@ pub struct Advert {
 }
 
 /// Advertise the native host on the LAN. `fingerprint` is the host cert SHA-256 (lowercase hex);
-/// `require_pairing` tells a discovering client whether it must pair before it can stream.
+/// `require_pairing` tells a discovering client whether it must pair before it can stream;
+/// `mgmt_port` is the management API's port (`Some` when this host serves one — the client browses
+/// the library there over mTLS on the advertised IP), `None` for a host with no mgmt API.
 pub fn advertise_native(
     hostname: &str,
     ip: IpAddr,
@@ -38,6 +43,7 @@ pub fn advertise_native(
     fingerprint: &str,
     require_pairing: bool,
     uniqueid: &str,
+    mgmt_port: Option<u16>,
 ) -> Result<Advert> {
     let daemon = ServiceDaemon::new().context("create mDNS daemon")?;
     let host_name = format!("{hostname}.local.");
@@ -54,6 +60,9 @@ pub fn advertise_native(
         .into(),
     );
     props.insert("id".into(), uniqueid.to_string());
+    if let Some(mgmt) = mgmt_port {
+        props.insert("mgmt".into(), mgmt.to_string());
+    }
     let service = ServiceInfo::new(NATIVE_SERVICE, hostname, &host_name, ip, port, props)
         .context("build native mDNS ServiceInfo")?;
     daemon

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

@@ -17,6 +17,10 @@ pub struct AppEntry {
     pub compositor: Option<crate::vdisplay::Compositor>,
     /// Command gamescope runs nested (gamescope entries only).
     pub cmd: Option<String>,
+    /// Store-qualified library id (`steam:570`, `epic:…`) for entries surfaced from the host's game
+    /// library ([`crate::library`]). When set, the launch path resolves + launches it against the
+    /// host's own library instead of running [`cmd`](Self::cmd). `None` for Desktop / apps.json entries.
+    pub library_id: Option<String>,
 }
 
 fn config_path() -> Option<std::path::PathBuf> {
@@ -35,9 +39,18 @@ fn parse_compositor(s: &str) -> Option<crate::vdisplay::Compositor> {
     }
 }
 
-/// The catalog: the user's `apps.json` if present, else defaults (Desktop, plus gamescope
-/// entries when gamescope is installed).
+/// The GameStream catalog Moonlight sees in `/applist`: the operator base ([`base_catalog`] — Desktop +
+/// apps.json) with the host's auto-detected game library ([`append_library`]) layered on top, so a
+/// Moonlight client sees the same Steam/Epic/GOG/Xbox titles the native clients do instead of just Desktop.
 pub fn catalog() -> Vec<AppEntry> {
+    let mut apps = base_catalog();
+    append_library(&mut apps);
+    apps
+}
+
+/// The operator base: the user's `apps.json` if present, else defaults (Desktop, plus gamescope
+/// entries when gamescope is installed). The installed game library is layered on by [`append_library`].
+fn base_catalog() -> Vec<AppEntry> {
     if let Some(path) = config_path() {
         if let Ok(raw) = std::fs::read_to_string(&path) {
             match serde_json::from_str::<Value>(&raw) {
@@ -53,6 +66,7 @@ pub fn catalog() -> Vec<AppEntry> {
                                     .and_then(|c| c.as_str())
                                     .and_then(parse_compositor),
                                 cmd: it.get("cmd").and_then(|c| c.as_str()).map(String::from),
+                                library_id: None,
                             })
                         })
                         .collect();
@@ -72,6 +86,7 @@ pub fn catalog() -> Vec<AppEntry> {
         title: "Desktop".into(),
         compositor: None,
         cmd: None,
+        library_id: None,
     }];
     if which("gamescope") {
         if which("steam") {
@@ -80,6 +95,7 @@ pub fn catalog() -> Vec<AppEntry> {
                 title: "Steam".into(),
                 compositor: Some(crate::vdisplay::Compositor::Gamescope),
                 cmd: Some("steam -gamepadui".into()),
+                library_id: None,
             });
         }
         if which("vkcube") {
@@ -88,23 +104,79 @@ pub fn catalog() -> Vec<AppEntry> {
                 title: "vkcube (test)".into(),
                 compositor: Some(crate::vdisplay::Compositor::Gamescope),
                 cmd: Some("vkcube".into()),
+                library_id: None,
             });
         }
     }
     apps
 }
 
