feat(audio): end-to-end 5.1/7.1 surround across the native path + all clients

Adds negotiated 5.1/7.1 surround to the punktfunk/1 protocol and every client
(previously stereo-only):

- core: new shared `audio` layout table (LAYOUT_51/71 + identity multistream
  mapping, canonical wire order FL FR FC LFE RL RR SL SR); Hello/Welcome
  `audio_channels` negotiation via the trailing-byte back-compat pattern (old
  peers fall back to stereo); C-ABI `punktfunk_connect_ex6`,
  `punktfunk_connection_audio_channels`, and in-core multistream decode
  `punktfunk_connection_next_audio_pcm` for embedders without a multistream
  Opus decoder. Real-libopus channel-identity round-trip test.
- host: native audio thread captures + Opus-(multi)stream-encodes at the
  negotiated count (with a cross-session cached-capturer channel-mismatch fix);
  GameStream surround unified onto the safe `opus::MSEncoder`, dropping
  `audiopus_sys` (~4 unsafe blocks) and un-gating Windows GameStream surround;
  WASAPI loopback capture relaxed to 2/6/8 with the correct dwChannelMask.
- clients: Linux (PipeWire), Windows (WASAPI), Android (AAudio) decode via
  `opus::MSDecoder` + render multichannel; Apple decodes in-core to PCM →
  AVAudioEngine with an explicit wire-order channel layout; each gains a
  Stereo/5.1/7.1 setting. `punktfunk-probe --audio-channels N` is the headless
  validator.

Verified on Linux: core/host/linux/probe test suites + the Android Rust
(cargo-ndk) build, clippy -D warnings, and rustfmt all green. Windows/Apple
builds, all on-glass checks, and the live native loopback are pending (CI / a
free box).

Also lands the concurrent in-tree HEVC 4:4:4 host work (PUNKTFUNK_444): it
shares the same touched files (quic.rs, punktfunk1.rs, encode/*, ...) and so
cannot be committed separately from the surround changes.

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

.gitea/workflows/android-screenshots.yml

@@ -0,0 +1,57 @@
+# Android client screenshots for the Play listing / marketing. Roborazzi renders the real Compose
+# UI with mock state on the host JVM via Robolectric — NO emulator, GPU, KVM, host, or JNI core
+# (`-PskipRustBuild` skips the cargo-ndk native build). The Android analogue of apple.yml's
+# `screenshots` job, gated to STABLE RELEASE tags only. Standalone + best-effort: a failure here
+# reds nothing else. PNGs land as a 30-day artifact; not committed or published.
+name: android-screenshots
+
+on:
+  push:
+    tags: ["v*"]
+  workflow_dispatch:
+
+jobs:
+  screenshots:
+    if: startsWith(github.ref, 'refs/tags/v') || github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-24.04
+    timeout-minutes: 45
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: JDK 21 (AGP 9.2 + Robolectric's SDK-36 android-all jar both want 17–21)
+        uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: "21"
+
+      - name: Android SDK
+        uses: android-actions/setup-android@v3
+
+      # No NDK/CMake — the screenshot unit tests are pure JVM. compileSdk 37 auto-downloads via AGP
+      # if the platform channel lacks it (same note as android.yml).
+      - name: platform-tools + platform 36 + build-tools
+        run: sdkmanager "platform-tools" "platforms;android-36" "build-tools;37.0.0"
+
+      - name: Cache (gradle)
+        uses: actions/cache@v4
+        with:
+          path: |
+            ~/.gradle/caches
+            ~/.gradle/wrapper
+          key: android-screenshots-${{ hashFiles('clients/android/**/*.gradle.kts') }}
+          restore-keys: android-screenshots-
+
+      # Roborazzi renders Compose on the JVM (Robolectric Native Graphics). `-PskipRustBuild` keeps
+      # the cargo-ndk native build out of the graph — the tests never load libpunktfunk_android.so.
+      - name: Capture screenshots (Roborazzi)
+        working-directory: clients/android
+        run: ./gradlew :app:testDebugUnitTest -PskipRustBuild --stacktrace
+
+      - name: Upload screenshots
+        if: always()
+        # v3: Gitea's API rejects upload-artifact@v4 (see apple.yml). Download is a zip.
+        uses: actions/upload-artifact@v3
+        with:
+          name: punktfunk-android-screenshots
+          path: clients/android/app/build/outputs/roborazzi
+          retention-days: 30

.gitea/workflows/decky.yml

@@ -11,12 +11,18 @@
 #   punktfunk.zip
 #     punktfunk/                  <- single top-level dir == plugin.json "name"
 #       plugin.json   [required]
-#       package.json  [required]
+#       package.json  [required; CI stamps "version" — Decky reads the installed version here]
 #       main.py       [required: python backend]
 #       dist/index.js [required: rollup output]
+#       update.json   [CI-baked {channel, manifest}: where the plugin's self-update check polls]
 #       README.md     (recommended)
 #       LICENSE       [required by the plugin store]
 #
+# SELF-UPDATE (no Decky store): alongside the zip we also publish a tiny per-channel
+# `manifest.json` ({version, artifact=<immutable per-version zip URL>, sha256}). The installed
+# plugin polls it (main.py check_update), and the frontend drives Decky's own install RPC to
+# apply a newer build. See clients/decky/README.md "Updating".
+#
 # REGISTRY_TOKEN: repo Actions secret, a PAT with write:package scope (shared with deb/rpm/docker).
 name: decky
 
@@ -56,20 +62,26 @@ jobs:
           pnpm install --frozen-lockfile
           pnpm run build          # rollup -> clients/decky/dist/index.js
 
-      - name: Version + channel
-        # Tag vX.Y.Z -> X.Y.Z (stable `latest/` alias + Gitea Release); main push -> 0.3.0-ciN.g<sha>
-        # (`canary/` alias). Used for the registry version path + the zip name (the plugin.json
-        # version is the source of truth Decky reads after install — bump it in the release commit).
+      - name: Version + channel + stamp
+        # Tag vX.Y.Z -> X.Y.Z (stable `latest/` alias + Gitea Release); main push -> 0.3.<run>
+        # (`canary/` alias). Decky reads a plugin's INSTALLED version from package.json (NOT
+        # plugin.json), and the plugin's own update check (clients/decky/main.py check_update)
+        # compares against it — so the build version is STAMPED into package.json here (mirrored
+        # into plugin.json for store parity). Canary is a PLAIN numeric semver, never a
+        # `-ci<N>` prerelease: compare-versions orders prerelease identifiers lexically
+        # (ci10 < ci9), which would break update detection; the run number is monotonic.
         working-directory: ${{ gitea.workspace }}
         run: |
-          SHORT=$(echo "$GITHUB_SHA" | cut -c1-8)
           case "$GITHUB_REF" in
             refs/tags/v*) V="${GITHUB_REF_NAME#v}"; ALIAS=latest ;;
-            *)            V="0.3.0-ci${GITHUB_RUN_NUMBER}.g${SHORT}"; ALIAS=canary ;;
+            *)            V="0.3.${GITHUB_RUN_NUMBER}"; ALIAS=canary ;;
           esac
+          BASE="https://$REGISTRY/api/packages/$OWNER/generic/$PACKAGE"
           echo "VERSION=$V" >> "$GITHUB_ENV"
           echo "ALIAS=$ALIAS" >> "$GITHUB_ENV"
+          echo "BASE=$BASE" >> "$GITHUB_ENV"
           echo "decky version $V -> alias '$ALIAS'"
+          VERSION="$V" node -e 'const fs=require("fs");for(const f of ["clients/decky/package.json","clients/decky/plugin.json"]){const j=JSON.parse(fs.readFileSync(f,"utf8"));j.version=process.env.VERSION;fs.writeFileSync(f,JSON.stringify(j,null,2)+"\n");}'
 
       - name: Assemble store-layout zip
         working-directory: ${{ gitea.workspace }}
@@ -89,9 +101,20 @@ jobs:
           chmod 0755 "$DEST/bin/punktfunkrun.sh"
           # Store requires a LICENSE in the plugin root; the project is MIT OR Apache-2.0.
           cp LICENSE-MIT                 "$DEST/LICENSE"
+          # Self-update channel pointer the backend reads (main.py check_update). It points at
+          # THIS channel's manifest.json (published below); that manifest in turn points at the
+          # immutable per-version zip, so its sha256 stays valid across future alias re-uploads.
+          printf '{"channel":"%s","manifest":"%s/%s/manifest.json"}\n' "$ALIAS" "$BASE" "$ALIAS" > "$DEST/update.json"
           ( cd "$STAGE" && zip -r "$RUNNER_TEMP/punktfunk.zip" "$PLUGIN" )
           ls -lh "$RUNNER_TEMP/punktfunk.zip"
           unzip -l "$RUNNER_TEMP/punktfunk.zip"
+          # The update manifest the plugin polls: the immutable per-version artifact + its
+          # sha256 (Decky's installer verifies the download against this hash, aborting on
+          # mismatch — so it MUST be the per-version URL, never the mutable alias).
+          SHA=$(sha256sum "$RUNNER_TEMP/punktfunk.zip" | cut -d' ' -f1)
+          printf '{"version":"%s","artifact":"%s/%s/punktfunk.zip","sha256":"%s"}\n' \
+            "$VERSION" "$BASE" "$VERSION" "$SHA" > "$RUNNER_TEMP/manifest.json"
+          cat "$RUNNER_TEMP/manifest.json"
 
       - name: Publish to the Gitea generic registry
         working-directory: ${{ gitea.workspace }}
@@ -99,18 +122,26 @@ jobs:
           TOKEN: ${{ secrets.REGISTRY_TOKEN }}
         run: |
           BASE="https://$REGISTRY/api/packages/$OWNER/generic/$PACKAGE"
-          # 1) Immutable, versioned URL.
+          # 1) Immutable, versioned URL + its update manifest (the manifest's `artifact` points
+          #    here, so the published sha256 keeps matching what Decky later downloads).
           curl -fsS --user "enricobuehler:$TOKEN" --upload-file "$RUNNER_TEMP/punktfunk.zip" \
             "$BASE/$VERSION/punktfunk.zip"
+          curl -fsS --user "enricobuehler:$TOKEN" --upload-file "$RUNNER_TEMP/manifest.json" \
+            "$BASE/$VERSION/manifest.json"
           echo "published $BASE/$VERSION/punktfunk.zip"
-          # 2) Channel alias (stable release -> latest/, canary main build -> canary/) — the link
-          #    to paste into Decky's "install from URL". The generic registry rejects re-uploading
-          #    an existing version/file (409), so delete the prior alias first (ignore 404 on run #1).
-          curl -fsS -o /dev/null --user "enricobuehler:$TOKEN" -X DELETE \
-            "$BASE/$ALIAS/punktfunk.zip" || true
+          # 2) Channel alias (stable release -> latest/, canary main build -> canary/) — the
+          #    zip is the "install from URL" link; manifest.json is what the installed plugin
+          #    polls for updates. The generic registry rejects re-uploading an existing
+          #    version/file (409), so delete the prior alias copies first (ignore 404 on run #1).
+          for f in punktfunk.zip manifest.json; do
+            curl -fsS -o /dev/null --user "enricobuehler:$TOKEN" -X DELETE "$BASE/$ALIAS/$f" || true
+          done
           curl -fsS --user "enricobuehler:$TOKEN" --upload-file "$RUNNER_TEMP/punktfunk.zip" \
             "$BASE/$ALIAS/punktfunk.zip"
+          curl -fsS --user "enricobuehler:$TOKEN" --upload-file "$RUNNER_TEMP/manifest.json" \
+            "$BASE/$ALIAS/manifest.json"
           echo "install-from-URL link: $BASE/$ALIAS/punktfunk.zip"
+          echo "update manifest:       $BASE/$ALIAS/manifest.json"
 
       - name: Attach zip to the Gitea release (stable tags only)
         if: startsWith(gitea.ref, 'refs/tags/v')

.gitea/workflows/linux-client-screenshots.yml

@@ -0,0 +1,67 @@
+# Native Linux client screenshots for the app/marketing listings. The client renders
+# host-free mock scenes (PUNKTFUNK_SHOT_SCENE) under a virtual X display; the driver
+# (clients/linux/tools/screenshots.sh) grabs each one — no host, GPU, or Wayland. The
+# Linux analogue of apple.yml's `screenshots` job, gated to STABLE RELEASE tags only.
+# Standalone + best-effort: a failure here reds nothing else. PNGs land as a 30-day
+# artifact; they are not committed or published.
+name: linux-client-screenshots
+
+on:
+  push:
+    tags: ["v*"]
+  workflow_dispatch:
+
+jobs:
+  screenshots:
+    if: startsWith(github.ref, 'refs/tags/v') || github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-24.04
+    # Same image as ci.yml/deb.yml — already carries the Rust toolchain + GTK/SDL build deps.
+    container:
+      image: git.unom.io/unom/punktfunk-rust-ci:latest
+    timeout-minutes: 90
+    steps:
+      - uses: actions/checkout@v4
+
+      # Client link deps (baked into the image; kept here so the job is green across image
+      # rebuilds — a no-op once present) PLUS the headless-render extras: a virtual X server,
+      # software GL+Vulkan (llvmpipe/lavapipe), the icon theme + fonts the UI draws with, and a
+      # root-window grab tool.
+      - name: Client link + headless-render deps
+        run: |
+          apt-get update
+          apt-get install -y --no-install-recommends \
+            libgtk-4-dev libadwaita-1-dev libsdl3-dev \
+            xvfb x11-utils imagemagick scrot \
+            libgl1-mesa-dri mesa-vulkan-drivers \
+            adwaita-icon-theme fonts-cantarell fonts-dejavu-core
+
+      # Reuse the workspace cargo caches (same keys as ci.yml/deb.yml).
+      - name: Cache keys
+        run: echo "rustc=$(rustc --version | cut -d' ' -f2)" >> "$GITHUB_ENV"
+      - uses: actions/cache@v4
+        with:
+          path: |
+            /usr/local/cargo/registry
+            /usr/local/cargo/git
+          key: cargo-home-${{ hashFiles('Cargo.lock') }}
+          restore-keys: cargo-home-
+      - uses: actions/cache@v4
+        with:
+          path: target
+          key: cargo-target-v3-${{ env.rustc }}-${{ hashFiles('Cargo.lock') }}
+          restore-keys: cargo-target-v3-${{ env.rustc }}-
+
+      - name: Build client
+        run: cargo build --release -p punktfunk-client-linux --locked
+
+      - name: Capture screenshots
+        run: bash clients/linux/tools/screenshots.sh
+
+      - name: Upload screenshots
+        if: always()
+        # v3: Gitea's API rejects upload-artifact@v4 (see apple.yml). Download is a zip.
+        uses: actions/upload-artifact@v3
+        with:
+          name: punktfunk-linux-client-screenshots
+          path: clients/linux/screenshots
+          retention-days: 30

.gitea/workflows/web-screenshots.yml

@@ -0,0 +1,53 @@
+# Management-console screenshots for the app/marketing listings. Captured from the
+# built Storybook with headless Chromium (web/tools/screenshots.mjs) — the page
+# stories render from fixtures, so no live mgmt API, login, or GPU is needed. This
+# is the web analogue of apple.yml's `screenshots` job, but gated to STABLE RELEASE
+# tags only (the console has no release workflow of its own — it ships inside the
+# host packaging). Best-effort: a standalone workflow, so a failure here reds
+# nothing else. PNGs land as a 30-day artifact; they are not committed or published.
+name: web-screenshots
+
+on:
+  push:
+    tags: ["v*"]
+  workflow_dispatch:
+
+jobs:
+  screenshots:
+    if: startsWith(github.ref, 'refs/tags/v') || github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-24.04
+    container:
+      image: oven/bun:1
+    timeout-minutes: 30
+    defaults:
+      run:
+        working-directory: web
+    steps:
+      # oven/bun ships neither git nor a real node (the driver runs under node), and
+      # the slim Debian base lacks a CA bundle — without it actions/checkout's HTTPS
+      # fetch dies with "Problem with the SSL CA cert" (same as ci.yml's web job).
+      - name: Install git + node + CA certs
+        working-directory: /
+        run: apt-get update && apt-get install -y --no-install-recommends ca-certificates git nodejs
+      - uses: actions/checkout@v4
+      # --ignore-scripts skips the prepare→codegen hook (mirrors ci.yml); run codegen
+      # explicitly since build-storybook has no prebuild hook of its own.
+      - name: Install dependencies
+        run: bun install --frozen-lockfile --ignore-scripts
+      - name: Generate API client + i18n messages
+        run: bun run codegen
+      # Pulls the matching Chromium build + the apt libs it needs (root in-container).
+      - name: Install Chromium
+        run: bunx playwright install --with-deps chromium
+      - name: Build Storybook
+        run: bun run build-storybook
+      - name: Capture screenshots
+        run: bun run screenshots
+      - name: Upload screenshots
+        if: always()
+        # v3: Gitea's API rejects upload-artifact@v4 (see apple.yml). Download is a zip.
+        uses: actions/upload-artifact@v3
+        with:
+          name: punktfunk-web-console-screenshots
+          path: web/screenshots
+          retention-days: 30

.gitignore

@@ -13,6 +13,7 @@ clients/apple/PunktfunkCore.xcframework/
 clients/apple/.swiftpm/
 # Generated App Store screenshots (tools/screenshots.sh output; uploaded as a CI artifact)
 clients/apple/screenshots/
+clients/linux/screenshots/
 # Xcode per-user state
 xcuserdata/
 

CLAUDE.md

@@ -346,7 +346,23 @@ FFI also link-needs `libGL`/`libgbm`/`libcuda` at build time). Env knobs: `PUNKT
 `PUNKTFUNK_COMPOSITOR=kwin|gamescope|mutter`, `PUNKTFUNK_ZEROCOPY=1`, `PUNKTFUNK_GAMESCOPE_APP=...`,
 `PUNKTFUNK_INPUT_BACKEND=...`, `PUNKTFUNK_PERF=1` (per-stage timing), `PUNKTFUNK_VIDEO_DROP=N` (FEC
 test), `PUNKTFUNK_FEC_PCT=N`, `PUNKTFUNK_DSCP=1` (opt-in DSCP/SO_PRIORITY media QoS on the data +
-GameStream video/audio sockets; no-op on the wire on Windows without a qWAVE policy).
+GameStream video/audio sockets; no-op on the wire on Windows without a qWAVE policy),
+`PUNKTFUNK_444=1` (full-chroma HEVC 4:4:4, see below).
+
+**HEVC 4:4:4 (full chroma, Range Extensions)**: opt-in via `PUNKTFUNK_444`, negotiated like 10-bit —
+the host emits 4:4:4 only when the client advertised `VIDEO_CAP_444` (wire bit `0x04` + ABI
+`PUNKTFUNK_VIDEO_CAP_444`), the codec is HEVC, the session is single-process, **and** a GPU probe
+(`encode::can_encode_444`, run before the Welcome) confirms support — else it resolves to 4:2:0 and
+`Welcome::chroma_format` reflects the real value (honest downgrade; the client reads it via
+`punktfunk_connection_chroma_format`). **punktfunk/1-native only** — GameStream/Moonlight stays 4:2:0
+(stock clients can't decode 4:4:4). **NVENC is the implemented path**: Linux `hevc_nvenc` feeds a
+swscale'd `yuv444p` (RGB-in is always 4:2:0 — verified on the RTX 5070 Ti — so the session forces CPU
+RGB capture for 4:4:4); Windows NVENC keeps ARGB input + FREXT profile + `chromaFormatIDC=3` and the
+DDA capturer delivers RGB. VAAPI / AMF / QSV **decline** (probe returns false — no validated 4:4:4
+hardware in the lab; they'd produce 4:2:0). Software (openh264) is 4:2:0-only. Test with
+`PUNKTFUNK_CLIENT_444=1 punktfunk-probe --out x.h265` then `ffprobe x.h265` (expect `pix_fmt yuv444p`).
+*Linux NVENC mechanism validated on the RTX 5070 Ti (ffmpeg CLI); Windows NVENC + 10-bit-4:4:4 not yet
+on-glass validated.*
 
 ## Conventions
 

Cargo.lock

@@ -2828,6 +2828,7 @@ dependencies = [
  "fec-rs",
  "hmac",
  "libc",
+ "opus",
  "proptest",
  "quinn",
  "rand 0.9.4",
@@ -2855,7 +2856,6 @@ dependencies = [
  "anyhow",
  "ash",
  "ashpd",
- "audiopus_sys",
  "axum",
  "axum-server",
  "base64",

clients/android/app/build.gradle.kts

@@ -62,6 +62,10 @@ android {
 
     buildFeatures { compose = true }
 
+    // Roborazzi/Robolectric render Compose on the host JVM (the CI screenshot harness) and need the
+    // merged Android resources + the app's manifest/theme available to the unit tests.
+    testOptions { unitTests { isIncludeAndroidResources = true } }
+
     compileOptions {
         sourceCompatibility = JavaVersion.VERSION_21
         targetCompatibility = JavaVersion.VERSION_21
@@ -99,4 +103,21 @@ dependencies {
     // Android TV components (we target phone + TV) land in the TV-UI milestone:
     //   implementation("androidx.tv:tv-material:1.1.0")
     // The manifest already declares leanback so the scaffold installs on TV.
+
+    // --- CI screenshot harness (Roborazzi on the JVM via Robolectric — no emulator/GPU). The
+    // screenshot tests render the real Compose UI with mock state; never load the JNI core, so the
+    // job runs `:app:testDebugUnitTest -PskipRustBuild` (see kit/build.gradle.kts). ---
+    testImplementation(composeBom)
+    testImplementation("androidx.compose.ui:ui-test-junit4")
+    debugImplementation("androidx.compose.ui:ui-test-manifest") // the ComponentActivity test host
+    testImplementation("junit:junit:4.13.2")
+    testImplementation("org.robolectric:robolectric:4.16.1")
+    testImplementation("io.github.takahirom.roborazzi:roborazzi:1.64.0")
+    testImplementation("io.github.takahirom.roborazzi:roborazzi-compose:1.64.0")
+}
+
+// Record (write) the screenshots when the unit tests run. These tests exist to GENERATE marketing
+// images, not to diff goldens, so always capture rather than verify.
+tasks.withType<Test>().configureEach {
+    systemProperty("roborazzi.test.record", "true")
 }