+/// The high half of the positive `i32` range — where library-derived GameStream ids live, kept clear of
+/// the small Desktop/apps.json ids so the two never collide.
+const LIBRARY_ID_BASE: u32 = 0x4000_0000;
+
+/// Append the host's installed game library ([`crate::library::all_games`] — Steam/Epic/GOG/Xbox/custom)
+/// to `apps`. Each title gets a STABLE GameStream `<ID>` derived from its store-qualified library id
+/// (Moonlight caches appids, so a title keeps its id across host restarts), carries that library id so
+/// the launch path resolves it against the host's own library, and is de-duplicated (by id) against the
+/// base catalog and the other library entries. Titles with no launch recipe are skipped (un-startable).
+fn append_library(apps: &mut Vec<AppEntry>) {
+    let mut used: std::collections::HashSet<u32> = apps.iter().map(|a| a.id).collect();
+    for g in crate::library::all_games() {
+        if g.launch.is_none() {
+            continue;
+        }
+        let mut id = stable_app_id(&g.id);
+        // Linear-probe within the library range on the (rare) hash collision — deterministic given the
+        // stable all_games() order, so a title keeps its id run to run.
+        while !used.insert(id) {
+            id = LIBRARY_ID_BASE | (id.wrapping_add(1) & 0x3FFF_FFFF);
+        }
+        apps.push(AppEntry {
+            id,
+            title: g.title,
+            compositor: None, // auto-detect the desktop session (Windows ignores the compositor)
+            cmd: None,
+            library_id: Some(g.id),
+        });
+    }
+}
+
+/// A STABLE GameStream `<ID>` for a store-qualified library id (`steam:570`): FNV-1a-32 folded into the
+/// high half of the positive `i32` range ([`LIBRARY_ID_BASE`]). Deterministic across runs and clear of
+/// the reserved small Desktop/apps.json ids.
+fn stable_app_id(library_id: &str) -> u32 {
+    let mut h: u32 = 0x811c_9dc5;
+    for b in library_id.bytes() {
+        h ^= b as u32;
+        h = h.wrapping_mul(0x0100_0193);
+    }
+    LIBRARY_ID_BASE | (h & 0x3FFF_FFFF)
+}
+
 pub fn by_id(id: u32) -> Option<AppEntry> {
     catalog().into_iter().find(|a| a.id == id)
 }
 
-/// Render the GameStream `/applist` XML.
+/// Box-art bytes for the GameStream `/appasset` cover proxy: resolve the Moonlight appid to its catalog
+/// entry, then (for a library title) fetch its cover from the host's library. `(bytes, content-type)`,
+/// or `None` for Desktop / apps.json entries (no art) or a fetch failure. Blocking (disk + network) —
+/// call off the async runtime.
+pub fn appasset_bytes(appid: u32) -> Option<(Vec<u8>, String)> {
+    let lib_id = by_id(appid)?.library_id?;
+    crate::library::fetch_box_art(&lib_id)
+}
+
+/// Render the GameStream `/applist` XML. `IsHdrSupported` reflects whether the host can actually deliver
+/// HDR (HEVC Main10 / PQ) for a title — host-wide today ([`crate::gamestream::host_hdr_capable`]); when
+/// true, Moonlight offers its per-app HDR toggle.
 pub fn applist_xml() -> String {
+    let hdr = u8::from(crate::gamestream::host_hdr_capable());
     let mut xml =
         String::from("<?xml version=\"1.0\" encoding=\"utf-8\"?>\n<root status_code=\"200\">\n");
     for app in catalog() {
         xml.push_str(&format!(
-            "<App>\n<IsHdrSupported>0</IsHdrSupported>\n<AppTitle>{}</AppTitle>\n<ID>{}</ID>\n</App>\n",
+            "<App>\n<IsHdrSupported>{hdr}</IsHdrSupported>\n<AppTitle>{}</AppTitle>\n<ID>{}</ID>\n</App>\n",
             xml_escape(&app.title),
             app.id
         ));
@@ -130,10 +202,46 @@ mod tests {
 