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

@@ -163,7 +163,7 @@ fun ConnectScreen(settings: Settings, onConnected: (Long) -> Unit) {
                     targetHost, targetPort, w, h, hz,
                     id.certPem, id.privateKeyPem, pinHex ?: "",
                     settings.bitrateKbps, settings.compositor, gamepadPref,
-                    hdrEnabled,
+                    hdrEnabled, settings.audioChannels,
                 )
             }
             connecting = false

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

@@ -16,9 +16,18 @@ data class Settings(
     val bitrateKbps: Int = 0,
     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
+     * can capture; the resolved count drives the decoder + AAudio layout. */
+    val audioChannels: Int = 2,
     val micEnabled: Boolean = false,
     /** Show the live stats overlay (FPS / throughput / latency) during a stream. */
     val statsHudEnabled: Boolean = true,
+    /**
+     * Touch input model. `true` (default) = trackpad: the cursor stays put on touch-down and moves
+     * by the finger's relative delta (swipe to nudge, lift and re-swipe to walk it across), tap to
+     * click where it is. `false` = direct pointing: the cursor jumps to the finger (the old behaviour).
+     */
+    val trackpadMode: Boolean = true,
 )
 
 /** Loads/saves [Settings] in the app-private `punktfunk_settings` prefs. */
@@ -33,8 +42,10 @@ class SettingsStore(context: Context) {
         bitrateKbps = prefs.getInt(K_BITRATE, 0),
         compositor = prefs.getInt(K_COMPOSITOR, 0),
         gamepad = prefs.getInt(K_GAMEPAD, 0),
+        audioChannels = prefs.getInt(K_AUDIO_CH, 2),
         micEnabled = prefs.getBoolean(K_MIC, false),
         statsHudEnabled = prefs.getBoolean(K_HUD, true),
+        trackpadMode = prefs.getBoolean(K_TRACKPAD, true),
     )
 
     fun save(s: Settings) {
@@ -45,8 +56,10 @@ class SettingsStore(context: Context) {
             .putInt(K_BITRATE, s.bitrateKbps)
             .putInt(K_COMPOSITOR, s.compositor)
             .putInt(K_GAMEPAD, s.gamepad)
+            .putInt(K_AUDIO_CH, s.audioChannels)
             .putBoolean(K_MIC, s.micEnabled)
             .putBoolean(K_HUD, s.statsHudEnabled)
+            .putBoolean(K_TRACKPAD, s.trackpadMode)
             .apply()
     }
 
@@ -57,8 +70,10 @@ class SettingsStore(context: Context) {
         const val K_BITRATE = "bitrate_kbps"
         const val K_COMPOSITOR = "compositor"
         const val K_GAMEPAD = "gamepad"
+        const val K_AUDIO_CH = "audio_channels"
         const val K_MIC = "mic_enabled"
         const val K_HUD = "stats_hud_enabled"
+        const val K_TRACKPAD = "trackpad_mode"
     }
 }
 
@@ -124,6 +139,13 @@ val REFRESH_OPTIONS = listOf(
     240 to "240 Hz",
 )
 
+/** (channel count, label). 2 = stereo (default), 6 = 5.1, 8 = 7.1. */
+val AUDIO_CHANNEL_OPTIONS = listOf(
+    2 to "Stereo",
+    6 to "5.1 Surround",
+    8 to "7.1 Surround",
+)
+
 /** (kbps, label). `0` = host default. */
 val BITRATE_OPTIONS = listOf(
     0 to "Automatic",

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

@@ -104,6 +104,12 @@ fun SettingsScreen(initial: Settings, onChange: (Settings) -> Unit, onBack: () -
         }
 
         SettingsGroup("Audio") {
+            SettingDropdown(
+                label = "Audio channels",
+                options = AUDIO_CHANNEL_OPTIONS,
+                selected = s.audioChannels,
+            ) { ch -> update(s.copy(audioChannels = ch)) }
+
             ToggleRow(
                 title = "Microphone",
                 subtitle = "Send your mic to the host's virtual microphone",
@@ -119,6 +125,16 @@ fun SettingsScreen(initial: Settings, onChange: (Settings) -> Unit, onBack: () -
             )
         }
 
+        SettingsGroup("Pointer") {
+            ToggleRow(
+                title = "Trackpad mode",
+                subtitle = "Relative cursor like a laptop touchpad — swipe to nudge, tap to click. " +
+                    "Off = the cursor jumps to your finger.",
+                checked = s.trackpadMode,
+                onCheckedChange = { on -> update(s.copy(trackpadMode = on)) },
+            )
+        }
+
         SettingsGroup("Overlay") {
             ToggleRow(
                 title = "Stats overlay",

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

@@ -41,6 +41,7 @@ import io.unom.punktfunk.kit.NativeBridge
 import java.util.concurrent.atomic.AtomicBoolean
 import kotlinx.coroutines.delay
 import kotlin.math.abs
+import kotlin.math.hypot
 import kotlin.math.roundToInt
 
 // Touch-gesture tuning (px / ms). TAP_SLOP: movement under this still counts as a tap, not a drag.
@@ -50,6 +51,15 @@ private const val TAP_SLOP = 12f
 private const val TAP_DRAG_MS = 250L
 private const val SCROLL_DIV = 4f
 
+// Trackpad-mode pointer ballistics (relative one-finger motion). POINTER_SENS: base finger-px →
+// host-px gain (~1:1, never twitchy). The rest is mild acceleration so a flick crosses the screen
+// while a slow drag stays precise: above ACCEL_SPEED_FLOOR px/ms the gain ramps by ACCEL_GAIN per
+// px/ms, capped at ACCEL_MAX (so a fast swipe can't fling the cursor uncontrollably).
+private const val POINTER_SENS = 1.3f
+private const val ACCEL_GAIN = 0.6f
+private const val ACCEL_SPEED_FLOOR = 0.3f
+private const val ACCEL_MAX = 3.0f
+
 @Composable
 fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
     val context = LocalContext.current
@@ -68,8 +78,11 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
     // Live decode stats for the HUD. Poll once a second for the whole stream (cheap, and each call
     // drains+resets the native window so it never grows unbounded even while the overlay is hidden);
     // `showStats` only gates rendering. A 3-finger tap toggles it live; the default comes from Settings.
+    val initialSettings = remember { SettingsStore(context).load() }
     var stats by remember { mutableStateOf<DoubleArray?>(null) }
-    var showStats by remember { mutableStateOf(SettingsStore(context).load().statsHudEnabled) }
+    var showStats by remember { mutableStateOf(initialSettings.statsHudEnabled) }
+    // Touch model is fixed per session (re-keys the gesture handler below if it ever changes).
+    val trackpad = initialSettings.trackpadMode
     LaunchedEffect(handle) {
         while (true) {
             delay(1000)
@@ -145,13 +158,18 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
         if (showStats) {
             stats?.let { StatsOverlay(it, Modifier.align(Alignment.TopStart).padding(12.dp)) }
         }
-        // Touch → mouse, absolute "direct pointing" like the Apple client: the host cursor follows
-        // your finger (MouseMoveAbs, host-normalized against the overlay size — which fills the video,
-        // so finger position maps straight onto the remote screen). Gestures: tap = left click;
-        // two-finger tap = right click; two-finger drag = scroll; tap-then-press-and-drag = left-drag
-        // (text selection / moving windows); three-finger tap = toggle the stats HUD.
+        // Touch → mouse. Two models, chosen by the Trackpad-mode setting:
+        //  • trackpad (default): the cursor STAYS where it is on touch-down and moves by the finger's
+        //    relative delta (MouseMove) with mild pointer acceleration — swipe to nudge, lift and
+        //    re-swipe to walk it across, tap to click where it is. This is what makes the cursor
+        //    reachable on a small screen.
+        //  • direct (opt-out): the cursor jumps to the finger and follows it (MouseMoveAbs,
+        //    host-normalized against the overlay size), the old "direct pointing" behaviour.
+        // Both share the same gesture vocabulary: tap = left click; two-finger tap = right click;
+        // two-finger drag = scroll; tap-then-press-and-drag = left-drag (text selection / moving
+        // windows); three-finger tap = toggle the stats HUD.
         Box(
-            Modifier.fillMaxSize().pointerInput(handle) {
+            Modifier.fillMaxSize().pointerInput(handle, trackpad) {
                 var lastTapUp = 0L
                 var lastTapX = 0f
                 var lastTapY = 0f
@@ -176,7 +194,9 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
                     val isDrag = down.uptimeMillis - lastTapUp < TAP_DRAG_MS &&
                         abs(startX - lastTapX) < TAP_SLOP && abs(startY - lastTapY) < TAP_SLOP
                     lastTapUp = 0L // consume the arming either way
-                    moveAbs(startX, startY) // cursor jumps to the finger immediately
+                    // Direct mode jumps the cursor to the finger; trackpad mode leaves it put (the
+                    // whole point — you nudge it with swipes instead).
+                    if (!trackpad) moveAbs(startX, startY)
                     if (isDrag) NativeBridge.nativeSendPointerButton(handle, 1, true)
 
                     var moved = false
@@ -185,6 +205,14 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
                     var prevCx = startX
                     var prevCy = startY
                     var upTime = down.uptimeMillis
+                    // Trackpad relative-motion state: the tracked finger, its last position/time, and
+                    // the sub-pixel remainder so a slow drag isn't lost to Int truncation.
+                    var trackId = down.id
+                    var prevX = startX
+                    var prevY = startY
+                    var prevT = down.uptimeMillis
+                    var accX = 0f
+                    var accY = 0f
 
                     while (true) {
                         val ev = awaitPointerEvent()
@@ -217,15 +245,46 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
                                 moved = true
                             }
                         } else if (!scrolling) {
-                            // One finger → the cursor follows it (skipped once a gesture turned into
-                            // a scroll, so dropping back to one finger doesn't jerk the cursor).
+                            // One finger (skipped once a gesture turned into a scroll, so dropping
+                            // back to one finger doesn't jerk the cursor).
                             val p = pressed.firstOrNull { it.id == down.id } ?: pressed.first()
                             if (abs(p.position.x - startX) > TAP_SLOP ||
                                 abs(p.position.y - startY) > TAP_SLOP
                             ) {
                                 moved = true
                             }
-                            moveAbs(p.position.x, p.position.y)
+                            if (trackpad) {
+                                // Relative: move by the finger delta × (sensitivity × acceleration),
+                                // carrying the sub-pixel remainder. Re-anchor (zero delta this frame)
+                                // if the tracked finger changed, so lifting one of several fingers
+                                // never jumps the cursor.
+                                if (p.id != trackId) {
+                                    trackId = p.id
+                                    prevX = p.position.x
+                                    prevY = p.position.y
+                                    prevT = p.uptimeMillis
+                                }
+                                val dx = p.position.x - prevX
+                                val dy = p.position.y - prevY
+                                val dt = (p.uptimeMillis - prevT).coerceAtLeast(1L)
+                                prevX = p.position.x
+                                prevY = p.position.y
+                                prevT = p.uptimeMillis
+                                val speed = hypot(dx, dy) / dt // finger px per ms
+                                val accel = (1f + ACCEL_GAIN * (speed - ACCEL_SPEED_FLOOR).coerceAtLeast(0f))
+                                    .coerceAtMost(ACCEL_MAX)
+                                accX += dx * POINTER_SENS * accel
+                                accY += dy * POINTER_SENS * accel
+                                val outX = accX.toInt() // truncates toward zero → remainder kept w/ sign
+                                val outY = accY.toInt()
+                                if (outX != 0 || outY != 0) {
+                                    NativeBridge.nativeSendPointerMove(handle, outX, outY)
+                                    accX -= outX
+                                    accY -= outY
+                                }
+                            } else {
+                                moveAbs(p.position.x, p.position.y) // direct: cursor follows the finger
+                            }
                         }
                         ev.changes.forEach { it.consume() }
                     }
@@ -239,7 +298,7 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
                                 NativeBridge.nativeSendPointerButton(handle, 3, true)
                                 NativeBridge.nativeSendPointerButton(handle, 3, false)
                             }
-                            else -> { // tap → left click, and arm tap-and-drag
+                            else -> { // tap → left click (at the cursor's current spot), arm tap-drag
                                 NativeBridge.nativeSendPointerButton(handle, 1, true)
                                 NativeBridge.nativeSendPointerButton(handle, 1, false)
                                 lastTapUp = upTime
@@ -260,7 +319,7 @@ fun StreamScreen(handle: Long, micEnabled: Boolean, onDisconnect: () -> Unit) {
  * `[fps, mbps, latP50Ms, latP95Ms, latValid, skew, w, h, hz, dropped]`.
  */
 @Composable
-private fun StatsOverlay(s: DoubleArray, modifier: Modifier = Modifier) {
+internal fun StatsOverlay(s: DoubleArray, modifier: Modifier = Modifier) {
     if (s.size < 10) return
     val w = s[6].toInt()
     val h = s[7].toInt()

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

@@ -10,7 +10,9 @@ import androidx.compose.ui.platform.LocalContext
 
 // punktfunk brand violets (from the app icon: #6C5BF3 / #A79FF8 / #D2C9FB on a #16132A indigo).
 // Used as the fallback dark scheme on pre-Android-12 devices; on 12+ we defer to Material You.
-private val BrandDark = darkColorScheme(
+// `internal` (not private) so the CI screenshot tests can force the deterministic brand palette —
+// Material You dynamic colour has no wallpaper to seed from under the Robolectric JVM renderer.
+internal val BrandDark = darkColorScheme(
     primary = Color(0xFFA79FF8),
     onPrimary = Color(0xFF1B1442),
     primaryContainer = Color(0xFF4C3FB3),

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

@@ -0,0 +1,74 @@
+package io.unom.punktfunk.screenshots
+
+import androidx.activity.ComponentActivity
+import androidx.compose.ui.test.junit4.createAndroidComposeRule
+import androidx.compose.ui.test.onRoot
+import com.github.takahirom.roborazzi.captureRoboImage
+import com.github.takahirom.roborazzi.captureScreenRoboImage
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.robolectric.RobolectricTestRunner
+import org.robolectric.annotation.Config
+import org.robolectric.annotation.GraphicsMode
+
+/**
+ * App-store / marketing screenshots of the native Android client, rendered on the JVM by Roborazzi
+ * (Robolectric Native Graphics) — no emulator, GPU, host, or JNI core. The scenes (ShotScenes.kt)
+ * render the REAL Compose UI with mock state.
+ *
+ * `sdk = [36]` is mandatory: Robolectric ships android-all jars only up to API 36 (Android 16), and
+ * the app's compileSdk is 37. PNGs land in build/outputs/roborazzi/.
+ */
+@RunWith(RobolectricTestRunner::class)
+@GraphicsMode(GraphicsMode.Mode.NATIVE)
+@Config(sdk = [36], qualifiers = "w360dp-h800dp-xxhdpi")
+class ScreenshotTest {
+    @get:Rule
+    val compose = createAndroidComposeRule<ComponentActivity>()
+
+    private val out = "build/outputs/roborazzi"
+
+    // Pausing the animation clock before composing (then advancing once past the entrance animation
+    // and freezing) is what makes a text-field-bearing scene capturable: a focused field blinks its
+    // cursor via an infinite animation that otherwise keeps Compose perpetually "busy", so
+    // setContent's wait-for-idle never returns. Frozen, the capture is also deterministic.
+
+    /** Full-screen content scenes: the compose root fills the device, so a root capture is the shot. */
+    private fun shootRoot(name: String, content: @androidx.compose.runtime.Composable () -> Unit) {
+        compose.mainClock.autoAdvance = false
+        compose.setContent { ShotTheme(content) }
+        compose.mainClock.advanceTimeBy(800)
+        compose.onRoot().captureRoboImage("$out/phone-$name.png")
+    }
+
+    /** Dialog scenes: the AlertDialog is a separate window, so capture the whole screen (all windows). */
+    private fun shootScreen(name: String, content: @androidx.compose.runtime.Composable () -> Unit) {
+        compose.mainClock.autoAdvance = false
+        compose.setContent { ShotTheme(content) }
+        compose.mainClock.advanceTimeBy(800)
+        captureScreenRoboImage("$out/phone-$name.png")
+    }
+
+    @Test
+    fun hosts() = shootRoot("hosts") { HostsScene() }
+
+    @Test
+    fun settings() = shootRoot("settings") { SettingsScene() }
+
+    @Test
+    @Config(sdk = [36], qualifiers = "w800dp-h360dp-xxhdpi") // landscape — the stream is immersive
+    fun stream() = shootRoot("stream") { StreamScene() }
+
+    @Test
+    fun trust() = shootScreen("trust") {
+        HostsScene()
+        TrustDialog()
+    }
+
+    @Test
+    fun pair() = shootScreen("pair") {
+        HostsScene()
+        PairDialog()
+    }
+}

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

@@ -0,0 +1,195 @@
+package io.unom.punktfunk.screenshots
+
+import androidx.compose.foundation.background
+import androidx.compose.foundation.layout.Arrangement
+import androidx.compose.foundation.layout.Box
+import androidx.compose.foundation.layout.Column
+import androidx.compose.foundation.layout.Spacer
+import androidx.compose.foundation.layout.fillMaxSize
+import androidx.compose.foundation.layout.fillMaxWidth
+import androidx.compose.foundation.layout.height
+import androidx.compose.foundation.layout.padding
+import androidx.compose.foundation.lazy.grid.GridCells
+import androidx.compose.foundation.lazy.grid.GridItemSpan
+import androidx.compose.foundation.lazy.grid.LazyVerticalGrid
+import androidx.compose.foundation.lazy.grid.items
+import androidx.compose.material3.AlertDialog
+import androidx.compose.material3.MaterialTheme
+import androidx.compose.material3.Surface
+import androidx.compose.material3.Text
+import androidx.compose.material3.TextButton
+import androidx.compose.runtime.Composable
+import androidx.compose.ui.Alignment
+import androidx.compose.ui.Modifier
+import androidx.compose.ui.graphics.Brush
+import androidx.compose.ui.graphics.Color
+import androidx.compose.ui.text.style.TextAlign
+import androidx.compose.ui.unit.dp
+import io.unom.punktfunk.BrandDark
+import io.unom.punktfunk.Settings
+import io.unom.punktfunk.SettingsScreen
+import io.unom.punktfunk.StatsOverlay
+import io.unom.punktfunk.components.HostCard
+import io.unom.punktfunk.components.SectionLabel
+import io.unom.punktfunk.models.HostStatus
+
+// The CI screenshot scenes: the REAL app composables, fed embedded mock state, under the forced
+// brand palette (Material You has no wallpaper to seed from on the JVM). The stream-video surface
+// and ConnectScreen/App are intentionally absent — they require the live JNI core / a session.
+
+/** Forces the deterministic punktfunk brand scheme (see Theme.kt) instead of dynamic colour. */
+@Composable
+internal fun ShotTheme(content: @Composable () -> Unit) {
+    MaterialTheme(colorScheme = BrandDark, content = content)
+}
+
+private data class MockHost(val name: String, val address: String, val status: HostStatus)
+
+private val SAVED = listOf(
+    MockHost("Living Room PC", "192.168.1.42:9777", HostStatus.PAIRED),
+    MockHost("Office", "192.168.1.50:9777", HostStatus.TOFU),
+)
+private val DISCOVERED = listOf(
+    MockHost("studio-deck", "192.168.1.61:9777", HostStatus.PAIRING),
+    MockHost("HTPC", "192.168.1.70:9777", HostStatus.TOFU),
+)
+
+/** The connect screen's host grid, reconstructed from the real HostCard/SectionLabel components. */
+@Composable
+internal fun HostsScene() {
+    Surface(Modifier.fillMaxSize(), color = MaterialTheme.colorScheme.background) {
+        LazyVerticalGrid(
+            columns = GridCells.Adaptive(minSize = 160.dp),
+            modifier = Modifier.fillMaxSize(),
+            contentPadding = androidx.compose.foundation.layout.PaddingValues(16.dp),
+            horizontalArrangement = Arrangement.spacedBy(8.dp),
+            verticalArrangement = Arrangement.spacedBy(8.dp),
+        ) {
+            item(span = { GridItemSpan(maxLineSpan) }) {
+                Column(
+                    horizontalAlignment = Alignment.CenterHorizontally,
+                    modifier = Modifier.fillMaxWidth(),
+                ) {
+                    Spacer(Modifier.height(8.dp))
+                    Text("Punktfunk", style = MaterialTheme.typography.headlineLarge)
+                    Text(
+                        "stream a remote desktop",
+                        style = MaterialTheme.typography.bodyMedium,
+                        color = MaterialTheme.colorScheme.onSurfaceVariant,
+                    )
+                    Spacer(Modifier.height(24.dp))
+                }
+            }
+            item(span = { GridItemSpan(maxLineSpan) }) { SectionLabel("Saved hosts") }
+            items(SAVED) { h ->
+                HostCard(h.name, h.address, h.status, enabled = true, onConnect = {}, onForget = {}, onRename = {})
+            }
+            item(span = { GridItemSpan(maxLineSpan) }) {
+                Spacer(Modifier.height(12.dp))
+                SectionLabel("Discovered on the network")
+            }
+            items(DISCOVERED) { h ->
+                HostCard(h.name, h.address, h.status, enabled = true, onConnect = {}, onForget = null)
+            }
+        }
+    }
+}
+
+/** The real SettingsScreen, fed a representative non-default Settings. */
+@Composable
+internal fun SettingsScene() {
+    Surface(Modifier.fillMaxSize(), color = MaterialTheme.colorScheme.background) {
+        SettingsScreen(
+            initial = Settings(
+                width = 1920,
+                height = 1080,
+                hz = 120,
+                bitrateKbps = 50_000,
+                compositor = 1,
+                gamepad = 2,
+                micEnabled = true,
+                statsHudEnabled = true,
+                trackpadMode = true,
+            ),
+            onChange = {},
+            onBack = {},
+        )
+    }
+}
+
+/** The real TOFU AlertDialog (mirrors ConnectScreen's PendingTrust.Kind.TRUST_NEW), shown over the host grid. */
+@Composable
+internal fun TrustDialog() {
+    AlertDialog(
+        onDismissRequest = {},
+        title = { Text("Trust this host?") },
+        text = {
+            Column {
+                Text("First connection to 192.168.1.61:9777.")
+                Text("Fingerprint 9f8e7d6c5b4a3928…")
+                Text(
+                    "This host allows trust-on-first-use, but that can't tell an impostor " +
+                        "from the real host. Pairing with a PIN is stronger — it proves both sides.",
+                )
+            }
+        },
+        confirmButton = { TextButton({}) { Text("Trust (TOFU)") } },
+        dismissButton = { TextButton({}) { Text("Pair with PIN…") } },
+    )
+}
+
+/** The PIN-pairing AlertDialog (mirrors ConnectScreen's PendingTrust.Kind.PAIR). The live screen
+ *  uses OutlinedTextFields, but a TextField inside a Dialog window never reaches idle under
+ *  Robolectric (its focus/cursor machinery animates forever) — so the PIN is shown as a static
+ *  display here, which also reads better in a marketing shot. */
+@Composable
+internal fun PairDialog() {
+    AlertDialog(
+        onDismissRequest = {},
+        title = { Text("Pair with PIN") },
+        text = {
+            Column {
+                Text("Enter the 4-digit PIN shown on the host.")
+                Spacer(Modifier.height(16.dp))
+                Surface(
+                    color = MaterialTheme.colorScheme.surfaceVariant,
+                    shape = MaterialTheme.shapes.medium,
+                    modifier = Modifier.fillMaxWidth(),
+                ) {
+                    Text(
+                        "4  8  2  7",
+                        style = MaterialTheme.typography.headlineMedium,
+                        textAlign = TextAlign.Center,
+                        modifier = Modifier.fillMaxWidth().padding(vertical = 16.dp),
+                    )
+                }
+                Spacer(Modifier.height(12.dp))
+                Text(
+                    "This device: Pixel 9 Pro",
+                    style = MaterialTheme.typography.bodyMedium,
+                    color = MaterialTheme.colorScheme.onSurfaceVariant,
+                )
+            }
+        },
+        confirmButton = { TextButton({}) { Text("Pair") } },
+        dismissButton = { TextButton({}) { Text("Cancel") } },
+    )
+}
+
+/** The live stats HUD (the real StatsOverlay) over a synthetic "streamed frame" gradient. */
+@Composable
+internal fun StreamScene() {
+    Box(
+        Modifier
+            .fillMaxSize()
+            .background(
+                Brush.linearGradient(listOf(Color(0xFF2A1E5C), Color(0xFF0E1B3D), Color(0xFF06122B))),
+            ),
+    ) {
+        // [fps, mbps, latP50, latP95, latValid, skew, w, h, hz, dropped]
+        StatsOverlay(
+            doubleArrayOf(238.0, 921.4, 1.3, 2.1, 1.0, 1.0, 5120.0, 1440.0, 240.0, 0.0),
+            Modifier.align(Alignment.TopStart).padding(12.dp),
+        )
+    }
+}