     #[test]
     fn default_catalog_has_desktop() {
+        // catalog() = base (Desktop + apps.json) + the installed library; Desktop (id 1) is always present.
         let apps = catalog();
         assert!(apps.iter().any(|a| a.id == 1 && a.title == "Desktop"));
     }
 
+    #[test]
+    fn stable_app_id_is_deterministic_and_in_library_range() {
+        // Same id every run (Moonlight caches appids), distinct per title, and always in the high
+        // half of the positive i32 range so it never collides with the small Desktop/apps.json ids.
+        let a = stable_app_id("steam:570");
+        let b = stable_app_id("steam:570");
+        let c = stable_app_id("steam:271590");
+        assert_eq!(a, b);
+        assert_ne!(a, c);
+        for id in [a, c] {
+            assert!(id >= LIBRARY_ID_BASE, "id {id:#x} below library base");
+            assert!(id <= 0x7FFF_FFFF, "id {id:#x} not a positive i32");
+            assert_ne!(id, 1, "must not collide with Desktop");
+        }
+    }
+
+    #[test]
+    fn append_library_dedups_against_base_ids() {
+        // A base app whose id happens to fall in the library range must not be clobbered by a library
+        // entry that hashes to it — append_library probes past any used id.
+        let mut apps = vec![AppEntry {
+            id: stable_app_id("steam:570"),
+            title: "Pinned".into(),
+            compositor: None,
+            cmd: None,
+            library_id: None,
+        }];
+        append_library(&mut apps);
+        let ids: Vec<u32> = apps.iter().map(|a| a.id).collect();
+        let mut uniq = ids.clone();
+        uniq.sort_unstable();
+        uniq.dedup();
+        assert_eq!(ids.len(), uniq.len(), "duplicate GameStream ids in catalog");
+    }
+
     #[test]
     fn applist_xml_is_wellformed_ish() {
         let xml = applist_xml();

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/mod.rs

@@ -48,13 +48,26 @@ pub const SCM_HEVC: u32 = 0x0000_0100;
 pub const SCM_HEVC_MAIN10: u32 = 0x0000_0200;
 pub const SCM_AV1_MAIN8: u32 = 0x0001_0000;
 pub const SCM_AV1_MAIN10: u32 = 0x0002_0000;
-/// What we actually encode via NVENC: H.264, HEVC Main, AV1 Main 8-bit (= 65793). The
-/// 10-bit flags are deliberately NOT advertised: Moonlight only selects Main10 profiles for
-/// HDR streaming, and our capture path is 8-bit SDR BGRx with no HDR metadata plumbing —
-/// advertising them would let clients enable an HDR mode we can't deliver. (The previous
-/// placeholder 3843 = 0xF03 wrongly claimed HEVC Main10 + 4:4:4 and *no* AV1.)
+/// The **SDR baseline** codec mask: H.264, HEVC Main, AV1 Main 8-bit (= 65793). HEVC Main10 (HDR) is
+/// layered on top of this at runtime by `serverinfo::codec_mode_support` when — and only when — the
+/// host can actually deliver it ([`host_hdr_capable`]); it is never a static claim, because a non-HDR
+/// host (Linux, or a Windows host without the `PUNKTFUNK_10BIT` opt-in) must not invite a client into
+/// an HDR mode it can't produce. (The previous placeholder 3843 = 0xF03 wrongly claimed HEVC Main10 +
+/// 4:4:4 and *no* AV1.) 4:4:4 stays off entirely: stock Moonlight is 4:2:0 and the Windows IDD-push
+/// capturer can't yet deliver full-chroma frames (`crate::capture::capturer_supports_444`).
 pub const SERVER_CODEC_MODE_SUPPORT: u32 = SCM_H264 | SCM_HEVC | SCM_AV1_MAIN8;
 