clients/android/kit/build.gradle.kts

@@ -99,6 +99,12 @@ val cargoNdkDebug = registerCargoNdk("cargoNdkDebug", release = false)
 val cargoNdkRelease = registerCargoNdk("cargoNdkRelease", release = true)
 
 afterEvaluate {
-    tasks.named("preDebugBuild").configure { dependsOn(cargoNdkDebug) }
-    tasks.named("preReleaseBuild").configure { dependsOn(cargoNdkRelease) }
+    // `-PskipRustBuild` skips the cargo-ndk native build — for JVM-only tasks (the Roborazzi
+    // screenshot unit tests render Compose on the JVM and never load libpunktfunk_android.so), so
+    // CI/local screenshot runs don't need the Rust toolchain or NDK. The native build stays wired
+    // for every normal APK/AAR build.
+    if (!project.hasProperty("skipRustBuild")) {
+        tasks.named("preDebugBuild").configure { dependsOn(cargoNdkDebug) }
+        tasks.named("preReleaseBuild").configure { dependsOn(cargoNdkRelease) }
+    }
 }

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

@@ -45,6 +45,7 @@ object NativeBridge {
         compositorPref: Int,
         gamepadPref: Int,
         hdrEnabled: Boolean,
+        audioChannels: Int,
     ): Long
 
     /** 64-hex SHA-256 of the cert the host presented on [handle]; valid after a successful connect. */

clients/android/native/src/audio.rs

@@ -1,8 +1,21 @@
 //! Android audio playback (android-only): pull Opus packets from the connector, decode to
-//! interleaved f32 stereo, and feed AAudio (LowLatency) via its realtime data callback through a
-//! jitter ring. Mirrors [`crate::decode`]: one thread we own (the Opus decode producer) plus a
-//! shutdown flag; the realtime callback thread is owned by AAudio. Ring logic ported from
-//! `punktfunk-client-linux/src/audio.rs` (prime ~3 quanta, drop-oldest cap, re-prime on drain).
+//! interleaved f32 (stereo or 5.1/7.1 surround), and feed AAudio (LowLatency) via its realtime data
+//! callback through a jitter ring. Mirrors [`crate::decode`]: one thread we own (the Opus decode
+//! producer) plus a shutdown flag; the realtime callback thread is owned by AAudio.
+//!
+//! The layout is the host-RESOLVED channel count (`NativeClient::audio_channels`, negotiated at
+//! connect), so an older/clamping host that can only capture stereo is decoded + played as stereo.
+//! 2 = stereo / 6 = 5.1 / 8 = 7.1, in the canonical wire order FL FR FC LFE RL RR SL SR.
+//!
+//! The ring started as a port of `punktfunk-client-linux/src/audio.rs`, but AAudio — unlike
+//! PipeWire, which adaptively rate-matches the stream and absorbs a shallow buffer — hands us a raw
+//! realtime callback and makes us own the buffer. So this client diverges deliberately to stop the
+//! Android-only crackle: (1) the callback is allocation/free-free — decoded buffers are recycled to
+//! the producer via a free-list instead of being freed on the audio thread (Android's Scudo `free`
+//! has unbounded tail latency); (2) the jitter ring is deeper (~40 ms prime / ~150 ms hard cap) and
+//! decoupled from the tiny LowLatency burst size, with de-prime hysteresis so a transient drain
+//! doesn't manufacture a silence; (3) the AAudio HW buffer is primed above its 2-burst default and
+//! grown on XRuns (Google's anti-glitch technique).
 
 use ndk::audio::{
     AudioCallbackResult, AudioDirection, AudioFormat, AudioPerformanceMode, AudioSharingMode,
@@ -13,16 +26,75 @@ use punktfunk_core::error::PunktfunkError;
 use std::collections::VecDeque;
 use std::ffi::c_void;
 use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
-use std::sync::mpsc::{sync_channel, SyncSender, TrySendError};
+use std::sync::mpsc::{sync_channel, Receiver, SyncSender, TrySendError};
 use std::sync::Arc;
 use std::time::Duration;
 
-const CHANNELS: usize = 2;
 const SAMPLE_RATE: i32 = 48_000;
 /// Decoded-chunk hand-off depth: 64 × 5 ms = 320 ms slack (matches the core's AUDIO_QUEUE).
 const RING_CHUNKS: usize = 64;
-/// Opus decode scratch: worst-case 120 ms stereo frame (5760 samples/ch × 2 ch).
-const PCM_SCRATCH: usize = 5760 * CHANNELS;
+
+// --- Jitter-ring depths, in MILLISECONDS (scaled to interleaved-f32 samples at runtime). --------
+// The channel count is negotiated, not a compile-time const, so these are kept in ms and multiplied
+// by `ms` (interleaved-f32 samples per millisecond at the resolved layout) inside `start`.
+// Unlike the Linux client (PipeWire adaptively rate-matches the stream to the graph clock, masking
+// host↔DAC drift + a shallow ring), AAudio hands us a raw callback and we own the buffer: drift and
+// WiFi power-save bunching land as underruns/overflows = crackle. So Android runs a deliberately
+// deeper, smoothly-managed ring than Linux — keep the two clients' depths intentionally divergent.
+/// Prime/target floor: fill to ~40 ms before playing (and after a sustained drain). Deep enough to
+/// ride out WiFi arrival jitter + clock drift; the dominant Android-only anti-crackle lever.
+const PRIME_FLOOR_MS: usize = 40;
+/// Ceiling for the burst-scaled target (so a large quantum can't push the prime depth too high).
+const PRIME_CEIL_MS: usize = 80;
+/// Drop-oldest headroom above the target before trimming — a ~80 ms band swallows an arrival burst
+/// without overflowing.
+const JITTER_HEADROOM_MS: usize = 80;
+/// Hard latency bound: never let the ring exceed ~150 ms (the only thing that caps added latency).
+const HARD_CAP_MS: usize = 150;
+/// Re-prime (go silent to refill) only after this many CONSECUTIVE empty callbacks, so one transient
+/// drain doesn't manufacture a fresh 40 ms silence (the old `if ring.is_empty()` re-primed instantly).
+const DEPRIME_AFTER_CALLBACKS: u32 = 5;
+/// Throttle the AAudio XRun-driven HW-buffer grow check (cheap, but no need to poll every quantum).
+const XRUN_CHECK_EVERY: u32 = 128;
+
+/// Opus decoder for the audio plane: a plain stereo decoder (the validated path) or a multistream
+/// decoder for 5.1/7.1, both behind one `decode_float`. Built from the host-RESOLVED channel count
+/// via the shared layout table. Mirrors the Linux client's `AudioDec`.
+enum AudioDec {
+    Stereo(opus::Decoder),
+    Surround(opus::MSDecoder),
+}
+
+impl AudioDec {
+    fn new(channels: u8) -> Result<AudioDec, opus::Error> {
+        if channels == 2 {
+            Ok(AudioDec::Stereo(opus::Decoder::new(
+                SAMPLE_RATE as u32,
+                opus::Channels::Stereo,
+            )?))
+        } else {
+            let l = punktfunk_core::audio::layout_for(channels, false);
+            Ok(AudioDec::Surround(opus::MSDecoder::new(
+                SAMPLE_RATE as u32,
+                l.streams,
+                l.coupled,
+                l.mapping,
+            )?))
+        }
+    }
+
+    fn decode_float(
+        &mut self,
+        input: &[u8],
+        out: &mut [f32],
+        fec: bool,
+    ) -> Result<usize, opus::Error> {
+        match self {
+            AudioDec::Stereo(d) => d.decode_float(input, out, fec),
+            AudioDec::Surround(d) => d.decode_float(input, out, fec),
+        }
+    }
+}
 
 /// Diagnostics — written by the decode thread + the realtime callback, logged periodically. The
 /// audio analogue of the video `fed`/`rendered` counters (we can't "screenshot" sound).
@@ -42,27 +114,57 @@ pub struct AudioPlayback {
 }
 
 impl AudioPlayback {
-    /// Open AAudio (LowLatency, 48 kHz/stereo/f32) with a realtime callback draining a jitter ring,
-    /// then spawn the Opus decode thread. `None` on failure (the caller leaves video streaming).
+    /// Open AAudio (LowLatency, 48 kHz/f32, the host-resolved channel layout) with a realtime
+    /// callback draining a jitter ring, then spawn the Opus decode thread. `None` on failure (the
+    /// caller leaves video streaming).
     pub fn start(client: Arc<NativeClient>) -> Option<AudioPlayback> {
+        // Build playback from the host-RESOLVED channel count (never the request): 2 = stereo /
+        // 6 = 5.1 / 8 = 7.1, canonical wire order FL FR FC LFE RL RR SL SR.
+        let channels = punktfunk_core::audio::normalize_channels(client.audio_channels) as usize;
+        // Interleaved f32 samples per millisecond at this layout (48 kHz × channels); the ms-
+        // denominated jitter-ring depths scale by it.
+        let ms = (SAMPLE_RATE as usize / 1000) * channels;
+        let prime_floor = PRIME_FLOOR_MS * ms;
+        let prime_ceil = PRIME_CEIL_MS * ms;
+        let jitter_headroom = JITTER_HEADROOM_MS * ms;
+        let hard_cap_max = HARD_CAP_MS * ms;
         let counters = Arc::new(Counters::default());
         let (tx, rx) = sync_channel::<Vec<f32>>(RING_CHUNKS);
+        // Recycle free-list: drained PCM buffers go BACK to the decode thread to be refilled, so the
+        // realtime callback never frees heap (Android's Scudo allocator has unbounded free() tail
+        // latency — a free on the audio thread is an XRun = a click) and the decode thread rarely
+        // allocates. Same depth as the data channel.
+        let (free_tx, free_rx) = sync_channel::<Vec<f32>>(RING_CHUNKS);
 
         // Realtime consumer state, owned by the callback (FnMut) — no lock: AAudio calls it from a
-        // single high-priority thread, and the decode thread only touches `tx`.
+        // single high-priority thread, and the decode thread only touches `tx`/`free_rx`.
         let cb_counters = counters.clone();
-        let mut ring: VecDeque<f32> = VecDeque::with_capacity(PCM_SCRATCH);
+        // Pre-reserve the ring so `extend` never reallocates on the realtime thread. Worst transient
+        // before the trim below = the hard cap plus one full channel of 5 ms (480-f32) frames — the
+        // punktfunk protocol always sends 5 ms Opus frames (host `audio_thread`); a larger frame
+        // would force a one-time realloc, asserted (not silently corrupted) in `decode_loop`.
+        let mut ring: VecDeque<f32> = VecDeque::with_capacity(hard_cap_max + RING_CHUNKS * 5 * ms);
         let mut primed = false;
-        let callback = move |_s: &AudioStream, data: *mut c_void, num_frames: i32| {
-            let want = num_frames as usize * CHANNELS;
+        let mut empties: u32 = 0; // consecutive empty callbacks (de-prime hysteresis)
+        let mut cb_count: u32 = 0; // callbacks since open (throttles the XRun grow check)
+        let mut last_xrun: i32 = 0; // last AAudio XRun count we grew the buffer for
+        let callback = move |s: &AudioStream, data: *mut c_void, num_frames: i32| {
+            let want = num_frames as usize * channels;
             // SAFETY: AAudio provides `num_frames * channel_count` F32 slots at `data`.
             let out = unsafe { std::slice::from_raw_parts_mut(data as *mut f32, want) };
-            while let Ok(chunk) = rx.try_recv() {
-                ring.extend(chunk);
+            // Drain decoded chunks into the ring WITHOUT freeing on the RT thread: `drain(..)` empties
+            // each Vec but keeps its capacity, then the empty buffer is handed back for reuse. The
+            // only RT-thread free is the rare case where the recycle channel is momentarily full.
+            while let Ok(mut chunk) = rx.try_recv() {
+                ring.extend(chunk.drain(..));
+                let _ = free_tx.try_send(chunk);
             }
-            // Prime to ~3 quanta (15 ms; floor 15 ms / ceiling 200 ms); drop OLDEST above the cap.
-            let target = (3 * want).clamp(720 * CHANNELS, 9600 * CHANNELS);
-            while ring.len() > target.max(want) + want {
+            // Jitter buffer: prime to ~40 ms (prime_floor) before playing and after a sustained drain;
+            // drop-oldest only above a wide ~120 ms band. Decoupled from the AAudio burst `want` (tiny
+            // on the LowLatency MMAP path) so the depth doesn't collapse to a single quantum.
+            let target = (3 * want).clamp(prime_floor, prime_ceil);
+            let hard_cap = (target + jitter_headroom).min(hard_cap_max);
+            while ring.len() > hard_cap {
                 ring.pop_front();
             }
             if !primed && ring.len() >= target {
@@ -79,12 +181,34 @@ impl AudioPlayback {
                 out.fill(0.0);
                 cb_counters.underruns.fetch_add(1, Ordering::Relaxed);
             }
+            // Re-prime only after a RUN of empty callbacks, not a single transient one — otherwise
+            // every momentary drain costs a fresh 40 ms silence (the old behaviour, self-inflicted
+            // crackle on any jitter spike).
             if ring.is_empty() {
-                primed = false; // re-prime after a genuine drain (avoids sustained crackle on loss)
+                empties += 1;
+                if empties >= DEPRIME_AFTER_CALLBACKS {
+                    primed = false;
+                }
+            } else {
+                empties = 0;
             }
             cb_counters
                 .ring_depth
                 .store(ring.len() as u64, Ordering::Relaxed);
+            // Google's AAudio anti-glitch technique: when the device reports new XRuns, grow the HW
+            // buffer by one burst (up to capacity). getXRunCount + setBufferSizeInFrames are both
+            // callback-safe / non-blocking, and set clamps to capacity so it self-limits. Throttled.
+            cb_count = cb_count.wrapping_add(1);
+            if cb_count % XRUN_CHECK_EVERY == 0 {
+                let xr = s.x_run_count();
+                if xr > last_xrun {
+                    last_xrun = xr;
+                    let burst = s.frames_per_burst().max(1);
+                    let grown =
+                        (s.buffer_size_in_frames() + burst).min(s.buffer_capacity_in_frames());
+                    let _ = s.set_buffer_size_in_frames(grown);
+                }
+            }
             AudioCallbackResult::Continue
         };
 
@@ -93,7 +217,11 @@ impl AudioPlayback {
             .ok()?
             .direction(AudioDirection::Output)
             .sample_rate(SAMPLE_RATE)
-            .channel_count(CHANNELS as i32)
+            // The wire order (FL FR FC LFE RL RR SL SR) is the standard AAudio/Android channel
+            // order, so this is an IDENTITY mapping — no permute. AAudio infers the 5.1/7.1 mask
+            // from `channel_count` (the ndk crate's builder exposes no setChannelMask); the host
+            // captures + Opus-encodes in exactly this order.
+            .channel_count(channels as i32)
             .format(AudioFormat::PCM_Float)
             .performance_mode(AudioPerformanceMode::LowLatency)
             .sharing_mode(AudioSharingMode::Shared)
@@ -109,19 +237,31 @@ impl AudioPlayback {
             log::error!("audio: request_start: {e}");
             return None;
         }
+        // Lift the AAudio HW buffer off its brittle ~2-burst LowLatency default so a single late
+        // callback doesn't immediately underrun; the in-callback XRun loop grows it further if the
+        // device still glitches. set_buffer_size_in_frames clamps to capacity.
+        let burst = stream.frames_per_burst().max(1);
+        let _ =
+            stream.set_buffer_size_in_frames((burst * 3).min(stream.buffer_capacity_in_frames()));
+        // perf != LowLatency or rate != 48000 means AAudio silently fell to a resampled legacy path
+        // (different burst behaviour) — surface it so the field can tell that apart from plain jitter.
         log::info!(
-            "audio: AAudio started rate={} ch={} fmt={:?} burst={}",
+            "audio: AAudio started rate={} ch={} fmt={:?} perf={:?} share={:?} burst={} buf={}/{}",
             stream.sample_rate(),
             stream.channel_count(),
             stream.format(),
+            stream.performance_mode(),
+            stream.sharing_mode(),
             stream.frames_per_burst(),
+            stream.buffer_size_in_frames(),
+            stream.buffer_capacity_in_frames(),
         );
 
         let shutdown = Arc::new(AtomicBool::new(false));
         let sd = shutdown.clone();
         let join = std::thread::Builder::new()
             .name("pf-audio".into())
-            .spawn(move || decode_loop(client, tx, sd, counters))
+            .spawn(move || decode_loop(client, tx, free_rx, sd, counters, channels))
             .ok();
 
         Some(AudioPlayback {
@@ -143,31 +283,53 @@ impl Drop for AudioPlayback {
 }
 
 /// Producer: `next_audio` → Opus `decode_float` → push interleaved f32 into the ring channel.
+/// Buffers come from (and return to) the realtime callback's recycle free-list so the steady state
+/// is allocation-free on both threads.
 fn decode_loop(
     client: Arc<NativeClient>,
     tx: SyncSender<Vec<f32>>,
+    free_rx: Receiver<Vec<f32>>,
     shutdown: Arc<AtomicBool>,
     counters: Arc<Counters>,
+    channels: usize,
 ) {
-    let mut dec = match opus::Decoder::new(SAMPLE_RATE as u32, opus::Channels::Stereo) {
+    // Interleaved f32 samples per millisecond at this layout — the ring's 5 ms reserve check below.
+    let ms = (SAMPLE_RATE as usize / 1000) * channels;
+    // Opus decode scratch: worst-case 120 ms frame (5760 samples/ch) × channels.
+    let pcm_scratch = 5760 * channels;
+    let mut dec = match AudioDec::new(channels as u8) {
         Ok(d) => d,
         Err(e) => {
             log::error!("audio: opus decoder init: {e} — audio disabled");
             return;
         }
     };
-    let mut pcm = vec![0f32; PCM_SCRATCH];
+    let mut pcm = vec![0f32; pcm_scratch];
     let mut window_peak = 0f32; // loudest |sample| since the last log — tells a tone from silence
     while !shutdown.load(Ordering::Relaxed) {
         match client.next_audio(Duration::from_millis(5)) {
             Ok(pkt) => match dec.decode_float(&pkt.data, &mut pcm, false) {
                 Ok(samples) => {
-                    let n = samples * CHANNELS;
+                    let n = samples * channels;
                     for &s in &pcm[..n] {
                         window_peak = window_peak.max(s.abs());
                     }
+                    // The ring's pre-reservation in `start` assumes the protocol's 5 ms (≤480-f32/ch)
+                    // frames; a larger frame would force a one-time realloc on the RT thread. Catch a
+                    // future host frame-size change here in debug, not as a silent audio glitch.
+                    debug_assert!(
+                        n <= 5 * ms,
+                        "audio frame {n} f32 exceeds the 5 ms ring reserve"
+                    );
                     let count = counters.opus_decoded.fetch_add(1, Ordering::Relaxed) + 1;
-                    match tx.try_send(pcm[..n].to_vec()) {
+                    // Reuse a recycled buffer if the callback handed one back; only allocate when the
+                    // free-list is momentarily empty (startup / after a backpressure drop).
+                    let mut buf = free_rx
+                        .try_recv()
+                        .unwrap_or_else(|_| Vec::with_capacity(pcm_scratch));
+                    buf.clear();
+                    buf.extend_from_slice(&pcm[..n]);
+                    match tx.try_send(buf) {
                         Ok(()) | Err(TrySendError::Full(_)) => {} // drop-newest under backpressure
                         Err(TrySendError::Disconnected(_)) => break,
                     }

clients/android/native/src/session.rs

@@ -140,10 +140,12 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeGenerateIde
 }
 
 /// `NativeBridge.nativeConnect(host, port, w, h, hz, certPem, keyPem, pinHex, bitrateKbps,
-/// compositorPref, gamepadPref): Long`. `certPem`/`keyPem` empty = anonymous, else presented as the
-/// persistent identity. `pinHex` empty = TOFU (read `nativeHostFingerprint` after), else 64-hex
-/// SHA-256 to pin the host (mismatch → 0). `bitrateKbps` 0 = host default. `compositorPref`/
-/// `gamepadPref` are `CompositorPref`/`GamepadPref` wire bytes (0 = Auto; unknown → Auto).
+/// compositorPref, gamepadPref, hdrEnabled, audioChannels): Long`. `certPem`/`keyPem` empty =
+/// anonymous, else presented as the persistent identity. `pinHex` empty = TOFU (read
+/// `nativeHostFingerprint` after), else 64-hex SHA-256 to pin the host (mismatch → 0). `bitrateKbps`
+/// 0 = host default. `compositorPref`/`gamepadPref` are `CompositorPref`/`GamepadPref` wire bytes
+/// (0 = Auto; unknown → Auto). `audioChannels` is the requested surround layout (2/6/8; normalized,
+/// anything else → stereo) — the host clamps it and the resolved count drives playback.
 /// Returns an opaque handle, or 0 on failure (logged).
 #[no_mangle]
 #[allow(clippy::too_many_arguments)]
@@ -162,6 +164,7 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeConnect<'lo
     compositor_pref: jint,
     gamepad_pref: jint,
     hdr_enabled: jboolean,
+    audio_channels: jint,
 ) -> jlong {
     let host: String = match env.get_string(&host) {
         Ok(s) => s.into(),
@@ -213,6 +216,11 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeConnect<'lo
         } else {
             0
         },
+        // Requested surround layout (2 = stereo / 6 = 5.1 / 8 = 7.1). The host clamps to what it can
+        // capture and echoes the resolved count in `connector.audio_channels`, which drives the
+        // decoder + AAudio layout (read in `crate::audio::AudioPlayback::start`). Anything else
+        // normalizes to stereo here.
+        punktfunk_core::audio::normalize_channels(audio_channels.clamp(0, u8::MAX as jint) as u8),
         None,     // launch: default app
         pin,      // Some → Crypto on host-fp mismatch
         identity, // owned (cert, key) PEM, or None (anonymous)

clients/apple/Sources/PunktfunkClient/ContentView.swift

@@ -25,6 +25,7 @@ struct ContentView: View {
     @AppStorage(DefaultsKey.compositor) private var compositor = 0
     @AppStorage(DefaultsKey.gamepadType) private var gamepadType = 0
     @AppStorage(DefaultsKey.bitrateKbps) private var bitrateKbps = 0
+    @AppStorage(DefaultsKey.audioChannels) private var audioChannels = 2
     @AppStorage(DefaultsKey.fullscreenWhileStreaming) private var fullscreenWhileStreaming = true
     @AppStorage(DefaultsKey.hudEnabled) private var hudEnabled = true
     @AppStorage(DefaultsKey.hudPlacement) private var hudPlacement = HUDPlacement.topTrailing.rawValue
@@ -252,6 +253,7 @@ struct ContentView: View {
                 setting: PunktfunkConnection.GamepadType(
                     rawValue: UInt32(clamping: gamepadType)) ?? .auto),
             bitrateKbps: UInt32(clamping: bitrateKbps),
+            audioChannels: UInt8(clamping: audioChannels),
             launchID: launchID,
             allowTofu: host.pinnedSHA256 == nil)
     }
@@ -351,6 +353,7 @@ struct ContentView: View {
             compositor: pref,
             gamepad: pad,
             bitrateKbps: bitrate,
+            audioChannels: UInt8(clamping: audioChannels),
             autoTrust: true)
     }
 }

clients/apple/Sources/PunktfunkClient/SessionModel.swift

@@ -99,6 +99,7 @@ final class SessionModel: ObservableObject {
                  compositor: PunktfunkConnection.Compositor = .auto,
                  gamepad: PunktfunkConnection.GamepadType = .auto,
                  bitrateKbps: UInt32 = 0,
+                 audioChannels: UInt8 = 2,
                  hdrEnabled: Bool = true,
                  launchID: String? = nil,
                  allowTofu: Bool = false,
@@ -137,7 +138,7 @@ final class SessionModel: ObservableObject {
                 width: width, height: height, refreshHz: hz,
                 pinSHA256: pin, identity: identity, compositor: compositor,
                 gamepad: gamepad, bitrateKbps: bitrateKbps, videoCaps: videoCaps,
-                launchID: launchID) }
+                audioChannels: audioChannels, launchID: launchID) }
             await MainActor.run { [weak self] in
                 guard let self else { return }
                 // The user may have abandoned this attempt (window closed, another host

clients/apple/Sources/PunktfunkClient/SettingsView.swift

@@ -25,6 +25,7 @@ struct SettingsView: View {
     @AppStorage(DefaultsKey.libraryEnabled) private var libraryEnabled = false
     @AppStorage(DefaultsKey.fullscreenWhileStreaming) private var fullscreenWhileStreaming = true
     @AppStorage(DefaultsKey.micEnabled) private var micEnabled = true
+    @AppStorage(DefaultsKey.audioChannels) private var audioChannels = 2
     @AppStorage(DefaultsKey.hudEnabled) private var hudEnabled = true
     @AppStorage(DefaultsKey.hudPlacement) private var hudPlacement = HUDPlacement.topTrailing.rawValue
     @ObservedObject private var gamepads = GamepadManager.shared
@@ -173,6 +174,10 @@ struct SettingsView: View {
                 TVSelectionRow(title: "Stream mode", options: options, selection: modeTag)
                 TVSelectionRow(
                     title: "Bitrate", options: bitrateOptions, selection: $bitrateKbps)
+                TVSelectionRow(
+                    title: "Audio channels",
+                    options: [("Stereo", 2), ("5.1 Surround", 6), ("7.1 Surround", 8)],
+                    selection: $audioChannels)
                 if bitrateKbps > 1_000_000 {
                     Label(Self.gigabitWarning, systemImage: "exclamationmark.triangle.fill")
                         .font(.caption)
@@ -271,6 +276,11 @@ struct SettingsView: View {
 
     @ViewBuilder private var audioSection: some View {
         Section {
+            Picker("Audio channels", selection: $audioChannels) {
+                Text("Stereo").tag(2)
+                Text("5.1 Surround").tag(6)
+                Text("7.1 Surround").tag(8)
+            }
             #if os(macOS)
             Picker("Speaker", selection: $speakerUID) {
                 Text("System default").tag("")

clients/apple/Sources/PunktfunkKit/DefaultsKeys.swift

@@ -15,6 +15,9 @@ public enum DefaultsKey {
     public static let gamepadType = "punktfunk.gamepadType"
     public static let gamepadID = "punktfunk.gamepadID"
     public static let bitrateKbps = "punktfunk.bitrateKbps"
+    /// Requested audio channel count: 2 (stereo), 6 (5.1) or 8 (7.1). The host clamps to what it
+    /// can capture; the resolved count drives the in-core decode + AVAudioEngine layout.
+    public static let audioChannels = "punktfunk.audioChannels"
     public static let micEnabled = "punktfunk.micEnabled"
     public static let speakerUID = "punktfunk.speakerUID"
     public static let micUID = "punktfunk.micUID"

clients/apple/Sources/PunktfunkKit/PunktfunkConnection.swift

@@ -235,6 +235,12 @@ public final class PunktfunkConnection {
     /// drain `nextHdrMeta`.
     public var isHDR: Bool { colorTransfer == 16 || colorTransfer == 18 }
 
+    /// The audio channel count the host resolved for this session (the Welcome's echo of the
+    /// requested `audioChannels`, clamped to what the host can capture): `2` (stereo), `6` (5.1)
+    /// or `8` (7.1). Build the playback layout from THIS, never the request. `2` for an older host.
+    /// PCM from `nextAudioPcm` is interleaved in the canonical wire order FL FR FC LFE RL RR SL SR.
+    public private(set) var resolvedAudioChannels: UInt8 = 2
+
     /// Connect and start a session at the requested mode (the host creates a native virtual
     /// output at exactly this size/refresh). Blocks up to `timeoutMs`.
     ///
@@ -264,6 +270,7 @@ public final class PunktfunkConnection {
         gamepad: GamepadType = .auto,
         bitrateKbps: UInt32 = 0,
         videoCaps: UInt8 = 0,
+        audioChannels: UInt8 = 2,
         launchID: String? = nil,
         timeoutMs: UInt32 = 10_000
     ) throws {
@@ -279,16 +286,16 @@ public final class PunktfunkConnection {
                     withOptionalCString(launchID) { launch in
                         if let pin = pinSHA256 {
                             return pin.withUnsafeBytes { p in
-                                punktfunk_connect_ex5(
+                                punktfunk_connect_ex6(
                                     cs, port, width, height, refreshHz, compositor.rawValue,
-                                    gamepad.rawValue, bitrateKbps, videoCaps, launch,
+                                    gamepad.rawValue, bitrateKbps, videoCaps, audioChannels, launch,
                                     p.bindMemory(to: UInt8.self).baseAddress, &observed,
                                     cert, key, timeoutMs)
                             }
                         }
-                        return punktfunk_connect_ex5(
+                        return punktfunk_connect_ex6(
                             cs, port, width, height, refreshHz, compositor.rawValue,
-                            gamepad.rawValue, bitrateKbps, videoCaps, launch,
+                            gamepad.rawValue, bitrateKbps, videoCaps, audioChannels, launch,
                             nil, &observed, cert, key, timeoutMs)
                     }
                 }
@@ -320,6 +327,9 @@ public final class PunktfunkConnection {
         colorMatrix = mtx
         colorFullRange = fullRange != 0
         bitDepth = depth
+        var ac: UInt8 = 2
+        _ = punktfunk_connection_audio_channels(handle, &ac)
+        resolvedAudioChannels = ac
     }
 
     /// A bandwidth speed-test measurement (see `startSpeedTest`). Partial until `done`.
@@ -468,6 +478,50 @@ public final class PunktfunkConnection {
         }
     }
 
+    /// One decoded audio frame from `nextAudioPcm`: interleaved 32-bit float at 48 kHz, in the
+    /// canonical wire channel order FL FR FC LFE RL RR SL SR (the first `channels`).
+    public struct AudioPCM: Sendable {
+        /// Interleaved f32 samples (`frameCount * channels` long), wire channel order.
+        public let samples: [Float]
+        /// Samples per channel.
+        public let frameCount: Int
+        /// Channel count (2/6/8) — `resolvedAudioChannels`.
+        public let channels: Int
+        public let ptsNs: UInt64
+        public let seq: UInt32
+    }
+
+    /// Pull the next audio frame, **decoded in-core** to interleaved f32 PCM — Apple's AudioToolbox
+    /// Opus path is stereo-only, so surround (and, for uniformity, stereo too) is decoded by the
+    /// Rust core (libopus multistream) and handed back as PCM. nil on timeout, throws `.closed` once
+    /// the session ended. Drain from a dedicated audio thread (do NOT also call `nextAudio` — they
+    /// share the underlying queue). The returned `samples` are copied out, so the buffer is owned.
+    public func nextAudioPcm(timeoutMs: UInt32 = 100) throws -> AudioPCM? {
+        audioLock.lock()
+        defer { audioLock.unlock() }
+        guard let h = liveHandle() else { throw PunktfunkClientError.closed }
+
+        var out = PunktfunkAudioPcm()
+        let rc = punktfunk_connection_next_audio_pcm(h, &out, timeoutMs)
+        switch rc {
+        case statusOK:
+            let channels = Int(out.channels)
+            let total = Int(out.frame_count) * channels
+            guard let base = out.samples, total > 0 else { return nil }
+            // Copy: the pointer borrows connection memory only until the next PCM call.
+            let samples = Array(UnsafeBufferPointer(start: base, count: total))
+            return AudioPCM(
+                samples: samples, frameCount: Int(out.frame_count),
+                channels: channels, ptsNs: out.pts_ns, seq: out.seq)
+        case statusNoFrame:
+            return nil
+        case statusClosed:
+            throw PunktfunkClientError.closed
+        default:
+            throw PunktfunkClientError.status(rc)
+        }
+    }
+
     /// Pull the next force-feedback update for the GCController haptics engine:
     /// `(pad, lowFrequency, highFrequency)` with 0...0xFFFF amplitudes, (0, 0) = stop.
     /// Drain from the (single) feedback thread, alongside `nextHidOutput`.

clients/apple/Sources/PunktfunkKit/SessionAudio.swift

@@ -19,13 +19,13 @@ import os
 
 private let log = Logger(subsystem: "io.unom.punktfunk", category: "audio")
 
-/// SPSC-ish jitter ring (interleaved stereo float), drain thread → render callback.
-/// The unfair lock is held for microseconds; fine at render-callback rates. Priming:
+/// SPSC-ish jitter ring (interleaved float, `channels` per frame), drain thread → render
+/// callback. The unfair lock is held for microseconds; fine at render-callback rates. Priming:
 /// reads return silence until enough is buffered (at least `prefill`, and at least one
 /// packet more than the device's render quantum — large-buffer devices would otherwise
 /// chronically out-demand the prefill and oscillate prime → dropout → re-prime), and an
 /// underrun re-primes, concealing jitter as one short dip instead of sustained crackle.
-/// All counts stay even (whole stereo frames), so L/R interleave can never flip.
+/// All counts stay whole frames (multiples of `channels`), so the interleave can never slip.
 final class AudioRing: @unchecked Sendable {
     private var buf: [Float]
     private var readIdx = 0
@@ -34,12 +34,14 @@ final class AudioRing: @unchecked Sendable {
     private var renderQuantum = 0
     private let prefill: Int
     private let highWater: Int
+    private let channels: Int
     private let lock = OSAllocatedUnfairLock()
 
-    /// `capacity`/`prefill` in samples (interleaved — 2 per frame, both must be even).
-    init(capacity: Int, prefill: Int) {
+    /// `capacity`/`prefill` in samples (interleaved — `channels` per frame, both whole frames).
+    init(capacity: Int, prefill: Int, channels: Int) {
         buf = [Float](repeating: 0, count: capacity)
         self.prefill = prefill
+        self.channels = channels
         highWater = prefill * 4
     }
 
@@ -74,8 +76,8 @@ final class AudioRing: @unchecked Sendable {
         renderQuantum = max(renderQuantum, count)
         let available = writeIdx - readIdx
         if !primed {
-            // 480 samples = one 5 ms host packet of slack beyond the device's demand.
-            if available >= max(prefill, renderQuantum + 480) {
+            // One 5 ms host packet (240 frames × channels) of slack beyond the device's demand.
+            if available >= max(prefill, renderQuantum + 240 * channels) {
                 primed = true
             } else {
                 for i in 0..<count { out[i] = 0 }
@@ -113,10 +115,55 @@ private final class StopFlag: @unchecked Sendable {
 /// Render-block-owned scratch storage: freed exactly when the closure (and thus the
 /// last possible render call) is released — never racing CoreAudio.
 private final class ScratchBuffer {
-    let ptr = UnsafeMutablePointer<Float>.allocate(capacity: 8192 * 2)
+    // 8192 frames × up to 8 channels (7.1) — the render block caps `frames` at 8192.
+    let ptr = UnsafeMutablePointer<Float>.allocate(capacity: 8192 * 8)
     deinit { ptr.deallocate() }
 }
 
+/// CoreAudio channel layout for the canonical wire order FL FR FC LFE RL RR [SL SR]. nil for
+/// stereo (the standard layout is correct). For 5.1/7.1 we list explicit channel labels via
+/// `kAudioChannelLayoutTag_UseChannelDescriptions` — preset tags (DTS_5_1 etc.) don't reliably
+/// match Moonlight's order. NB the 7.1 mapping (verified against the WASAPI 0x63F + SPA orderings):
+/// wire idx 4-5 = RL/RR = the WAVE *back* pair → LeftSurround/RightSurround; idx 6-7 = SL/SR = the
+/// WAVE *side* pair → LeftSurroundDirect/RightSurroundDirect. (Using RearSurround* for 6-7 would
+/// swap side/back vs the Windows/Linux clients.)
+private func wireChannelLayout(channels: Int) -> AVAudioChannelLayout? {
+    let labels: [AudioChannelLabel]
+    switch channels {
+    case 6:
+        labels = [
+            kAudioChannelLabel_Left, kAudioChannelLabel_Right, kAudioChannelLabel_Center,
+            kAudioChannelLabel_LFEScreen, kAudioChannelLabel_LeftSurround,
+            kAudioChannelLabel_RightSurround,
+        ]
+    case 8:
+        labels = [
+            kAudioChannelLabel_Left, kAudioChannelLabel_Right, kAudioChannelLabel_Center,
+            kAudioChannelLabel_LFEScreen,
+            kAudioChannelLabel_LeftSurround, kAudioChannelLabel_RightSurround, // wire RL/RR (back)
+            kAudioChannelLabel_LeftSurroundDirect, kAudioChannelLabel_RightSurroundDirect, // wire SL/SR (side)
+        ]
+    default:
+        return nil
+    }
+    let size = MemoryLayout<AudioChannelLayout>.size
+        + (labels.count - 1) * MemoryLayout<AudioChannelDescription>.stride
+    let raw = UnsafeMutableRawPointer.allocate(byteCount: size, alignment: 16)
+    defer { raw.deallocate() }
+    let layout = raw.bindMemory(to: AudioChannelLayout.self, capacity: 1)
+    layout.pointee.mChannelLayoutTag = kAudioChannelLayoutTag_UseChannelDescriptions
+    layout.pointee.mChannelBitmap = AudioChannelBitmap(rawValue: 0)
+    layout.pointee.mNumberChannelDescriptions = UInt32(labels.count)
+    let descs = UnsafeMutableBufferPointer(
+        start: &layout.pointee.mChannelDescriptions, count: labels.count)
+    for (i, lbl) in labels.enumerated() {
+        descs[i] = AudioChannelDescription(
+            mChannelLabel: lbl, mChannelFlags: AudioChannelFlags(rawValue: 0),
+            mCoordinates: (0, 0, 0))
+    }
+    return AVAudioChannelLayout(layout: layout)
+}
+
 public final class SessionAudio {
     private let connection: PunktfunkConnection
     private let flag = StopFlag()
@@ -229,9 +276,13 @@ public final class SessionAudio {
     // MARK: - Playback (host → speaker)
 
     private func startPlayback(speakerUID: String) {
-        // 1 s of interleaved stereo capacity, ~20 ms prefill: four 5 ms host packets of
-        // jitter absorption before the first sample plays.
-        let ring = AudioRing(capacity: 96_000, prefill: 1920)
+        // Build the playback layout from the host-RESOLVED channel count (never the request):
+        // 2 = stereo / 6 = 5.1 / 8 = 7.1, canonical wire order FL FR FC LFE RL RR SL SR.
+        let channels = Int(connection.resolvedAudioChannels)
+        // 1 s interleaved capacity, ~20 ms prefill (four 5 ms host packets of jitter absorption
+        // before the first sample plays), both scaled by the channel count.
+        let ring = AudioRing(
+            capacity: 48_000 * channels, prefill: 960 * channels, channels: channels)
 
         let engine = AVAudioEngine()
         #if os(macOS)
@@ -247,21 +298,32 @@ public final class SessionAudio {
         }
         #endif
 