+/// Whether this host can deliver an **HDR** (HEVC Main10 / BT.2020 PQ) GameStream — the single gate
+/// for advertising [`SCM_HEVC_MAIN10`] in serverinfo and `IsHdrSupported` per app, and for honoring a
+/// client's `dynamicRangeMode` request. HDR capture+encode is **Windows-only** (the Linux host is
+/// 8-bit, blocked upstream) and behind the operator's `PUNKTFUNK_10BIT` opt-in — the same policy gate
+/// the native punktfunk/1 plane honors. When this is true the IDD-push capturer streams HEVC Main10 PQ
+/// whenever the desktop is HDR, and a client HDR request makes the GameStream video path proactively
+/// enable advanced color on the per-session virtual display so PQ flows even from an SDR desktop.
+pub fn host_hdr_capable() -> bool {
+    cfg!(target_os = "windows") && crate::config::config().ten_bit
+}
+
 /// Stable host identity + advertised capabilities, shared across control-plane handlers.
 pub struct Host {
     pub hostname: String,
@@ -225,7 +238,7 @@ pub fn serve(
             tokio::try_join!(
                 nvhttp::run(state.clone()),
                 crate::mgmt::run(state.clone(), mgmt, Some(np.clone()), stats.clone()),
-                crate::punktfunk1::serve(native_opts, np, stats.clone()),
+                crate::punktfunk1::serve(native_opts, native.mgmt_port, np, stats.clone()),
             )?;
         } else {
             // Secure default: native punktfunk/1 + management API only (no GameStream surface).
@@ -236,7 +249,7 @@ pub fn serve(
             );
             tokio::try_join!(
                 crate::mgmt::run(state.clone(), mgmt, Some(np.clone()), stats.clone()),
-                crate::punktfunk1::serve(native_opts, np, stats.clone()),
+                crate::punktfunk1::serve(native_opts, native.mgmt_port, np, stats.clone()),
             )?;
         }
         Ok(())

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

@@ -13,8 +13,8 @@ use super::{serverinfo, AppState, LaunchSession, HTTPS_PORT, HTTP_PORT, RTSP_POR
 use anyhow::{anyhow, Context, Result};
 use axum::{
     extract::{Query, State},
-    http::header,
-    response::IntoResponse,
+    http::{header, StatusCode},
+    response::{IntoResponse, Response},
     routing::get,
     Extension, Router,
 };
@@ -64,6 +64,7 @@ fn router(state: Arc<AppState>, https: bool) -> Router {
         .route("/serverinfo", get(h_serverinfo))
         .route("/pair", get(h_pair))
         .route("/applist", get(h_applist))
+        .route("/appasset", get(h_appasset))
         .route("/launch", get(h_launch))
         .route("/resume", get(h_resume))
         .route("/cancel", get(h_cancel))
@@ -94,10 +95,32 @@ async fn h_applist(
         tracing::warn!("applist rejected — client is not paired");
         return xml(error_xml());
     }
-    // One app for now: the headless desktop (the wlroots virtual output).
     xml(super::apps::applist_xml())
 }
 
+/// Box-art cover proxy (`/appasset?appid=N&AssetType=2&AssetIdx=0`). Moonlight fetches per-app covers
+/// from the HOST, so we resolve the appid to its library title and proxy the cover image bytes (Steam/
+/// Epic CDN, etc.). 404 for Desktop / apps.json entries (no art) or any fetch failure — Moonlight then
+/// shows its title-only placeholder. Paired clients only (same gate as `/applist`). The resolve+fetch is
+/// blocking (disk + network), so it runs on a blocking thread off the async runtime.
+async fn h_appasset(
+    State(st): State<Arc<AppState>>,
+    peer: Option<Extension<PeerCertFingerprint>>,
+    Query(q): Query<HashMap<String, String>>,
+) -> Response {
+    if !peer_is_paired(&peer, &st) {
+        tracing::warn!("appasset rejected — client is not paired");
+        return StatusCode::FORBIDDEN.into_response();
+    }
+    let Some(appid) = q.get("appid").and_then(|s| s.parse::<u32>().ok()) else {
+        return StatusCode::BAD_REQUEST.into_response();
+    };
+    match tokio::task::spawn_blocking(move || super::apps::appasset_bytes(appid)).await {
+        Ok(Some((bytes, ctype))) => ([(header::CONTENT_TYPE, ctype)], bytes).into_response(),
+        _ => StatusCode::NOT_FOUND.into_response(),
+    }
+}
+
 async fn h_launch(
     State(st): State<Arc<AppState>>,
     peer: Option<Extension<PeerCertFingerprint>>,