-        // Engine-native deinterleaved float; the render block deinterleaves from the ring.
-        guard let format = AVAudioFormat(standardFormatWithSampleRate: 48_000, channels: 2)
-        else { return }
+        // Engine-native deinterleaved float; the render block deinterleaves from the ring. Surround
+        // uses an explicit wire-order channel layout; the mixer downmixes to the output device when
+        // it has fewer speakers (e.g. an iPhone's stereo built-ins). (Explicit if/else rather than
+        // map/flatMap so it's correct whether the channelLayout initializer is failable or not.)
+        var format: AVAudioFormat?
+        if channels == 2 {
+            format = AVAudioFormat(standardFormatWithSampleRate: 48_000, channels: 2)
+        } else if let layout = wireChannelLayout(channels: channels) {
+            format = AVAudioFormat(standardFormatWithSampleRate: 48_000, channelLayout: layout)
+        }
+        guard let format else {
+            log.error("could not build \(channels)-channel audio format — audio disabled")
+            return
+        }
         let scratch = ScratchBuffer() // block-owned; freed with the closure
         let source = AVAudioSourceNode(format: format) { _, _, frameCount, abl -> OSStatus in
             let frames = Int(frameCount)
             guard frames <= 8192 else { return kAudioUnitErr_TooManyFramesToProcess }
-            ring.read(into: scratch.ptr, count: frames * 2)
+            ring.read(into: scratch.ptr, count: frames * channels)
             let buffers = UnsafeMutableAudioBufferListPointer(abl)
-            if buffers.count >= 2,
-               let left = buffers[0].mData?.assumingMemoryBound(to: Float.self),
-               let right = buffers[1].mData?.assumingMemoryBound(to: Float.self) {
-                for f in 0..<frames {
-                    left[f] = scratch.ptr[f * 2]
-                    right[f] = scratch.ptr[f * 2 + 1]
+            // Deinterleave the wire-order interleaved ring into the engine's per-channel buses.
+            if buffers.count >= channels {
+                for ch in 0..<channels {
+                    if let dst = buffers[ch].mData?.assumingMemoryBound(to: Float.self) {
+                        for f in 0..<frames { dst[f] = scratch.ptr[f * channels + ch] }
+                    }
                 }
             }
             return noErr
@@ -292,29 +354,20 @@ public final class SessionAudio {
         stateLock.unlock()
         let thread = Thread { [connection, flag, drainDone] in
             defer { drainDone.signal() }
-            guard let decoder = try? OpusDecoder(framesPerPacket: 240),
-                  let pcm = AVAudioPCMBuffer(
-                      pcmFormat: decoder.pcmFormat, frameCapacity: 5760)
-            else {
-                log.error("Opus decoder unavailable — audio playback disabled")
-                return
-            }
+            // Decode happens IN-CORE (libopus multistream) — AudioToolbox's Opus path is
+            // stereo-only — and is handed back as interleaved f32 PCM in wire channel order.
             while !flag.isStopped {
-                let packet: AudioPacket?
+                let pcm: PunktfunkConnection.AudioPCM?
                 do {
-                    packet = try connection.nextAudio(timeoutMs: 100)
+                    pcm = try connection.nextAudioPcm(timeoutMs: 100)
                 } catch {
                     break // session closed
                 }
-                guard let packet else { continue }
-                do {
-                    let frames = try decoder.decode(packet.data, into: pcm)
-                    if frames > 0, let p = pcm.floatChannelData?[0] {
-                        ring.write(p, count: Int(frames) * 2)
+                guard let pcm, pcm.frameCount > 0 else { continue }
+                pcm.samples.withUnsafeBufferPointer { p in
+                    if let base = p.baseAddress {
+                        ring.write(base, count: pcm.frameCount * pcm.channels)
                     }
-                } catch {
-                    // One corrupt packet ≠ a dead stream; skip it.
-                    log.warning("audio decode failed: \(error.localizedDescription)")
                 }
             }
         }

clients/decky/README.md

@@ -45,8 +45,9 @@ Gaming Mode automatically.
 | `src/steam.ts` | Steam-shortcut launch (`AddShortcut` / `SetAppLaunchOptions` / `RunGame`) — the focus-correct stream start. |
 | `src/backend.ts` | Typed `callable` bridges to `main.py`. |
 | `bin/punktfunkrun.sh` | The launch wrapper the Steam shortcut targets (so the window is focusable). |
-| `main.py` | Backend: `discover` / `pair` / `runner_info` / `get_settings` / `set_settings` / `kill_stream`. |
+| `main.py` | Backend: `discover` / `pair` / `runner_info` / `get_settings` / `set_settings` / `kill_stream` / `check_update`. |
 | `plugin.json` | Decky plugin manifest. |
+| `update.json` | CI-baked `{channel, manifest}` — where `check_update()` polls (absent on dev builds). |
 | `decky.pyi` | Type stub for the injected `decky` module (vendored from the template). |
 
 ### Discovery (`discover()`)
@@ -140,6 +141,40 @@ shows up in the Quick Access Menu.
 > [`../../packaging/flatpak/README.md`](../../packaging/flatpak/README.md)) — install that on
 > the Deck too, or the panel's Connect surfaces a `client-not-found` error.
 
+## Updating (self-update, no store)
+
+The plugin updates itself without the official Decky store. CI (`decky.yml`) publishes a tiny
+per-channel `manifest.json` next to the zip in the Gitea registry:
+
+```json
+{"version":"0.3.123","artifact":".../punktfunk-decky/0.3.123/punktfunk.zip","sha256":"…"}
+```
+
+and bakes an `update.json` (`{channel, manifest}`) into the plugin so it knows which channel it was
+installed from. The backend `check_update()` reads the **installed** version from `package.json` —
+the value Decky itself reports (it does **not** read `plugin.json`) — fetches the channel manifest,
+and compares. When a newer build exists the frontend shows an **Update to vX** button that drives
+Decky Loader's own install RPC:
+
+```ts
+window.DeckyBackend.callable("utilities/install_plugin")(artifact, "punktfunk", version, hash, /*UPDATE=*/2)
+```
+
+The loader (root) downloads the immutable per-version zip, **SHA-256-verifies** it against `hash`,
+replaces `~/homebrew/plugins/punktfunk`, and hot-reloads — the unprivileged backend never writes the
+root-owned plugins dir itself. `window.DeckyBackend` / `utilities/install_plugin` are loader
+internals (not `@decky/api`), so every access is guarded; missing them, the button falls back to a
+toast pointing at **Install Plugin from URL**.
+
+> CI stamps a **plain numeric** semver per channel (`0.3.<run>` canary, `X.Y.Z` stable) into
+> `package.json`. Decky's `compare-versions` orders pre-release identifiers lexically (so `ci10 < ci9`)
+> — a `-ciN` suffix would mis-detect updates.
+
+**Optional — native Updates tab:** Decky's store is single-source (a custom store URL *replaces* the
+official catalog), so punktfunk doesn't ship one by default. A user who wants the native update badge
+can point Decky → Settings → **Custom store** at a punktfunk-only store JSON — not recommended if you
+use other plugins, since it hides the official catalog.
+
 ## Limitations / next steps
 
 - **Needs on-Deck validation in Gaming Mode**: the Steam-shortcut launch (`AddShortcut` /

clients/decky/bin/punktfunkrun.sh

@@ -31,4 +31,6 @@ fi
 echo "punktfunkrun: streaming $APPID --connect $PF_HOST" >&2
 # exec so the flatpak client IS the game process — when it exits, Steam ends the "game" and
 # Gaming Mode reclaims focus automatically (no manual refocus needed).
-exec "$FLATPAK" run --arch=x86_64 "$APPID" --connect "$PF_HOST"
+# --fullscreen: present the stream chrome-less and fullscreen (the client also auto-detects the
+# Deck/gamescope env, and ignores the flag harmlessly on older builds that predate it).
+exec "$FLATPAK" run --arch=x86_64 "$APPID" --connect "$PF_HOST" --fullscreen

clients/decky/main.py

@@ -17,6 +17,8 @@ The backend's jobs are the things Steam can't do:
 * **get_settings() / set_settings()** — read/write the flatpak client's stream settings JSON
   (resolution / bitrate / gamepad), so the Deck UI configures the stream the client reads.
 * **kill_stream()** — force-stop a wedged stream (``flatpak kill``).
+* **check_update()** — poll the registry's per-channel ``manifest.json`` and report whether a
+  newer build is available (the frontend then drives Decky's own install RPC to apply it).
 
 The TXT-record keys parsed (``proto`` / ``fp`` / ``pair`` / ``id``) are defined by the host
 advert in ``crates/punktfunk-host/src/discovery.rs``.
@@ -26,7 +28,10 @@ import asyncio
 import json
 import os
 import shutil
+import ssl
 import stat
+import time
+import urllib.request
 from pathlib import Path
 
 import decky
@@ -37,22 +42,99 @@ APP_ID = "io.unom.Punktfunk"
 # Service type advertised by punktfunk/1 hosts (matches NATIVE_SERVICE in the Rust host).
 SERVICE_TYPE = "_punktfunk._udp"
 
-# The flatpak client persists identity / known-hosts / settings under HOME/.config/punktfunk;
-# inside the flatpak sandbox HOME is ~/.var/app/<APP_ID>, so the real on-disk location is this.
-# The backend writes settings here so the (sandboxed) client reads them.
+# The flatpak client persists identity / known-hosts / settings under HOME/.config/punktfunk.
+# The sandbox HOME resolves to the REAL user home (== DECKY_USER_HOME), NOT the per-app
+# ~/.var/app/<APP_ID> dir — verified on-device (`flatpak run … sh -c 'echo $HOME'` prints
+# /home/deck, and the manifest's `--filesystem=~/.config/punktfunk` grants exactly that path;
+# we also pass HOME=DECKY_USER_HOME into `flatpak run`, see _flatpak_env). Pointing here is what
+# lets plugin settings actually reach the client AND lets us read the client's known-hosts to
+# tell whether THIS device is already paired with a given host.
 def _client_config_dir() -> Path:
-    return Path(decky.DECKY_USER_HOME) / ".var" / "app" / APP_ID / ".config" / "punktfunk"
+    return Path(decky.DECKY_USER_HOME) / ".config" / "punktfunk"
 
 
 def _settings_path() -> Path:
     return _client_config_dir() / "client-gtk-settings.json"
 
 
+def _paired_fingerprints() -> set[str]:
+    """Host cert fingerprints (lowercase hex) this client has PIN-paired, from the client's
+    known-hosts store. Keyed by fingerprint so it survives a host changing IP address."""
+    try:
+        data = json.loads((_client_config_dir() / "client-known-hosts.json").read_text())
+    except (OSError, json.JSONDecodeError):
+        return set()
+    hosts = data.get("hosts", []) if isinstance(data, dict) else []
+    return {
+        h["fp_hex"].lower()
+        for h in hosts
+        if isinstance(h, dict) and h.get("paired") and isinstance(h.get("fp_hex"), str)
+    }
+
+
 def _runner_path() -> str:
     """Absolute path to the launch wrapper shipped with the plugin (bin/punktfunkrun.sh)."""
     return str(Path(decky.DECKY_PLUGIN_DIR) / "bin" / "punktfunkrun.sh")
 
 
+# ----------------------------------------------------------------------------------------
+# Self-update check (no Decky store). The plugin is distributed via "Install Plugin from
+# URL" pointing at our Gitea generic registry, so the official store never sees it and
+# can't offer updates. Instead the backend polls a tiny per-channel ``manifest.json`` the
+# CI publishes next to the zip, compares it to the installed version, and the frontend
+# offers a one-tap update that drives Decky's own (root, privileged) install RPC. The
+# channel + manifest URL are baked into ``update.json`` by CI (.gitea/workflows/decky.yml);
+# a dev/sideload build has no ``update.json`` and update checks are simply disabled.
+_UPDATE_TTL_S = 1800.0  # cache a successful check for 30 min (the QAM remounts often)
+_update_cache: dict = {"at": 0.0, "data": None}
+
+
+def _update_config() -> dict:
+    """The CI-baked ``{channel, manifest}`` next to the plugin (absent on dev builds)."""
+    try:
+        return json.loads((Path(decky.DECKY_PLUGIN_DIR) / "update.json").read_text())
+    except (OSError, json.JSONDecodeError):
+        return {}
+
+
+def _installed_version() -> str:
+    """The version Decky itself reports for this plugin — it reads ``package.json`` (NOT
+    plugin.json), so the CI stamps the build version there."""
+    try:
+        pkg = json.loads((Path(decky.DECKY_PLUGIN_DIR) / "package.json").read_text())
+        return str(pkg.get("version", "0.0.0"))
+    except (OSError, json.JSONDecodeError):
+        return "0.0.0"
+
+
+def _semver_tuple(v: str) -> tuple[int, int, int]:
+    """A tolerant (major, minor, patch) tuple for ``>`` comparison. We control the version
+    format (plain numeric ``X.Y.Z`` on both channels), so leading-int-per-component is
+    enough; any pre-release suffix is dropped before comparing."""
+    parts: list[int] = []
+    for comp in str(v).split("-", 1)[0].split(".")[:3]:
+        digits = ""
+        for ch in comp:
+            if ch.isdigit():
+                digits += ch
+            else:
+                break
+        parts.append(int(digits) if digits else 0)
+    while len(parts) < 3:
+        parts.append(0)
+    return (parts[0], parts[1], parts[2])
+
+
+def _fetch_json(url: str, timeout: float = 8.0) -> dict:
+    """Blocking HTTPS GET of a small JSON document (run in an executor)."""
+    req = urllib.request.Request(
+        url, headers={"Accept": "application/json", "User-Agent": "punktfunk-decky"}
+    )
+    ctx = ssl.create_default_context()
+    with urllib.request.urlopen(req, timeout=timeout, context=ctx) as resp:
+        return json.loads(resp.read().decode("utf-8", errors="replace"))
+
+
 def _flatpak() -> str | None:
     return shutil.which("flatpak") or (
         "/usr/bin/flatpak" if Path("/usr/bin/flatpak").exists() else None
@@ -179,6 +261,13 @@ class Plugin:
         if stderr:
             decky.logger.debug("avahi-browse stderr: %s", stderr.decode(errors="replace"))
         hosts = _parse_avahi_browse(stdout.decode(errors="replace"))
+        # Mark which hosts THIS device has already paired (by cert fingerprint), so the UI can
+        # show "Stream" instead of "Pair" — the mDNS `pair` field is the host's policy, not our
+        # per-device pairing state.
+        paired = _paired_fingerprints()
+        for h in hosts:
+            fp = h.get("fp") or ""
+            h["paired"] = bool(fp) and fp.lower() in paired
         decky.logger.info("discovered %d punktfunk host(s)", len(hosts))
         return hosts
 
@@ -279,6 +368,54 @@ class Plugin:
             return {"ok": False}
         return {"ok": True}
 
+    async def check_update(self, force: bool = False) -> dict:
+        """Is a newer build available in our registry? Compares the installed version
+        (``package.json``) against the per-channel ``manifest.json`` the CI publishes, and
+        returns everything the frontend needs to drive Decky's install RPC. Non-fatal: any
+        failure (no channel baked in, network down) returns ``update_available: False``.
+        """
+        current = _installed_version()
+        cfg = _update_config()
+        result = {
+            "current": current,
+            "latest": current,
+            "artifact": "",
+            "hash": "",
+            "channel": str(cfg.get("channel", "")),
+            "update_available": False,
+        }
+
+        manifest_url = cfg.get("manifest")
+        if not manifest_url:
+            result["error"] = "update-channel-unknown"  # dev / sideloaded build
+            return result
+
+        now = time.monotonic()
+        cached = _update_cache["data"]
+        if not force and cached and (now - _update_cache["at"]) < _UPDATE_TTL_S:
+            return cached
+
+        try:
+            loop = asyncio.get_running_loop()
+            manifest = await loop.run_in_executor(None, _fetch_json, manifest_url)
+        except Exception as exc:  # noqa: BLE001
+            decky.logger.warning("update check failed: %s", exc)
+            result["error"] = "fetch-failed"
+            return result  # transient — don't cache, retry next open
+
+        latest = str(manifest.get("version", current))
+        result["latest"] = latest
+        result["artifact"] = str(manifest.get("artifact", ""))
+        result["hash"] = str(manifest.get("sha256", ""))
+        result["update_available"] = bool(result["artifact"]) and (
+            _semver_tuple(latest) > _semver_tuple(current)
+        )
+        if result["update_available"]:
+            decky.logger.info("update available: %s -> %s (%s)", current, latest, result["channel"])
+        _update_cache["at"] = now
+        _update_cache["data"] = result
+        return result
+
     # ---- Decky lifecycle ----
 
     async def _main(self):

clients/decky/src/backend.ts

@@ -5,8 +5,9 @@ export interface Host {
   name: string;
   host: string;
   port: number;
-  pair: string; // "required" | "optional"
+  pair: string; // "required" | "optional" — the HOST's policy
   fp: string;
+  paired: boolean; // whether THIS device has already PIN-paired this host (by fingerprint)
 }
 
 export interface PairResult {
@@ -32,6 +33,16 @@ export interface StreamSettings {
   mic_enabled: boolean;
 }
 
+export interface UpdateInfo {
+  current: string; // installed version (package.json)
+  latest: string; // newest version in our registry for this channel
+  artifact: string; // immutable zip URL Decky should install
+  hash: string; // sha256 of that zip (Decky verifies it)
+  channel: string; // "latest" (stable) | "canary"
+  update_available: boolean;
+  error?: string; // "update-channel-unknown" (dev build) | "fetch-failed"
+}
+
 export const discover = callable<[], Host[]>("discover");
 export const pair = callable<
   [host: string, port: number, pin: string, name: string],
@@ -43,3 +54,4 @@ export const setSettings = callable<[settings: StreamSettings], { ok: boolean }>
   "set_settings",
 );
 export const killStream = callable<[], { ok: boolean }>("kill_stream");
+export const checkUpdate = callable<[force: boolean], UpdateInfo>("check_update");

clients/decky/src/index.tsx

@@ -10,12 +10,22 @@ import {
   PanelSectionRow,
   SliderField,
   Spinner,
+  Tabs,
   ToggleField,
   showModal,
   staticClasses,
 } from "@decky/ui";
 import { definePlugin, routerHook, toaster } from "@decky/api";
-import { FC, useCallback, useEffect, useState } from "react";
+import {
+  Component,
+  CSSProperties,
+  ErrorInfo,
+  FC,
+  ReactNode,
+  useCallback,
+  useEffect,
+  useState,
+} from "react";
 import {
   FaTv,
   FaSyncAlt,
@@ -23,19 +33,130 @@ import {
   FaLockOpen,
   FaPlay,
   FaArrowLeft,
+  FaDownload,
 } from "react-icons/fa";
 import {
   discover,
   getSettings,
   pair,
   setSettings,
+  checkUpdate,
   Host,
   StreamSettings,
+  UpdateInfo,
 } from "./backend";
 import { launchStream } from "./steam";
 
 const ROUTE = "/punktfunk";
 
+// Decky Loader exposes its already-authenticated WSRouter as a global. This is NOT part of
+// @decky/api (it's a loader internal), so we treat it as optional and guard every use — on a
+// loader without it we fall back to manual "Install Plugin from URL". We use it to drive
+// Decky's own privileged install path (the root loader does the download + SHA-256 verify +
+// extract + hot-reload), which is the only way a plugin can update itself: ~/homebrew/plugins
+// is root-owned, so our unprivileged backend can't swap its own files.
+declare global {
+  interface Window {
+    DeckyBackend?: {
+      callable: (route: string) => (...args: unknown[]) => Promise<unknown>;
+    };
+  }
+}
+
+// PluginInstallType.UPDATE in decky-loader's browser.py (INSTALL=0/REINSTALL=1/UPDATE=2/…).
+const INSTALL_TYPE_UPDATE = 2;
+
+// ----------------------------------------------------------------------------------------
+// Error boundary — contains ANY render failure in our UI so a single bad render can never take
+// down the whole Quick Access "Decky" section (Decky's tab-level boundary shows the generic
+// "Something went wrong while displaying this content" for the entire tab when one plugin
+// throws). The realistic trigger is a future Steam client update that makes a @decky/ui
+// component resolve to `undefined` (React then throws "Element type is invalid"). The fallback
+// is built from ONLY plain DOM elements + inline styles, so it cannot itself depend on a
+// (possibly broken) Steam-internal component — it is guaranteed to render.
+// ----------------------------------------------------------------------------------------
+class PluginErrorBoundary extends Component<
+  { children: ReactNode },
+  { error: Error | null }
+> {
+  state: { error: Error | null } = { error: null };
+
+  static getDerivedStateFromError(error: Error) {
+    return { error };
+  }
+
+  componentDidCatch(error: Error, info: ErrorInfo) {
+    // Surface it for diagnosis, but never rethrow — containment is the whole point.
+    // eslint-disable-next-line no-console
+    console.error("[punktfunk] contained UI render error:", error, info?.componentStack);
+  }
+
+  render() {
+    const { error } = this.state;
+    if (!error) return this.props.children;
+    return (
+      <div style={{ padding: "1em", lineHeight: 1.45 }}>
+        <div style={{ fontWeight: "bold", marginBottom: "0.4em" }}>
+          punktfunk couldn’t draw this view
+        </div>
+        <div style={{ opacity: 0.8, marginBottom: "0.6em" }}>
+          The plugin hit a display error — your Steam Deck is fine. Reload punktfunk from
+          Decky&apos;s plugin list, or update the plugin.
+        </div>
+        <div
+          style={{
+            opacity: 0.55,
+            fontFamily: "monospace",
+            fontSize: "0.8em",
+            wordBreak: "break-word",
+          }}
+        >
+          {String(error?.message ?? error)}
+        </div>
+      </div>
+    );
+  }
+}
+
+// Checks our registry for a newer build on mount (the backend caches + is non-fatal offline).
+function useUpdate() {
+  const [info, setInfo] = useState<UpdateInfo | null>(null);
+  useEffect(() => {
+    void checkUpdate(false)
+      .then(setInfo)
+      .catch(() => {});
+  }, []);
+  return info;
+}
+
+async function applyUpdate(info: UpdateInfo) {
+  try {
+    const backend = window.DeckyBackend;
+    if (backend?.callable) {
+      // Fire-and-forget: the loader reinstalls + reloads THIS plugin, tearing the panel down
+      // before any result could arrive — so never await it. Decky shows its own confirm prompt.
+      void backend.callable("utilities/install_plugin")(
+        info.artifact,
+        "punktfunk",
+        info.latest,
+        info.hash,
+        INSTALL_TYPE_UPDATE,
+      );
+      toaster.toast({
+        title: "punktfunk",
+        body: `Updating to v${info.latest}… confirm the Decky prompt.`,
+      });
+      return;
+    }
+  } catch {
+    // fall through to the manual path
+  }
+  toaster.toast({
+    title: "punktfunk",
+    body: "Update from Decky → Developer → Install Plugin from URL.",
+  });
+}
+
 // ----------------------------------------------------------------------------------------
 // Discovery hook — shared by the QAM panel and the full page.
 // ----------------------------------------------------------------------------------------
@@ -255,20 +376,24 @@ const SettingsSection: FC = () => {
 // One host row on the full page.
 // ----------------------------------------------------------------------------------------
 const HostRow: FC<{ host: Host }> = ({ host }) => {
-  const pairRequired = host.pair === "required";
+  // The host's policy is `pair=required`, but if THIS device is already paired we don't need to
+  // pair again — show it as trusted and go straight to Stream.
+  const needsPair = host.pair === "required" && !host.paired;
   return (
     <Field
       label={
         <span style={{ display: "inline-flex", alignItems: "center", gap: "0.4em" }}>
-          {pairRequired ? <FaLock /> : <FaLockOpen />}
+          {needsPair ? <FaLock /> : <FaLockOpen />}
           {host.name}
         </span>
       }
-      description={`${host.host}:${host.port}${pairRequired ? " · pairing required" : ""}`}
+      description={`${host.host}:${host.port}${
+        needsPair ? " · pairing required" : host.paired ? " · paired" : ""
+      }`}
       childrenContainerWidth="max"
     >
       <Focusable style={{ display: "flex", gap: "0.5em" }}>
-        {pairRequired && (
+        {needsPair && (
           <DialogButton
             style={{ minWidth: "5em" }}
             onClick={() =>
@@ -288,52 +413,129 @@ const HostRow: FC<{ host: Host }> = ({ host }) => {
 };
 
 // ----------------------------------------------------------------------------------------
-// The fullscreen page (registered as the /punktfunk route).
+// The fullscreen page (registered as the /punktfunk route) — a tabbed Hosts / Settings view.
 // ----------------------------------------------------------------------------------------
+
+// Bottom inset so the last control clears Gaming Mode's footer hint bar. Routed pages render
+// *under* that bar otherwise — that's why the last Stream-settings row was getting hidden. The
+// value is generous on purpose (and harmless where the tab area already insets); tune to taste.
+const SAFE_BOTTOM = "80px";
+
+// Each tab is its own scroll area so long content is always reachable above the footer.
+const tabScroll: CSSProperties = {
+  height: "100%",
+  overflowY: "auto",
+  padding: "0.5em 2.5em",
+  paddingBottom: SAFE_BOTTOM,
+  boxSizing: "border-box",
+};
+
+const HostsTab: FC<{
+  hosts: Host[];
+  scanning: boolean;
+  refresh: () => void;
+}> = ({ hosts, scanning, refresh }) => (
+  <div style={tabScroll}>
+    <Field
+      label="Discover"
+      description={
+        scanning
+          ? "Scanning the LAN…"
+          : `${hosts.length} host${hosts.length === 1 ? "" : "s"} on your network`
+      }
+      childrenContainerWidth="max"
+      bottomSeparator={hosts.length ? "standard" : "none"}
+    >
+      <DialogButton style={{ minWidth: "8em" }} disabled={scanning} onClick={refresh}>
+        {scanning ? (
+          <Spinner style={{ height: "1em", marginRight: "0.5em" }} />
+        ) : (
+          <FaSyncAlt style={{ marginRight: "0.5em" }} />
+        )}
+        {scanning ? "Scanning…" : "Refresh"}
+      </DialogButton>
+    </Field>
+
+    {hosts.length === 0 && !scanning && (
+      <Field
+        focusable={false}
+        description="No punktfunk hosts found. Make sure a host is running on the same network."
+      >
+        No hosts found
+      </Field>
+    )}
+    {hosts.map((h) => (
+      <HostRow key={h.fp || `${h.host}:${h.port}`} host={h} />
+    ))}
+  </div>
+);
+
+const SettingsTab: FC = () => (
+  <div style={tabScroll}>
+    <SettingsSection />
+  </div>
+);
+
 const PunktfunkPage: FC = () => {
   const { hosts, scanning, refresh } = useHosts();
+  const update = useUpdate();
+  const [tab, setTab] = useState("hosts");
 
   return (
     <div
       style={{
         marginTop: "40px",
         height: "calc(100% - 40px)",
-        overflowY: "auto",
-        padding: "0 2.5em 2.5em",
+        display: "flex",
+        flexDirection: "column",
       }}
     >
-      <Focusable style={{ display: "flex", alignItems: "center", gap: "1em", marginBottom: "1em" }}>
+      <Focusable
+        style={{
+          display: "flex",
+          alignItems: "center",
+          gap: "1em",
+          padding: "0 2.5em",
+          marginBottom: "0.4em",
+          flexShrink: 0,
+        }}
+      >
         <DialogButton
-          style={{ width: "3em", minWidth: "3em" }}
+          style={{ width: "3em", minWidth: "3em", padding: 0 }}
           onClick={() => Navigation.NavigateBack()}
         >
           <FaArrowLeft />
         </DialogButton>
-        <div className={staticClasses.Title} style={{ flex: 1 }}>
+        <div className={staticClasses?.Title} style={{ flex: 1, margin: 0 }}>
           punktfunk
         </div>
-        <DialogButton style={{ width: "10em" }} disabled={scanning} onClick={refresh}>
-          {scanning ? (
-            <Spinner style={{ height: "1em", marginRight: "0.5em" }} />
-          ) : (
-            <FaSyncAlt style={{ marginRight: "0.5em" }} />
-          )}
-          {scanning ? "Scanning…" : "Refresh"}
-        </DialogButton>
+        {update?.update_available && (
+          <DialogButton style={{ minWidth: "9em" }} onClick={() => applyUpdate(update)}>
+            <FaDownload style={{ marginRight: "0.4em" }} />
+            Update v{update.latest}
+          </DialogButton>
+        )}
       </Focusable>
 
-      <div style={{ fontSize: "1.1em", fontWeight: "bold", margin: "0.5em 0" }}>Hosts</div>
-      {hosts.length === 0 && !scanning && (
-        <Field focusable={false}>No hosts discovered on the LAN.</Field>
-      )}
-      {hosts.map((h) => (
-        <HostRow key={h.fp || `${h.host}:${h.port}`} host={h} />
-      ))}
-
-      <div style={{ fontSize: "1.1em", fontWeight: "bold", margin: "1.5em 0 0.5em" }}>
-        Stream settings
+      <div style={{ flex: 1, minHeight: 0 }}>
+        <Tabs
+          activeTab={tab}
+          onShowTab={(id: string) => setTab(id)}
+          autoFocusContents
+          tabs={[
+            {
+              id: "hosts",
+              title: "Hosts",
+              content: <HostsTab hosts={hosts} scanning={scanning} refresh={refresh} />,
+            },
+            {
+              id: "settings",
+              title: "Settings",
+              content: <SettingsTab />,
+            },
+          ]}
+        />
       </div>
-      <SettingsSection />
     </div>
   );
 };
@@ -343,9 +545,25 @@ const PunktfunkPage: FC = () => {
 // ----------------------------------------------------------------------------------------
 const QamPanel: FC = () => {
   const { hosts, scanning, refresh } = useHosts();
+  const update = useUpdate();
 
   return (
     <>
+      {update?.update_available && (
+        <PanelSection title="Update">
+          <PanelSectionRow>
+            <ButtonItem
+              layout="below"
+              onClick={() => applyUpdate(update)}
+              label={`v${update.current} → v${update.latest}`}
+            >
+              <FaDownload style={{ marginRight: "0.5em" }} />
+              Update punktfunk
+            </ButtonItem>
+          </PanelSectionRow>
+        </PanelSection>
+      )}
+
       <PanelSection title="punktfunk">
         <PanelSectionRow>
           <ButtonItem
@@ -378,25 +596,25 @@ const QamPanel: FC = () => {
           </PanelSectionRow>
         )}
         {hosts.map((h) => {
-          const pairRequired = h.pair === "required";
+          const needsPair = h.pair === "required" && !h.paired;
           return (
             <PanelSectionRow key={h.fp || `${h.host}:${h.port}`}>
               <ButtonItem
                 layout="below"
                 onClick={() =>
-                  pairRequired
+                  needsPair
                     ? showModal(<PairModal host={h} onPaired={() => startStream(h)} />)
                     : startStream(h)
                 }
                 label={
                   <span style={{ display: "inline-flex", alignItems: "center", gap: "0.4em" }}>
-                    {pairRequired ? <FaLock /> : <FaLockOpen />}
+                    {needsPair ? <FaLock /> : <FaLockOpen />}
                     {h.name}
                   </span>
                 }
-                description={`${h.host}:${h.port}`}
+                description={`${h.host}:${h.port}${h.paired ? " · paired" : ""}`}
               >
-                {pairRequired ? "Pair & Stream" : "Stream"}
+                {needsPair ? "Pair & Stream" : "Stream"}
               </ButtonItem>
             </PanelSectionRow>
           );
@@ -406,12 +624,25 @@ const QamPanel: FC = () => {
   );
 };
 
+// Full page behind the boundary — registered as the /punktfunk route.
+const PunktfunkRoute: FC = () => (
+  <PluginErrorBoundary>
+    <PunktfunkPage />
+  </PluginErrorBoundary>
+);
+
 export default definePlugin(() => {
-  routerHook.addRoute(ROUTE, PunktfunkPage, { exact: true });
+  routerHook.addRoute(ROUTE, PunktfunkRoute, { exact: true });
   return {
     name: "punktfunk",
-    titleView: <div className={staticClasses.Title}>punktfunk</div>,
-    content: <QamPanel />,
+    // `staticClasses?.Title` is guarded so a future client that drops the export can't throw
+    // at plugin-load time (an error boundary only catches render-time, not load-time, errors).
+    titleView: <div className={staticClasses?.Title}>punktfunk</div>,
+    content: (
+      <PluginErrorBoundary>
+        <QamPanel />
+      </PluginErrorBoundary>
+    ),
     icon: <FaTv />,
     onDismount() {
       routerHook.removeRoute(ROUTE);

clients/decky/src/steam.ts

@@ -24,12 +24,31 @@ declare const SteamClient: {
     SetShortcutExe(appId: number, exe: string): void;
     SetShortcutStartDir(appId: number, dir: string): void;
     SetAppLaunchOptions(appId: number, options: string): void;
-    SetAppHidden(appId: number, hidden: boolean): void;
     RunGame(gameId: string, _unused: string, _i: number, _j: number): void;
     TerminateApp(gameId: string, _b: boolean): void;
   };
 };
 
+// Steam removed `SteamClient.Apps.SetAppHidden`. Hiding a non-Steam shortcut now goes through
+// `collectionStore.SetAppsAsHidden([appId], true)` — but that looks the app up in appStore, which
+// only registers a freshly-created shortcut a moment later (calling it immediately throws on a
+// null overview). So hiding is BEST-EFFORT + DEFERRED and must NEVER block the launch.
+declare const collectionStore:
+  | { SetAppsAsHidden?: (appIds: number[], hidden: boolean) => void }
+  | undefined;
+
+function hideShortcut(appId: number): void {
+  const attempt = () => {
+    try {
+      collectionStore?.SetAppsAsHidden?.([appId], true);
+    } catch {
+      /* overview not registered yet, or the API changed — cosmetic, ignore */
+    }
+  };
+  attempt(); // succeeds immediately for an already-registered (reused) shortcut
+  setTimeout(attempt, 2500); // fresh shortcut: retry once its app overview lands
+}
+
 const SHORTCUT_NAME = "punktfunk";
 
 // The 64-bit "gameid" RunGame wants, derived from a 32-bit non-Steam shortcut appId: the
@@ -88,7 +107,8 @@ async function ensureShortcut(): Promise<number> {
   );
   SteamClient.Apps.SetShortcutName(appId, SHORTCUT_NAME);
   // Hide it from the library — it's an implementation detail, launched programmatically.
-  SteamClient.Apps.SetAppHidden(appId, true);
+  // Best-effort + deferred (see hideShortcut); never let it block the launch.
+  hideShortcut(appId);
   rememberAppId(appId);
   return appId;
 }

clients/linux/src/app.rs

@@ -22,6 +22,8 @@ struct App {
     gamepad: crate::gamepad::GamepadService,
     /// One session at a time — ignore connects while one is starting/running.
     busy: std::cell::Cell<bool>,
+    /// Steam Deck / Gaming-Mode launch: fullscreen the window (chrome-less) when a stream starts.
+    fullscreen: bool,
 }
 
 impl App {
@@ -41,7 +43,13 @@ pub fn run() -> glib::ExitCode {
     if let Some(pin) = arg_value("--pair") {
         return headless_pair(&pin);
     }
-    let app = adw::Application::builder().application_id(APP_ID).build();
+    let mut builder = adw::Application::builder().application_id(APP_ID);
+    // Screenshot mode launches the app once per scene back-to-back; NON_UNIQUE keeps each
+    // launch its own primary instance instead of forwarding to a still-registered name.
+    if shot_scene().is_some() {
+        builder = builder.flags(gtk::gio::ApplicationFlags::NON_UNIQUE);
+    }
+    let app = builder.build();
     app.connect_activate(build_ui);
     // GTK doesn't see our argv (`--connect` is handled in `build_ui`); an empty argv also
     // keeps GApplication from rejecting unknown options.
@@ -56,6 +64,20 @@ fn arg_value(flag: &str) -> Option<String> {
         .filter(|v| !v.starts_with("--"))
 }
 
+/// True if argv contains `flag` (a valueless switch).
+fn arg_flag(flag: &str) -> bool {
+    std::env::args().any(|a| a == flag)
+}
+
+/// Run the stream fullscreen with no window chrome — the Steam Deck / Gaming-Mode launch path.
+/// The Decky wrapper passes `--fullscreen`; we also honor the Deck/gamescope env as a fallback
+/// so a manual launch under Gaming Mode does the right thing too.
+fn fullscreen_mode() -> bool {
+    arg_flag("--fullscreen")
+        || std::env::var_os("SteamDeck").is_some()
+        || std::env::var_os("GAMESCOPE_WAYLAND_DISPLAY").is_some()
+}
+
 /// Run the SPAKE2 PIN ceremony without a GTK window and persist the verified host to the
 /// known-hosts store as paired, so a later `--connect` connects silently. Same identity
 /// store the streaming path uses (same binary), so pairing here makes the stream work.
@@ -161,6 +183,7 @@ fn build_ui(gtk_app: &adw::Application) {
         identity,
         gamepad: crate::gamepad::GamepadService::start(),
         busy: std::cell::Cell::new(false),
+        fullscreen: fullscreen_mode(),
     });
 
     let hosts_page = crate::ui_hosts::new(
@@ -182,11 +205,65 @@ fn build_ui(gtk_app: &adw::Application) {
     nav.add(&hosts_page);
     window.present();
 
+    // CI screenshot mode: render one scripted, host-free scene and signal readiness
+    // (clients/linux/tools/screenshots.sh). Mutually exclusive with a real connect.
+    if let Some(scene) = shot_scene() {
+        run_shot(app, &scene);
+        return;
+    }
+
     if let Some(req) = cli_connect_request() {
         initiate_connect(app, req);
     }
 }
 
+/// `PUNKTFUNK_SHOT_SCENE`, when set, selects a scripted host-free scene for CI screenshots.
+fn shot_scene() -> Option<String> {
+    std::env::var("PUNKTFUNK_SHOT_SCENE")
+        .ok()
+        .filter(|s| !s.is_empty())
+}
+
+/// Render one mock-populated, host-free scene over the already-presented window, then print
+/// `PF_SHOT_READY` once it has had a moment to map + settle so the driver knows when to capture.
+/// No `NativeClient` or session is created. The stream scene is deliberately absent — its page
+/// requires a live connector (`ui_stream::new` takes an `Arc<NativeClient>`).
+fn run_shot(app: Rc<App>, scene: &str) {
+    // A plausible host for the trust/pair dialogs (fp_hex is 64 hex chars, like a real SHA-256).
+    let mock_req = || ConnectRequest {
+        name: "Living Room PC".to_string(),
+        addr: "192.168.1.42".to_string(),
+        port: 9777,
+        fp_hex: Some(
+            "9f8e7d6c5b4a39281706f5e4d3c2b1a0998877665544332211ffeeddccbbaa00".to_string(),
+        ),
+        pair_optional: true,
+    };
+
+    match scene {
+        // The saved-hosts grid reads ~/.config/punktfunk/client-known-hosts.json, which the
+        // driver seeds — so the already-shown hosts page is the scene; nothing to do here.
+        "hosts" | "02-hosts" => {}
+        "settings" | "03-settings" => {
+            crate::ui_settings::show(&app.window, app.settings.clone(), &app.gamepad);
+        }
+        "trust" | "04-trust" => tofu_dialog(app.clone(), mock_req()),
+        "pair" | "05-pair" => pin_dialog(app.clone(), mock_req()),
+        other => tracing::warn!("unknown PUNKTFUNK_SHOT_SCENE={other:?}; showing hosts only"),
+    }
+
+    let settle_ms = std::env::var("PUNKTFUNK_SHOT_SETTLE_MS")
+        .ok()
+        .and_then(|v| v.parse().ok())
+        .unwrap_or(900);
+    let scene = scene.to_string();
+    glib::timeout_add_local_once(std::time::Duration::from_millis(settle_ms), move || {
+        use std::io::Write as _;
+        println!("PF_SHOT_READY scene={scene}");
+        let _ = std::io::stdout().flush();
+    });
+}
+
 /// The trust gate in front of every connect. The host is the policy authority (it
 /// advertises `pair=optional` only when it accepts unpaired clients); the client renders
 /// its trust UI from that:
@@ -375,6 +452,7 @@ fn speed_test(app: Rc<App>, req: ConnectRequest) {
                 GamepadPref::Auto,
                 0,    // bitrate_kbps (host default)
                 0,    // video_caps: the Linux client has no 10-bit/HDR present path yet
+                2,    // audio_channels: speed-test probe, stereo
                 None, // launch: speed-test probe connect, no game
                 pin,
                 Some(identity),
@@ -443,11 +521,19 @@ fn resolve_mode(app: &App) -> punktfunk_core::config::Mode {
         refresh_hz: s.refresh_hz,
     };
     if mode.width == 0 || mode.refresh_hz == 0 {
+        // Prefer the monitor the window is on; fall back to the display's first monitor. On a
+        // `--connect` launch the window may not be mapped yet when this runs, and without the
+        // fallback we'd drop to the 1920×1080 floor below — wrong on the Deck (1280×800).
         let monitor = app
             .window
             .surface()
             .zip(gdk::Display::default())
-            .and_then(|(surf, d)| d.monitor_at_surface(&surf));
+            .and_then(|(surf, d)| d.monitor_at_surface(&surf))
+            .or_else(|| {
+                gdk::Display::default()
+                    .and_then(|d| d.monitors().item(0))
+                    .and_then(|o| o.downcast::<gdk::Monitor>().ok())
+            });
         if let Some(m) = monitor {
             let geo = m.geometry();
             let scale = m.scale_factor().max(1);
@@ -488,6 +574,7 @@ fn start_session(app: Rc<App>, req: ConnectRequest, pin: Option<[u8; 32]>) {
         },
         bitrate_kbps: s.bitrate_kbps,
         mic_enabled: s.mic_enabled,
+        audio_channels: s.audio_channels,
         pin,
         identity: app.identity.clone(),
     };
@@ -540,6 +627,12 @@ fn start_session(app: Rc<App>, req: ConnectRequest, pin: Option<[u8; 32]>) {
                         &title,
                     );
                     app.nav.push(&p.page);
+                    // Steam Deck / Gaming Mode: gamescope fullscreens the window but GTK doesn't
+                    // know it, so its header bar stays drawn. Enter GTK fullscreen explicitly —
+                    // the stream page's `connect_fullscreened_notify` then hides all chrome.
+                    if app.fullscreen {
+                        app.window.fullscreen();
+                    }
                     page = Some(p);
                 }
                 SessionEvent::Stats(s) => {

clients/linux/src/audio.rs

@@ -27,16 +27,17 @@ pub struct AudioPlayer {
 }
 
 impl AudioPlayer {
-    /// Spawn the PipeWire playback thread. Failure (no PipeWire in the session) is
-    /// survivable — the caller streams video-only.
-    pub fn spawn() -> Result<AudioPlayer> {
+    /// Spawn the PipeWire playback thread for `channels` (2/6/8, canonical wire order
+    /// FL FR FC LFE RL RR SL SR). Failure (no PipeWire in the session) is survivable — the
+    /// caller streams video-only.
+    pub fn spawn(channels: u32) -> Result<AudioPlayer> {
         // 64 × 5 ms = 320 ms of slack between the pump and the PipeWire loop.
         let (pcm_tx, pcm_rx) = std::sync::mpsc::sync_channel::<Vec<f32>>(64);
         let (quit_tx, quit_rx) = pipewire::channel::channel::<Terminate>();
         let thread = std::thread::Builder::new()
             .name("punktfunk-audio".into())
             .spawn(move || {
-                if let Err(e) = pw_thread(pcm_rx, quit_rx) {
+                if let Err(e) = pw_thread(pcm_rx, quit_rx, channels as usize) {
                     tracing::warn!(error = %e, "audio playback thread ended");
                 }
             })
@@ -48,8 +49,8 @@ impl AudioPlayer {
         })
     }
 