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/rtsp.rs

@@ -350,19 +350,34 @@ fn stream_config(map: &HashMap<String, String>) -> Option<StreamConfig> {
     let fps = parse_u("x-nv-video[0].maxFPS")
         .filter(|&f| f > 0)
         .unwrap_or(60);
-    let bitrate_kbps = parse_u("x-nv-vqos[0].bw.maximumBitrateKbps").unwrap_or(20_000);
+    // Bitrate: Moonlight caps the legacy `x-nv-vqos[0].bw.*` fields at 100 Mbps for old-GFE
+    // compatibility and carries the user's REAL (uncapped) configured bitrate in the moonlight-specific
+    // `x-ml-video.configuredBitrateKbps`. Read that first — exactly like Sunshine — so a 500 Mbps client
+    // setting isn't silently floored to 100. Fall back to the legacy max for clients that don't send it,
+    // then a conservative default; clamp to a sane ceiling (the RTSP ANNOUNCE is attacker-controlled).
+    const MAX_BITRATE_KBPS: u32 = 1_000_000; // 1 Gbps — well above Moonlight's 500 Mbps slider
+    let bitrate_kbps = parse_u("x-ml-video.configuredBitrateKbps")
+        .filter(|&b| b > 0)
+        .or_else(|| parse_u("x-nv-vqos[0].bw.maximumBitrateKbps").filter(|&b| b > 0))
+        .unwrap_or(20_000)
+        .min(MAX_BITRATE_KBPS);
     // Client codec choice (moonlight-common-c SdpGenerator.c): 0=H264, 1=HEVC, 2=AV1.
     let codec = match map.get("x-nv-vqos[0].bitStreamFormat").map(|s| s.trim()) {
         Some("1") => Codec::H265,
         Some("2") => Codec::Av1,
         _ => Codec::H264,
     };
-    // 10-bit/HDR request flag. We never advertise the Main10 SCM bits, so a compliant
-    // client can't ask — if one does anyway, stream 8-bit SDR rather than failing.
-    if parse_u("x-nv-video[0].dynamicRangeMode").unwrap_or(0) != 0 {
+    // 10-bit/HDR request (Moonlight sets `dynamicRangeMode != 0` only when it both saw our Main10 SCM
+    // bit AND the user enabled HDR). Honor it only when the host can actually deliver Main10 (Windows +
+    // PUNKTFUNK_10BIT, `host_hdr_capable`); when honored, the video path proactively enables advanced
+    // color on the virtual display so a PQ stream flows even from an SDR desktop. A request we can't
+    // honor degrades to 8-bit SDR (and a desktop that is ALREADY HDR still streams PQ regardless, since
+    // the IDD-push capturer follows the display).
+    let hdr_requested = parse_u("x-nv-video[0].dynamicRangeMode").unwrap_or(0) != 0;
+    let hdr = hdr_requested && crate::gamestream::host_hdr_capable();
+    if hdr_requested && !hdr {
         tracing::warn!(
-            "client requested HDR/10-bit (dynamicRangeMode != 0) — not advertised/supported, \
-             streaming 8-bit SDR"
+            "client requested HDR (dynamicRangeMode != 0) but host is not HDR-capable — streaming 8-bit SDR"
         );
     }
     // Parity floor the client asks for (protects small frames); clamp to a sane max.
@@ -377,6 +392,7 @@ fn stream_config(map: &HashMap<String, String>) -> Option<StreamConfig> {
         bitrate_kbps,
         codec,
         min_fec,
+        hdr,
     })
 }
 