-    /// Queue one interleaved-stereo f32 chunk. Drops the chunk if the PipeWire side is
-    /// wedged (the renderer conceals the gap; never block the session pump).
+    /// Queue one interleaved f32 chunk (in the session's channel layout). Drops the chunk if the
+    /// PipeWire side is wedged (the renderer conceals the gap; never block the session pump).
     pub fn push(&self, pcm: Vec<f32>) {
         if let Err(TrySendError::Disconnected(_)) = self.pcm_tx.try_send(pcm) {
             // Thread already dead — Drop will reap it; nothing to do per-chunk.
@@ -71,11 +72,14 @@ struct PlayerData {
     rx: Receiver<Vec<f32>>,
     ring: VecDeque<f32>,
     primed: bool,
+    /// Interleaved channel count this stream was opened with (2/6/8).
+    channels: usize,
 }
 
 fn pw_thread(
     pcm_rx: Receiver<Vec<f32>>,
     quit_rx: pipewire::channel::Receiver<Terminate>,
+    channels: usize,
 ) -> Result<()> {
     use pipewire as pw;
     use pw::{properties::properties, spa};
@@ -115,6 +119,7 @@ fn pw_thread(
         rx: pcm_rx,
         ring: VecDeque::new(),
         primed: false,
+        channels,
     };
 
     let _listener = stream
@@ -130,19 +135,19 @@ fn pw_thread(
                 while let Ok(chunk) = ud.rx.try_recv() {
                     ud.ring.extend(chunk);
                 }
-                let stride = 4 * CHANNELS; // F32LE interleaved
+                let stride = 4 * ud.channels; // F32LE interleaved
                 let datas = buffer.datas_mut();
                 if datas.is_empty() {
                     return;
                 }
                 let data = &mut datas[0];
                 let want_frames = data.data().map(|s| s.len() / stride).unwrap_or(0);
-                let want = want_frames * CHANNELS;
+                let want = want_frames * ud.channels;
 
                 // Adaptive jitter buffer (same shape as the host's virtual mic): prime to
                 // ~3 quanta, cap at ~1 quantum of slack beyond that, re-prime after a
                 // genuine drain.
-                let target = (3 * want).clamp(720 * CHANNELS, 9600 * CHANNELS);
+                let target = (3 * want).clamp(720 * ud.channels, 9600 * ud.channels);
                 while ud.ring.len() > target.max(want) + want {
                     ud.ring.pop_front();
                 }
@@ -182,7 +187,13 @@ fn pw_thread(
     let mut info = AudioInfoRaw::new();
     info.set_format(AudioFormat::F32LE);
     info.set_rate(SAMPLE_RATE);
-    info.set_channels(CHANNELS as u32);
+    info.set_channels(channels as u32);
+    // Channel positions in canonical wire order (FL FR FC LFE RL RR SL SR) so PipeWire routes each
+    // slot to the matching speaker (and downmixes when the sink has fewer). Identity, no permute.
+    let order = punktfunk_core::audio::spa_positions(channels as u8);
+    let mut positions = [0u32; 64];
+    positions[..order.len()].copy_from_slice(order);
+    info.set_position(positions);
     let obj = pw::spa::pod::Object {
         type_: pw::spa::utils::SpaTypes::ObjectParamFormat.as_raw(),
         id: pw::spa::param::ParamType::EnumFormat.as_raw(),

clients/linux/src/session.rs

@@ -20,6 +20,8 @@ pub struct SessionParams {
     pub compositor: CompositorPref,
     pub gamepad: GamepadPref,
     pub bitrate_kbps: u32,
+    /// Requested audio channel count (2/6/8); the host echoes the resolved value.
+    pub audio_channels: u8,
     /// Stream the default microphone to the host's virtual mic source.
     pub mic_enabled: bool,
     /// Pinned host fingerprint; `None` = trust on first use (caller persists the observed one).
@@ -83,6 +85,42 @@ fn now_ns() -> u64 {
         .unwrap_or(0)
 }
 
+/// Opus decoder for the audio plane: a plain stereo decoder (the validated path) or a multistream
+/// decoder for 5.1/7.1, both behind one `decode_float`. Built from the host-RESOLVED channel count
+/// via the shared layout table.
+enum AudioDec {
+    Stereo(opus::Decoder),
+    Surround(opus::MSDecoder),
+}
+
+impl AudioDec {
+    fn new(channels: u8) -> Result<AudioDec, opus::Error> {
+        if channels == 2 {
+            Ok(AudioDec::Stereo(opus::Decoder::new(
+                48_000,
+                opus::Channels::Stereo,
+            )?))
+        } else {
+            let l = punktfunk_core::audio::layout_for(channels, false);
+            Ok(AudioDec::Surround(opus::MSDecoder::new(
+                48_000, l.streams, l.coupled, l.mapping,
+            )?))
+        }
+    }
+
+    fn decode_float(
+        &mut self,
+        input: &[u8],
+        out: &mut [f32],
+        fec: bool,
+    ) -> Result<usize, opus::Error> {
+        match self {
+            AudioDec::Stereo(d) => d.decode_float(input, out, fec),
+            AudioDec::Surround(d) => d.decode_float(input, out, fec),
+        }
+    }
+}
+
 fn pump(
     params: SessionParams,
     ev_tx: async_channel::Sender<SessionEvent>,
@@ -96,7 +134,8 @@ fn pump(
         params.compositor,
         params.gamepad,
         params.bitrate_kbps,
-        0,    // video_caps: the Linux client has no 10-bit/HDR present path yet
+        0, // video_caps: the Linux client has no 10-bit/HDR present path yet
+        params.audio_channels,
         None, // launch: the Linux client has no library picker yet
         params.pin,
         Some(params.identity),
@@ -134,11 +173,14 @@ fn pump(
         }
     };
     // Audio is best-effort: a session without it still streams. Gamepads are the
-    // app-lifetime service's job (the UI attaches it on Connected).
-    let player = audio::AudioPlayer::spawn()
+    // app-lifetime service's job (the UI attaches it on Connected). Build the decoder + playback
+    // from the host-RESOLVED channel count (never the request), so an older/clamping host that
+    // resolves stereo is decoded as stereo.
+    let channels = connector.audio_channels;
+    let player = audio::AudioPlayer::spawn(channels as u32)
         .map_err(|e| tracing::warn!(error = %e, "audio disabled"))
         .ok();
-    let mut opus_dec = opus::Decoder::new(48_000, opus::Channels::Stereo)
+    let mut opus_dec = AudioDec::new(channels)
         .map_err(|e| tracing::warn!(error = %e, "opus decoder failed — audio disabled"))
         .ok();
     let _mic = params
@@ -157,8 +199,8 @@ fn pump(
     let mut bytes_n = 0u64;
     let mut decode_us_sum = 0u64;
     let mut lat_us: Vec<u64> = Vec::with_capacity(256);
-    let mut pcm = vec![0f32; 5760 * 2]; // decode scratch: max Opus frame (120 ms stereo)
-                                        // Loss recovery: watch the host→client unrecoverable-drop count and ask for an IDR when it climbs.
+    let mut pcm = vec![0f32; 5760 * channels as usize]; // scratch: max Opus frame (120 ms) × channels
+                                                        // Loss recovery: watch the host→client unrecoverable-drop count and ask for an IDR when it climbs.
     let mut last_dropped = connector.frames_dropped();
     let mut last_kf_req: Option<Instant> = None;
 
@@ -221,7 +263,8 @@ fn pump(
         while let Ok(pkt) = connector.next_audio(Duration::ZERO) {
             if let (Some(player), Some(dec)) = (&player, opus_dec.as_mut()) {
                 match dec.decode_float(&pkt.data, &mut pcm, false) {
-                    Ok(samples) => player.push(pcm[..samples * 2].to_vec()),
+                    // `samples` is per-channel; the interleaved frame is `samples * channels`.
+                    Ok(samples) => player.push(pcm[..samples * channels as usize].to_vec()),
                     Err(e) => tracing::debug!(error = %e, "opus decode"),
                 }
             }

clients/linux/src/trust.rs

@@ -90,6 +90,14 @@ impl KnownHosts {
         self.hosts.iter().find(|h| h.addr == addr && h.port == port)
     }
 
+    /// Forget the entry with this fingerprint. Returns true if one was removed (the user
+    /// will have to pair/trust again to reconnect).
+    pub fn remove_by_fp(&mut self, fp_hex: &str) -> bool {
+        let before = self.hosts.len();
+        self.hosts.retain(|h| h.fp_hex != fp_hex);
+        self.hosts.len() != before
+    }
+
     /// Insert or refresh an entry, keyed by fingerprint. `paired` only ever upgrades
     /// (a later TOFU connect must not demote a PIN-paired host).
     pub fn upsert(&mut self, entry: KnownHost) {
@@ -124,6 +132,9 @@ pub struct Settings {
     pub inhibit_shortcuts: bool,
     /// Stream the default microphone to the host's virtual mic source.
     pub mic_enabled: bool,
+    /// Requested audio channel count: 2 (stereo), 6 (5.1) or 8 (7.1). The host clamps to what it
+    /// can capture; the resolved count drives the decoder + playback layout.
+    pub audio_channels: u8,
 }
 
 impl Default for Settings {
@@ -137,6 +148,7 @@ impl Default for Settings {
             compositor: "auto".into(),
             inhibit_shortcuts: true,
             mic_enabled: false,
+            audio_channels: 2,
         }
     }
 }

clients/linux/src/ui_hosts.rs

@@ -181,6 +181,52 @@ pub fn new(
                     // pinned connect; TOFU eligibility is irrelevant.
                     pair_optional: false,
                 };
+                // Forget this host (drops the pinned fingerprint — a later connect re-pairs).
+                // Confirmed first, since it's destructive and a misclick on the Deck is easy.
+                let remove_btn = gtk::Button::from_icon_name("user-trash-symbolic");
+                remove_btn.set_tooltip_text(Some("Remove saved host"));
+                remove_btn.set_valign(gtk::Align::Center);
+                remove_btn.add_css_class("flat");
+                {
+                    let fp = k.fp_hex.clone();
+                    let name = k.name.clone();
+                    let saved_list = saved_list.clone();
+                    let saved_label = saved_label.clone();
+                    let row = row.clone();
+                    remove_btn.connect_clicked(move |_| {
+                        let dialog = adw::AlertDialog::new(
+                            Some("Remove saved host?"),
+                            Some(&format!(
+                                "Forget “{name}”? You'll need to pair (or trust) it again to reconnect."
+                            )),
+                        );
+                        dialog.add_responses(&[("cancel", "Cancel"), ("remove", "Remove")]);
+                        dialog.set_response_appearance(
+                            "remove",
+                            adw::ResponseAppearance::Destructive,
+                        );
+                        dialog.set_default_response(Some("cancel"));
+                        dialog.set_close_response("cancel");
+                        {
+                            // Scoped clones for the response handler so `row` survives for present().
+                            let fp = fp.clone();
+                            let saved_list = saved_list.clone();
+                            let saved_label = saved_label.clone();
+                            let row = row.clone();
+                            dialog.connect_response(Some("remove"), move |_, _| {
+                                let mut known = KnownHosts::load();
+                                known.remove_by_fp(&fp);
+                                let _ = known.save();
+                                saved_list.remove(&row);
+                                let empty = known.hosts.is_empty();
+                                saved_list.set_visible(!empty);
+                                saved_label.set_visible(!empty);
+                            });
+                        }
+                        dialog.present(Some(&row));
+                    });
+                }
+                row.add_suffix(&remove_btn);
                 let speed_btn = gtk::Button::from_icon_name("network-transmit-receive-symbolic");
                 speed_btn.set_tooltip_text(Some("Test network speed"));
                 speed_btn.set_valign(gtk::Align::Center);

clients/linux/src/ui_settings.rs

@@ -140,6 +140,16 @@ pub fn show(
     input.add(&inhibit_row);
 
     let audio = adw::PreferencesGroup::builder().title("Audio").build();
+    let surround_row = adw::ComboRow::builder()
+        .title("Audio channels")
+        .subtitle("Request stereo or surround (the host downmixes if its output has fewer)")
+        .model(&gtk::StringList::new(&[
+            "Stereo",
+            "5.1 Surround",
+            "7.1 Surround",
+        ]))
+        .build();
+    audio.add(&surround_row);
     let mic_row = adw::SwitchRow::builder()
         .title("Stream microphone")
         .subtitle("Send the default input device to the host's virtual microphone")
@@ -170,6 +180,11 @@ pub fn show(
         compositor_row.set_selected(comp_i as u32);
         inhibit_row.set_active(s.inhibit_shortcuts);
         mic_row.set_active(s.mic_enabled);
+        surround_row.set_selected(match s.audio_channels {
+            6 => 1,
+            8 => 2,
+            _ => 0,
+        });
     }
 
     let dialog = adw::PreferencesDialog::new();
@@ -186,6 +201,11 @@ pub fn show(
             .to_string();
         s.inhibit_shortcuts = inhibit_row.is_active();
         s.mic_enabled = mic_row.is_active();
+        s.audio_channels = match surround_row.selected() {
+            1 => 6,
+            2 => 8,
+            _ => 2,
+        };
         s.save();
     });
     dialog.present(Some(parent));

clients/linux/tools/screenshots.sh

@@ -0,0 +1,123 @@
+#!/usr/bin/env bash
+# Capture host-free UI screenshots of the native Linux client under a virtual X
+# display. Mirrors the iOS harness (clients/apple/tools/screenshots.sh): one app
+# launch per scene (PUNKTFUNK_SHOT_SCENE), the app renders a mock-populated REAL
+# view and prints `PF_SHOT_READY`, then we grab the X root window. No host, GPU, or
+# live stream — only the chrome scenes (the stream page needs a live connector).
+#
+#   cargo build --release -p punktfunk-client-linux
+#   bash clients/linux/tools/screenshots.sh            # → clients/linux/screenshots/<scene>.png
+#   bash clients/linux/tools/screenshots.sh hosts pair # a subset
+#
+# Env knobs: BIN (client binary), OUT (output dir), GEOMETRY (Xvfb WxHxDepth),
+# SETTLE (extra seconds after PF_SHOT_READY), SHOT_DISPLAY (X display), GSK_RENDERER
+# (gl|ngl|cairo — gl/llvmpipe by default for full libadwaita fidelity).
+set -euo pipefail
+
+here="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" # clients/linux
+BIN="${BIN:-$here/../../target/release/punktfunk-client}"
+OUT="${OUT:-$here/screenshots}"
+# The client window maps at its 1100x720 default; with no WM under Xvfb it lands at the
+# top-left, so keep the root just larger so the full window (incl. its CSD shadow) is
+# captured by a root grab with only a thin margin to crop.
+GEOMETRY="${GEOMETRY:-1280x800x24}"
+SETTLE="${SETTLE:-1.2}"
+SHOT_DISPLAY="${SHOT_DISPLAY:-:99}"
+
+if [ "$#" -gt 0 ]; then SCENES=("$@"); else SCENES=(hosts settings trust pair); fi
+
+[ -x "$BIN" ] || {
+	echo "client binary not found: $BIN (build it first: cargo build --release -p punktfunk-client-linux)" >&2
+	exit 1
+}
+
+# Isolated scratch HOME: the client generates its identity here on first run, and the
+# saved-hosts grid is read from client-known-hosts.json, so seed mock hosts for the
+# `hosts` scene (the dialogs/settings build their own mock state in-app).
+WORK="$(mktemp -d)"
+export HOME="$WORK"
+mkdir -p "$HOME/.config/punktfunk"
+cat >"$HOME/.config/punktfunk/client-known-hosts.json" <<'JSON'
+{
+  "hosts": [
+    { "name": "Living Room PC", "addr": "192.168.1.42", "port": 9777,
+      "fp_hex": "9f8e7d6c5b4a39281706f5e4d3c2b1a0998877665544332211ffeeddccbbaa00",
+      "paired": true },
+    { "name": "Office", "addr": "192.168.1.50", "port": 9777,
+      "fp_hex": "a1b2c3d4e5f60718293a4b5c6d7e8f90112233445566778899aabbccddeeff00",
+      "paired": false }
+  ]
+}
+JSON
+
+# Software-rendered X session — no GPU/Wayland. GL/llvmpipe runs the real NGL renderer
+# (cairo is documented-incomplete for 3D-transformed content / libadwaita transitions).
+unset WAYLAND_DISPLAY
+export DISPLAY="$SHOT_DISPLAY"
+export GDK_BACKEND=x11
+export LIBGL_ALWAYS_SOFTWARE=1
+export GALLIUM_DRIVER="${GALLIUM_DRIVER:-llvmpipe}"
+export GSK_RENDERER="${GSK_RENDERER:-gl}"
+
+Xvfb "$SHOT_DISPLAY" -screen 0 "$GEOMETRY" -nolisten tcp >"$WORK/xvfb.log" 2>&1 &
+XVFB_PID=$!
+cleanup() {
+	kill "$XVFB_PID" 2>/dev/null || true
+	rm -rf "$WORK"
+}
+trap cleanup EXIT
+
+# Wait for the display to accept connections.
+for _ in $(seq 1 50); do
+	if command -v xdpyinfo >/dev/null 2>&1; then
+		xdpyinfo -display "$SHOT_DISPLAY" >/dev/null 2>&1 && break
+	else
+		[ -e "/tmp/.X11-unix/X${SHOT_DISPLAY#:}" ] && break
+	fi
+	sleep 0.1
+done
+
+capture() {
+	local out="$1"
+	if command -v import >/dev/null 2>&1; then
+		import -silent -window root "$out"
+	elif command -v scrot >/dev/null 2>&1; then
+		scrot -o "$out"
+	else
+		echo "no screenshot tool — install imagemagick or scrot" >&2
+		return 1
+	fi
+}
+
+mkdir -p "$OUT"
+rc=0
+for scene in "${SCENES[@]}"; do
+	: >"$WORK/log"
+	PUNKTFUNK_SHOT_SCENE="$scene" "$BIN" >"$WORK/log" 2>&1 &
+	pid=$!
+	ready=0
+	for _ in $(seq 1 200); do # up to ~20s
+		if grep -q "PF_SHOT_READY" "$WORK/log"; then
+			ready=1
+			break
+		fi
+		if ! kill -0 "$pid" 2>/dev/null; then break; fi
+		sleep 0.1
+	done
+	if [ "$ready" = 1 ]; then
+		sleep "$SETTLE"
+		if capture "$OUT/$scene.png"; then
+			echo "✓ $scene → $OUT/$scene.png"
+		else
+			rc=1
+		fi
+	else
+		echo "✗ $scene: client never signalled PF_SHOT_READY" >&2
+		sed 's/^/    /' "$WORK/log" >&2 || true
+		rc=1
+	fi
+	kill "$pid" 2>/dev/null || true
+	wait "$pid" 2>/dev/null || true
+done
+
+exit "$rc"

clients/probe/Cargo.toml

@@ -18,8 +18,7 @@ tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 # LAN host discovery (`--discover`): browse the native `_punktfunk._udp` mDNS service the host
 # advertises (same crate/version the host advertises with).
 mdns-sd = "0.20"
-
-# Linux-only: --mic-test's Opus encoder (libopus). The mic UPLINK itself is portable —
-# only this synthetic-tone test rig needs the encoder.
-[target.'cfg(target_os = "linux")'.dependencies]
+# Opus: multistream DECODE of the host's audio plane (the surround validator) + `--mic-test`'s
+# encoder. libopus is already in the graph via `punktfunk-core`'s quic feature; this exposes the
+# name directly. Cross-platform (cmake-vendored), so the probe builds + validates everywhere.
 opus = "0.3"

clients/probe/src/main.rs

@@ -78,6 +78,10 @@ struct Args {
     gamepad: GamepadPref,
     /// `--bitrate KBPS` — request this encoder bitrate (kilobits/s); 0 = host default.
     bitrate_kbps: u32,
+    /// `--audio-channels N` — request stereo (2), 5.1 (6) or 7.1 (8) audio; default 2. The probe
+    /// multistream-decodes the host's frames and asserts the per-channel sample count, so it's the
+    /// headless validator for the surround encode path.
+    audio_channels: u8,
     /// `--launch ID` — ask the host to launch a library title in this session (a store-qualified
     /// id from the host's `GET /api/v1/library`, e.g. `steam:570`). Host resolves it; `None` = none.
     launch: Option<String>,
@@ -201,6 +205,11 @@ fn parse_args() -> Args {
         compositor,
         gamepad,
         bitrate_kbps: get("--bitrate").and_then(|s| s.parse().ok()).unwrap_or(0),
+        audio_channels: punktfunk_core::audio::normalize_channels(
+            get("--audio-channels")
+                .and_then(|s| s.parse().ok())
+                .unwrap_or(2),
+        ),
         launch: get("--launch").map(str::to_string),
         speed_test: get("--speed-test").and_then(|s| {
             let (kbps, ms) = s.split_once(':')?;
@@ -385,13 +394,23 @@ async fn session(args: Args) -> Result<()> {
             // `--launch ID` — host resolves it against its own library and runs it this session.
             launch: args.launch.clone(),
             // This headless tool just dumps the bitstream (no decode), so it can always claim
-            // 10-bit support. Gated by env so latency runs stay on the 8-bit baseline:
-            // PUNKTFUNK_CLIENT_10BIT=1 advertises VIDEO_CAP_10BIT to exercise the host Main10 path.
-            video_caps: if std::env::var_os("PUNKTFUNK_CLIENT_10BIT").is_some() {
-                punktfunk_core::quic::VIDEO_CAP_10BIT
-            } else {
-                0
+            // 10-bit / 4:4:4 support. Gated by env so latency runs stay on the 8-bit 4:2:0 baseline:
+            // PUNKTFUNK_CLIENT_10BIT=1 advertises VIDEO_CAP_10BIT (host Main10 path);
+            // PUNKTFUNK_CLIENT_444=1 advertises VIDEO_CAP_444 (host HEVC 4:4:4 path) — verify the
+            // resulting chroma with `ffprobe` on the `--out` .h265.
+            video_caps: {
+                let mut caps = 0u8;
+                if std::env::var_os("PUNKTFUNK_CLIENT_10BIT").is_some() {
+                    caps |= punktfunk_core::quic::VIDEO_CAP_10BIT;
+                }
+                if std::env::var_os("PUNKTFUNK_CLIENT_444").is_some() {
+                    caps |= punktfunk_core::quic::VIDEO_CAP_444;
+                }
+                caps
             },
+            // `--audio-channels` (default stereo); the probe multistream-decodes + validates the
+            // host's frames to exercise the surround encode path headlessly.
+            audio_channels: args.audio_channels,
         }
         .encode(),
     )
@@ -408,6 +427,8 @@ async fn session(args: Args) -> Result<()> {
         bit_depth = welcome.bit_depth,
         color = ?welcome.color,
         hdr = welcome.color.is_hdr(),
+        chroma_444 = welcome.chroma_format == punktfunk_core::quic::CHROMA_IDC_444,
+        chroma_format_idc = welcome.chroma_format,
         "session offer"
     );
 