@@ -490,6 +506,26 @@ mod tests {
         }
     }
 
+    /// Bitrate precedence: the moonlight-specific `x-ml-video.configuredBitrateKbps` (the user's real,
+    /// uncapped setting) wins over the legacy `x-nv-vqos[0].bw.maximumBitrateKbps` (which Moonlight floors
+    /// at 100 Mbps for old-GFE compat). Without this a 500 Mbps client streamed at 100.
+    #[test]
+    fn announce_prefers_configured_bitrate() {
+        // Real Moonlight shape: legacy max floored at 100 Mbps, configured carrying the true 500 Mbps.
+        let map = announce(&[
+            ("x-nv-vqos[0].bw.maximumBitrateKbps", "100000"),
+            ("x-ml-video.configuredBitrateKbps", "500000"),
+        ]);
+        assert_eq!(stream_config(&map).unwrap().bitrate_kbps, 500_000);
+        // No configured field (older client) → fall back to the legacy max (the base announce's 40 Mbps).
+        assert_eq!(stream_config(&announce(&[])).unwrap().bitrate_kbps, 40_000);
+        // A zero configured value is ignored (falls back), and an absurd value is clamped to the ceiling.
+        let zero = announce(&[("x-ml-video.configuredBitrateKbps", "0")]);
+        assert_eq!(stream_config(&zero).unwrap().bitrate_kbps, 40_000);
+        let huge = announce(&[("x-ml-video.configuredBitrateKbps", "9000000")]);
+        assert_eq!(stream_config(&huge).unwrap().bitrate_kbps, 1_000_000);
+    }
+
     /// Missing required video keys → no config (the PLAY handler then refuses to stream).
     #[test]
     fn announce_missing_required_keys() {

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

@@ -43,11 +43,33 @@ pub fn serverinfo_xml(host: &Host, https: bool, paired: bool) -> String {
     )
 }
 
-/// The `<ServerCodecModeSupport>` mask to advertise. On the VAAPI (AMD/Intel) backend it reflects
-/// what the GPU can ACTUALLY encode (probed — AV1 is narrow, and an old iGPU might lack HEVC), so a
-/// Moonlight client never negotiates a codec the encoder can't open. NVENC and Windows keep the
-/// Moonlight-validated static superset.
+/// The `<ServerCodecModeSupport>` mask to advertise: the SDR baseline ([`base_codec_mode_support`]) plus
+/// the HEVC Main10 (HDR) bit when the host can actually deliver HDR ([`apply_hdr`] /
+/// [`crate::gamestream::host_hdr_capable`]). Without the Main10 bit Moonlight never offers its HDR
+/// toggle; with it, enabling HDR client-side negotiates Main10 and the IDD-push path streams BT.2020 PQ.
 fn codec_mode_support() -> u32 {
+    apply_hdr(
+        base_codec_mode_support(),
+        crate::gamestream::host_hdr_capable(),
+    )
+}
+
+/// Add the HEVC Main10 (HDR) bit to `base` when `hdr` and HEVC is advertised — pure so the
+/// HDR-layering is unit-testable without a GPU. (HDR streaming uses HEVC Main10; AV1 Main10 is left
+/// off until the GameStream AV1 path is live-confirmed.)
+fn apply_hdr(base: u32, hdr: bool) -> u32 {
+    if hdr && base & super::SCM_HEVC != 0 {
+        base | super::SCM_HEVC_MAIN10
+    } else {
+        base
+    }
+}
+
+/// The **SDR baseline** mask. On the VAAPI (AMD/Intel) backend it reflects what the GPU can ACTUALLY
+/// encode (probed — AV1 is narrow, and an old iGPU might lack HEVC), so a Moonlight client never
+/// negotiates a codec the encoder can't open. NVENC and the GPU-less software path keep the
+/// Moonlight-validated static superset. HDR (Main10) is layered on by [`codec_mode_support`].
+fn base_codec_mode_support() -> u32 {
     #[cfg(target_os = "linux")]
     if crate::encode::linux_zero_copy_is_vaapi() {
         if let Some(m) = probed_mask(crate::encode::vaapi_codec_support()) {
@@ -108,6 +130,22 @@ mod tests {
         );
     }
 
+    #[test]
+    fn apply_hdr_adds_main10_only_when_capable_and_hevc() {
+        // HDR-capable + HEVC advertised → Main10 added.
+        assert_eq!(
+            apply_hdr(SCM_H264 | SCM_HEVC | SCM_AV1_MAIN8, true),
+            SCM_H264 | SCM_HEVC | SCM_AV1_MAIN8 | SCM_HEVC_MAIN10
+        );
+        // Not HDR-capable → baseline unchanged (no HDR claim).
+        assert_eq!(
+            apply_hdr(SCM_H264 | SCM_HEVC | SCM_AV1_MAIN8, false),
+            SCM_H264 | SCM_HEVC | SCM_AV1_MAIN8
+        );
+        // HDR-capable but a GPU with no HEVC at all → no Main10 (you can't do Main10 without HEVC).
+        assert_eq!(apply_hdr(SCM_H264, true), SCM_H264);
+    }
+
     #[test]
     fn serverinfo_xml_carries_codec_mask() {
         let host = Host {

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

@@ -28,6 +28,10 @@ pub struct StreamConfig {
     pub codec: Codec,
     /// Client's `x-nv-vqos[0].fec.minRequiredFecPackets` — parity floor per FEC block.
     pub min_fec: u8,
+    /// Client requested HDR (`dynamicRangeMode != 0`) AND the host can deliver it ([`host_hdr_capable`]).
+    /// Drives the capturer's proactive advanced-color enable; the encoder picks Main10 from the captured
+    /// (P010) frame format. Always `false` on a non-HDR host, so the SDR path is unchanged.
+    pub hdr: bool,
 }
 
 /// Slot for the persistent screen capturer, shared with the control plane and reused across
@@ -137,7 +141,15 @@ fn run(
         let launch_here = compositor != crate::vdisplay::Compositor::Gamescope;
         #[cfg(any(windows, target_os = "linux"))]
         if launch_here {
-            if let Some(cmd) = app
+            // A library title (Steam/Epic/GOG/Xbox/custom, surfaced in /applist) carries its
+            // store-qualified id — resolve + launch it against the host's OWN library (the client can
+            // only pick an existing title, never inject a command). An apps.json entry instead carries
+            // an operator-typed `cmd`. Library id wins when both are set.
+            if let Some(lib_id) = app.and_then(|a| a.library_id.as_deref()) {
+                if let Err(e) = crate::library::launch_gamestream_library(lib_id) {
+                    tracing::warn!(library_id = lib_id, error = %e, "gamestream: could not launch library title");
+                }
+            } else if let Some(cmd) = app
                 .and_then(|a| a.cmd.as_deref())
                 .filter(|c| !c.trim().is_empty())
             {
@@ -213,14 +225,26 @@ fn open_gs_virtual_source(
     let compositor = if let Some(c) = app.and_then(|a| a.compositor) {
         c
     } else {
-        let active = crate::vdisplay::detect_active_session();
-        crate::vdisplay::apply_session_env(&active);
-        let c = crate::vdisplay::compositor_for_kind(active.kind)
-            .map(Ok)
-            .unwrap_or_else(crate::vdisplay::detect)
-            .context("detect compositor")?;
-        crate::vdisplay::apply_input_env(c);
-        c
+        // 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)
+                .map(Ok)
+                .unwrap_or_else(crate::vdisplay::detect)
+                .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
@@ -233,11 +257,13 @@ fn open_gs_virtual_source(
             refresh_hz: cfg.fps,
         })
         .context("create virtual output at client resolution")?;
-    // want_hdr=false: GameStream HDR is not negotiated into StreamConfig here (the default WGC backend
-    // still auto-detects HDR from the output colorspace; only the opt-in IDD-push path streams SDR).
+    // HDR: pass the negotiated `cfg.hdr` (client asked for HDR AND the host can deliver it). On the
+    // Windows IDD-push path this proactively enables advanced color on the virtual display so a Main10
+    // PQ stream flows even from an SDR desktop; an already-HDR desktop streams PQ regardless (the
+    // capturer follows the display). No-op on Linux (8-bit, and `cfg.hdr` is always false there).
     let capturer = capture::capture_virtual_output(
         vout,
-        capture::OutputFormat::resolve(false),
+        capture::OutputFormat::resolve(cfg.hdr),
         crate::session_plan::CaptureBackend::resolve(),
     )
     .context("capture virtual output")?;
@@ -245,6 +271,19 @@ fn open_gs_virtual_source(
     Ok((capturer, compositor))
 }
 
+/// The encoder bit depth implied by the captured frame's pixel format: a 10-bit (HDR) source — the
+/// Windows IDD-push capturer's `P010`/`Rgb10a2` when the desktop is HDR — opens NVENC as HEVC Main10
+/// (BT.2020 PQ); everything else is 8-bit. The encoder backends already key the real profile off the
+/// `format`, so this just keeps the `bit_depth` argument honest (the old hard-coded `8` mislabeled an
+/// HDR stream that the format had already promoted to 10-bit).
+fn gs_bit_depth(format: crate::capture::PixelFormat) -> u8 {
+    use crate::capture::PixelFormat;
+    match format {
+        PixelFormat::P010 | PixelFormat::Rgb10a2 => 10,
+        _ => 8,
+    }
+}
+
 /// One frame's packets, handed from the encode thread to the send thread.
 type PacketBatch = Vec<Vec<u8>>;
 
@@ -430,9 +469,10 @@ fn stream_body(
         cfg.fps,
         cfg.bitrate_kbps as u64 * 1000,
         frame.is_cuda(),
-        8, // GameStream/Moonlight path: 8-bit (its own codec negotiation)
+        // 8-bit SDR, or 10-bit when the captured frame is HDR (P010) — see `gs_bit_depth`.
+        gs_bit_depth(frame.format),
         // GameStream/Moonlight stays 4:2:0 — stock Moonlight clients can't decode 4:4:4, and the
-        // protocol has no chroma negotiation. 4:4:4 is punktfunk/1-native only.
+        // Windows IDD-push capturer can't yet deliver full-chroma frames. 4:4:4 is punktfunk/1-native only.
         encode::ChromaFormat::Yuv420,
     )
     .context("open video encoder for stream")?;
@@ -562,7 +602,7 @@ fn stream_body(
                     cfg.fps,
                     cfg.bitrate_kbps as u64 * 1000,
                     frame.is_cuda(),
-                    8,
+                    gs_bit_depth(frame.format),
                     encode::ChromaFormat::Yuv420, // GameStream stays 4:2:0
                 )
                 .context("reopen encoder after rebuild")?;

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
-//! on a headless box (single per-session virtual output at the origin, scale 1) equals the streamed
-//! output's pixels.
+//! virtual-keyboard path), and absolute pointer/touch coordinates are global compositor space.
+//!
+//! 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 {
-            name,
-            interface,
-            version,
-        } = event
-        {
-            if interface == "org_kde_kwin_fake_input" {
-                state.fake = Some(registry.bind(name, version.min(MAX_VERSION), qh, ()));
+        match event {
+            wl_registry::Event::Global {
+                name,
+                interface,
+                version,
+            } => match interface.as_str() {
+                "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)");
-        Ok(Self {
+        // Settle output geometry (wl_output + xdg-output were bound during the registry roundtrip
+        // 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 h = event.flags & 0xffff;
+                let w = ((event.flags >> 16) & 0xffff) as i32;
+                let h = (event.flags & 0xffff) as i32;
                 if w > 0 && h > 0 {
-                    let x = event.x.clamp(0, w as i32) as f64;
-                    let y = event.y.clamp(0, h as i32) as f64;
-                    self.fake.pointer_motion_absolute(x, y);
+                    self.refresh_geometry();
+                    let (lx, ly, lw, lh) = self.logical_target(w, h);
+                    // 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 h = event.flags & 0xffff;
+                let w = ((event.flags >> 16) & 0xffff) as i32;
+                let h = (event.flags & 0xffff) as i32;
                 if w > 0 && h > 0 {
-                    let x = event.x.clamp(0, w as i32) as f64;
-                    let y = event.y.clamp(0, h as i32) as f64;
+                    self.refresh_geometry();
+                    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 {