@@ -830,13 +851,37 @@ async fn session(args: Args) -> Result<()> {
             hidout_pkts.clone(),
         );
         let conn2 = conn.clone();
+        // Build a multistream decoder for the host-RESOLVED layout so the probe actually decodes
+        // the surround stream (not just counts bytes) — the headless validator for the encode path.
+        let audio_channels = welcome.audio_channels;
         tokio::spawn(async move {
             use std::sync::atomic::Ordering::Relaxed;
             let mut hdr_logged = false;
+            let layout = punktfunk_core::audio::layout_for(audio_channels, false);
+            let mut audio_dec =
+                opus::MSDecoder::new(48_000, layout.streams, layout.coupled, layout.mapping).ok();
+            let mut pcm = vec![0f32; 5760 * audio_channels as usize];
+            let mut audio_decoded_logged = false;
             while let Ok(d) = conn2.read_datagram().await {
                 if let Some((_, _, opus)) = punktfunk_core::quic::decode_audio_datagram(&d) {
                     a.fetch_add(1, Relaxed);
                     ab.fetch_add(opus.len() as u64, Relaxed);
+                    // Decode + validate: the per-channel sample count must be a legal Opus frame
+                    // size; log the first success so a loopback test can assert surround decoded.
+                    if let Some(dec) = audio_dec.as_mut() {
+                        match dec.decode_float(opus, &mut pcm, false) {
+                            Ok(samples) if !audio_decoded_logged => {
+                                audio_decoded_logged = true;
+                                tracing::info!(
+                                    channels = audio_channels,
+                                    samples_per_channel = samples,
+                                    "audio decoded (Opus multistream)"
+                                );
+                            }
+                            Ok(_) => {}
+                            Err(e) => tracing::debug!(error = %e, "probe audio decode"),
+                        }
+                    }
                 } else if punktfunk_core::quic::decode_rumble_datagram(&d).is_some() {
                     r.fetch_add(1, Relaxed);
                 } else if let Some(meta) = punktfunk_core::quic::decode_hdr_meta_datagram(&d) {

clients/windows/src/app.rs

@@ -39,6 +39,9 @@ const DECODERS: &[(&str, &str)] = &[
 ];
 /// Bitrate presets in Mb/s; `0` = host default.
 const BITRATES_MBPS: &[u32] = &[0, 10, 20, 30, 50, 80, 150];
+/// Audio channel presets: `(channel count, display label)`. The host clamps to what it can
+/// capture; the resolved count drives the decoder + WASAPI render layout.
+const AUDIO_CHANNELS: &[(u8, &str)] = &[(2, "Stereo"), (6, "5.1 Surround"), (8, "7.1 Surround")];
 
 #[derive(Clone, PartialEq)]
 enum Screen {
@@ -598,6 +601,7 @@ fn connect(
         compositor: CompositorPref::Auto,
         gamepad: gamepad_pref,
         bitrate_kbps: s.bitrate_kbps,
+        audio_channels: s.audio_channels,
         mic_enabled: s.mic_enabled,
         hdr_enabled: s.hdr_enabled,
         decoder: DecoderPref::from_name(&s.decoder),
@@ -886,6 +890,23 @@ fn settings_page(ctx: &Arc<AppCtx>, set_screen: &AsyncSetState<Screen>) -> Eleme
                 s.save();
             })
     };
+    let ac_i = AUDIO_CHANNELS
+        .iter()
+        .position(|&(v, _)| v == s.audio_channels)
+        .unwrap_or(0) as i32;
+    let ac_names: Vec<String> = AUDIO_CHANNELS.iter().map(|&(_, l)| l.to_string()).collect();
+    let channels_combo = {
+        let ctx = ctx.clone();
+        ComboBox::new(ac_names)
+            .header("Audio channels")
+            .selected_index(ac_i)
+            .on_selection_changed(move |i: i32| {
+                let (v, _) = AUDIO_CHANNELS[(i.max(0) as usize).min(AUDIO_CHANNELS.len() - 1)];
+                let mut s = ctx.settings.lock().unwrap();
+                s.audio_channels = v;
+                s.save();
+            })
+    };
 
     let header = grid((
         text_block("Settings")
@@ -934,8 +955,17 @@ fn settings_page(ctx: &Arc<AppCtx>, set_screen: &AsyncSetState<Screen>) -> Eleme
         .spacing(10.0),
     );
 
-    let audio_card =
-        card(vstack((text_block("Audio").font_size(15.0).semibold(), mic_toggle)).spacing(10.0));
+    let audio_card = card(
+        vstack((
+            text_block("Audio").font_size(15.0).semibold(),
+            text_block("Request stereo or surround — the host downmixes if its output has fewer.")
+                .font_size(12.0)
+                .foreground(ThemeRef::SecondaryText),
+            channels_combo,
+            mic_toggle,
+        ))
+        .spacing(10.0),
+    );
 
     page(vec![
         header.into(),

clients/windows/src/audio.rs

@@ -21,9 +21,9 @@ use std::time::Duration;
 use wasapi::{DeviceEnumerator, Direction, SampleType, StreamMode, WaveFormat};
 
 const SAMPLE_RATE: usize = 48_000;
+/// The microphone uplink stays stereo (the host's virtual mic is stereo). The render path is
+/// multichannel — its channel count + block align are runtime, driven by the host-resolved layout.
 const CHANNELS: usize = 2;
-/// 48 kHz stereo f32: 2 channels * 4 bytes = 8 bytes per frame.
-const BLOCK_ALIGN: usize = CHANNELS * 4;
 /// Mic frames are 20 ms (960 samples/channel) — any size ≤ 120 ms is fine host-side.
 const MIC_FRAME: usize = 960;
 
@@ -34,9 +34,10 @@ pub struct AudioPlayer {
 }
 
 impl AudioPlayer {
-    /// Spawn the WASAPI render thread. Failure (no render endpoint on this box) is
-    /// survivable — the caller streams video-only.
-    pub fn spawn() -> Result<AudioPlayer> {
+    /// Spawn the WASAPI render thread for `channels` (2/6/8, canonical wire order
+    /// FL FR FC LFE RL RR SL SR). Failure (no render endpoint on this box) is survivable — the
+    /// caller streams video-only.
+    pub fn spawn(channels: u8) -> Result<AudioPlayer> {
         // 64 × 5 ms = 320 ms of slack between the pump and the WASAPI loop.
         let (pcm_tx, pcm_rx) = std::sync::mpsc::sync_channel::<Vec<f32>>(64);
         let stop = Arc::new(AtomicBool::new(false));
@@ -45,14 +46,14 @@ impl AudioPlayer {
         let thread = std::thread::Builder::new()
             .name("punktfunk-audio".into())
             .spawn(move || {
-                if let Err(e) = render_thread(pcm_rx, stop_t, ready_tx) {
+                if let Err(e) = render_thread(pcm_rx, stop_t, ready_tx, channels) {
                     tracing::warn!(error = format!("{e:#}"), "audio playback thread ended");
                 }
             })
             .context("spawn audio thread")?;
         match ready_rx.recv_timeout(Duration::from_secs(3)) {
             Ok(Ok(())) => {
-                tracing::info!("WASAPI render: 48 kHz stereo f32 (default endpoint)");
+                tracing::info!(channels, "WASAPI render: 48 kHz f32 (default endpoint)");
                 Ok(AudioPlayer {
                     pcm_tx,
                     stop,
@@ -66,8 +67,8 @@ impl AudioPlayer {
         }
     }
 
-    /// Queue one interleaved-stereo f32 chunk. Drops the chunk if the WASAPI side is wedged
-    /// (the renderer conceals the gap; never block the session pump).
+    /// Queue one interleaved f32 chunk (in the session's channel layout). Drops the chunk if the
+    /// WASAPI side is wedged (the renderer conceals the gap; never block the session pump).
     pub fn push(&self, pcm: Vec<f32>) {
         if let Err(TrySendError::Disconnected(_)) = self.pcm_tx.try_send(pcm) {
             // Thread already dead — Drop will reap it; nothing to do per-chunk.
@@ -88,6 +89,7 @@ fn render_thread(
     pcm_rx: Receiver<Vec<f32>>,
     stop: Arc<AtomicBool>,
     ready: SyncSender<Result<()>>,
+    channels: u8,
 ) -> Result<()> {
     if let Err(e) = wasapi::initialize_mta()
         .ok()
@@ -97,12 +99,26 @@ fn render_thread(
         return Ok(());
     }
     let res = (|| -> Result<()> {
+        // F32LE interleaved: channels × 4 bytes/sample. Stereo (channels == 2) is byte-identical
+        // to the old fixed path (mask 0x3, block align 8).
+        let block_align = channels as usize * 4;
         let device = DeviceEnumerator::new()
             .context("DeviceEnumerator")?
             .get_default_device(&Direction::Render)
             .context("default render endpoint")?;
         let mut audio_client = device.get_iaudioclient().context("IAudioClient")?;
-        let desired = WaveFormat::new(32, 32, &SampleType::Float, SAMPLE_RATE, CHANNELS, None);
+        // The explicit dwChannelMask is the wire order (FL FR FC LFE RL RR SL SR); 5.1 = 0x3F,
+        // 7.1 = 0x63F. WASAPI delivers channels in ascending mask-bit order, which equals the wire
+        // order, so the render mapping is the identity — no permute. `autoconvert` (below) lets the
+        // audio engine downmix when the endpoint has fewer speakers.
+        let desired = WaveFormat::new(
+            32,
+            32,
+            &SampleType::Float,
+            SAMPLE_RATE,
+            channels as usize,
+            Some(punktfunk_core::audio::wasapi_channel_mask(channels)),
+        );
         let (default_period, _min_period) =
             audio_client.get_device_period().context("device period")?;
         let mode = StreamMode::EventsShared {
@@ -139,10 +155,10 @@ fn render_thread(
             if avail_frames == 0 {
                 continue;
             }
-            let want_bytes = avail_frames * BLOCK_ALIGN;
+            let want_bytes = avail_frames * block_align;
 
             // Prime to ~3 quanta; cap at ~1 quantum of slack beyond that; re-prime on drain.
-            let target = (3 * want_bytes).clamp(720 * BLOCK_ALIGN, 9600 * BLOCK_ALIGN);
+            let target = (3 * want_bytes).clamp(720 * block_align, 9600 * block_align);
             while ring.len() > target.max(want_bytes) + want_bytes {
                 ring.pop_front();
             }

clients/windows/src/session.rs

@@ -23,6 +23,8 @@ pub struct SessionParams {
     pub compositor: CompositorPref,
     pub gamepad: GamepadPref,
     pub bitrate_kbps: u32,
+    /// Requested audio channel count (2/6/8); the host echoes the resolved value.
+    pub audio_channels: u8,
     /// Stream the default microphone to the host's virtual mic source.
     pub mic_enabled: bool,
     /// Advertise 10-bit + HDR10 so the host may upgrade HDR content to a Main10/PQ stream.
@@ -94,6 +96,42 @@ fn now_ns() -> u64 {
         .unwrap_or(0)
 }
 
+/// Opus decoder for the audio plane: a plain stereo decoder (the validated path) or a multistream
+/// decoder for 5.1/7.1, both behind one `decode_float`. Built from the host-RESOLVED channel count
+/// via the shared layout table.
+enum AudioDec {
+    Stereo(opus::Decoder),
+    Surround(opus::MSDecoder),
+}
+
+impl AudioDec {
+    fn new(channels: u8) -> Result<AudioDec, opus::Error> {
+        if channels == 2 {
+            Ok(AudioDec::Stereo(opus::Decoder::new(
+                48_000,
+                opus::Channels::Stereo,
+            )?))
+        } else {
+            let l = punktfunk_core::audio::layout_for(channels, false);
+            Ok(AudioDec::Surround(opus::MSDecoder::new(
+                48_000, l.streams, l.coupled, l.mapping,
+            )?))
+        }
+    }
+
+    fn decode_float(
+        &mut self,
+        input: &[u8],
+        out: &mut [f32],
+        fec: bool,
+    ) -> Result<usize, opus::Error> {
+        match self {
+            AudioDec::Stereo(d) => d.decode_float(input, out, fec),
+            AudioDec::Surround(d) => d.decode_float(input, out, fec),
+        }
+    }
+}
+
 fn pump(
     params: SessionParams,
     ev_tx: async_channel::Sender<SessionEvent>,
@@ -122,6 +160,7 @@ fn pump(
             }
             0
         },
+        params.audio_channels,
         None, // launch: the Windows client has no library picker yet
         params.pin,
         Some(params.identity),
@@ -161,11 +200,14 @@ fn pump(
     let mut hardware = decoder.is_hardware();
     let mut hdr = false;
     // Audio is best-effort: a session without it still streams. Gamepads are the
-    // app-lifetime service's job (the UI attaches it on Connected).
-    let player = audio::AudioPlayer::spawn()
+    // app-lifetime service's job (the UI attaches it on Connected). Build the decoder + playback
+    // from the host-RESOLVED channel count (never the request), so an older/clamping host that
+    // resolves stereo is decoded as stereo.
+    let channels = connector.audio_channels;
+    let player = audio::AudioPlayer::spawn(channels)
         .map_err(|e| tracing::warn!(error = %e, "audio disabled"))
         .ok();
-    let mut opus_dec = opus::Decoder::new(48_000, opus::Channels::Stereo)
+    let mut opus_dec = AudioDec::new(channels)
         .map_err(|e| tracing::warn!(error = %e, "opus decoder failed — audio disabled"))
         .ok();
     let _mic = params
@@ -184,8 +226,8 @@ fn pump(
     let mut bytes_n = 0u64;
     let mut decode_us_sum = 0u64;
     let mut lat_us: Vec<u64> = Vec::with_capacity(256);
-    let mut pcm = vec![0f32; 5760 * 2]; // decode scratch: max Opus frame (120 ms stereo)
-                                        // Loss recovery: watch the host→client unrecoverable-drop count and ask for an IDR when it climbs.
+    let mut pcm = vec![0f32; 5760 * channels as usize]; // scratch: max Opus frame (120 ms) × channels
+                                                        // Loss recovery: watch the host→client unrecoverable-drop count and ask for an IDR when it climbs.
     let mut last_dropped = connector.frames_dropped();
     let mut last_kf_req: Option<Instant> = None;
 
@@ -253,7 +295,8 @@ fn pump(
         while let Ok(pkt) = connector.next_audio(Duration::ZERO) {
             if let (Some(player), Some(dec)) = (&player, opus_dec.as_mut()) {
                 match dec.decode_float(&pkt.data, &mut pcm, false) {
-                    Ok(samples) => player.push(pcm[..samples * 2].to_vec()),
+                    // `samples` is per-channel; the interleaved frame is `samples * channels`.
+                    Ok(samples) => player.push(pcm[..samples * channels as usize].to_vec()),
                     Err(e) => tracing::debug!(error = %e, "opus decode"),
                 }
             }

clients/windows/src/trust.rs

@@ -130,6 +130,9 @@ pub struct Settings {
     pub inhibit_shortcuts: bool,
     /// Stream the default microphone to the host's virtual mic source.
     pub mic_enabled: bool,
+    /// Requested audio channel count: 2 (stereo), 6 (5.1) or 8 (7.1). The host clamps to what it
+    /// can capture; the resolved count drives the decoder + WASAPI render layout.
+    pub audio_channels: u8,
     /// Advertise 10-bit + HDR10 so the host upgrades HDR content to a Main10/PQ stream (the client
     /// presents it on a 10-bit ST.2084 swapchain). No effect on SDR content.
     pub hdr_enabled: bool,
@@ -148,6 +151,7 @@ impl Default for Settings {
             compositor: "auto".into(),
             inhibit_shortcuts: true,
             mic_enabled: false,
+            audio_channels: 2,
             hdr_enabled: true,
             decoder: "auto".into(),
         }

crates/punktfunk-core/Cargo.toml

@@ -19,7 +19,7 @@ crate-type = ["lib", "cdylib", "staticlib"]
 default = []
 # Control-plane QUIC (pairing, config, reverse audio). tokio is permitted ONLY here,
 # never on the per-frame hot path. Off by default so the core stays runtime-free.
-quic = ["dep:quinn", "dep:tokio", "dep:rustls", "dep:rcgen", "dep:rustls-pki-types", "dep:sha2", "dep:hmac", "dep:spake2"]
+quic = ["dep:quinn", "dep:tokio", "dep:rustls", "dep:rcgen", "dep:rustls-pki-types", "dep:sha2", "dep:hmac", "dep:spake2", "dep:opus"]
 
 [dependencies]
 reed-solomon-simd = "3.1"      # GF(2^16) Leopard-RS, SIMD, O(n log n) — the wall-breaker (P2)
@@ -51,6 +51,12 @@ sha2 = { version = "0.10", optional = true }
 hmac = { version = "0.12", optional = true }
 spake2 = { version = "0.4", optional = true }
 tokio = { version = "1", optional = true, features = ["rt-multi-thread", "net", "sync", "macros"] }
+# In-core Opus (multistream) DECODE for the C-ABI `punktfunk_connection_next_audio_pcm` path —
+# used by embedders without a multistream-capable Opus decoder (Apple's AudioToolbox is
+# stereo-only). The Rust clients link `opus` themselves and decode the raw `next_audio` frames,
+# so this only matters when the connection API (quic) is built. Same libopus the host vendors;
+# cargo unifies the build. Multistream API: `opus::MSDecoder` (lib.rs:1187).
+opus = { version = "0.3", optional = true }
 
 # `libc` for batched UDP syscalls: `sendmmsg`/`recvmmsg` on Linux (the 1 Gbps+ lever) and the
 # `recv(MSG_DONTWAIT)` drain on the other unix (Apple/BSD) targets, which have no `recvmmsg`

crates/punktfunk-core/src/abi.rs

@@ -467,6 +467,23 @@ pub struct PunktfunkConnection {
     last: std::sync::Mutex<Option<crate::session::Frame>>,
     /// Same, for `punktfunk_connection_next_audio` (independent of the video slot).
     last_audio: std::sync::Mutex<Option<crate::client::AudioPacket>>,
+    /// Decode-in-core state for `punktfunk_connection_next_audio_pcm` (Apple / any embedder
+    /// without a multistream Opus decoder). The decoder is built lazily from the negotiated
+    /// `inner.audio_channels`; `pcm` is a fixed-capacity reusable buffer the returned pointer
+    /// borrows until the next PCM call (same contract as `last_audio`).
+    audio_pcm: std::sync::Mutex<AudioPcmState>,
+}
+
+/// Lazily-initialized in-core Opus decode state. A coupled-1-stream multistream decoder is
+/// equivalent to a plain stereo decoder, so one [`opus::MSDecoder`] handles 2/6/8 channels.
+#[cfg(feature = "quic")]
+#[derive(Default)]
+struct AudioPcmState {
+    decoder: Option<opus::MSDecoder>,
+    /// Interleaved f32 PCM, wire channel order. Pre-sized to the largest legal Opus frame
+    /// (120 ms @ 48 kHz = 5760 samples/ch) × 8 channels so decode never reallocates (which would
+    /// dangle the pointer handed to the embedder).
+    pcm: Vec<f32>,
 }
 
 /// `PunktfunkHidOutput::kind` — lightbar RGB (`r`/`g`/`b` valid).
@@ -708,12 +725,18 @@ pub const PUNKTFUNK_VIDEO_CAP_10BIT: u8 = 0x01;
 /// Video-capability bit for [`punktfunk_connect_ex5`] (`video_caps`): the client can present
 /// BT.2020 PQ HDR10 (implies 10-bit). (Mirrors `quic::VIDEO_CAP_HDR`.)
 pub const PUNKTFUNK_VIDEO_CAP_HDR: u8 = 0x02;
+/// Video-capability bit for [`punktfunk_connect_ex5`] (`video_caps`): the client can decode a
+/// full-chroma 4:4:4 HEVC stream (Range Extensions). The host emits 4:4:4 only when this is set,
+/// the host opted in, the codec is HEVC, and the GPU supports it — else the stream stays 4:2:0 and
+/// [`punktfunk_connection_chroma_format`] reports the real value. (Mirrors `quic::VIDEO_CAP_444`.)
+pub const PUNKTFUNK_VIDEO_CAP_444: u8 = 0x04;
 
 // Keep the ABI cap bits in lockstep with the wire constants (compile-time guard against drift).
 #[cfg(feature = "quic")]
 const _: () = {
     assert!(PUNKTFUNK_VIDEO_CAP_10BIT == crate::quic::VIDEO_CAP_10BIT);
     assert!(PUNKTFUNK_VIDEO_CAP_HDR == crate::quic::VIDEO_CAP_HDR);
+    assert!(PUNKTFUNK_VIDEO_CAP_444 == crate::quic::VIDEO_CAP_444);
 };
 
 // Keep the ABI gamepad constants in lockstep with the wire enum (compile-time guard against drift).
@@ -980,6 +1003,58 @@ pub unsafe extern "C" fn punktfunk_connect_ex5(
     client_cert_pem: *const std::os::raw::c_char,
     client_key_pem: *const std::os::raw::c_char,
     timeout_ms: u32,
+) -> *mut PunktfunkConnection {
+    // Delegate to the surround-aware variant requesting stereo (the pre-surround behaviour).
+    unsafe {
+        punktfunk_connect_ex6(
+            host,
+            port,
+            width,
+            height,
+            refresh_hz,
+            compositor,
+            gamepad,
+            bitrate_kbps,
+            video_caps,
+            2, // audio_channels = stereo
+            launch_id,
+            pin_sha256,
+            observed_sha256_out,
+            client_cert_pem,
+            client_key_pem,
+            timeout_ms,
+        )
+    }
+}
+
+/// Like [`punktfunk_connect_ex5`], but additionally requests the audio channel count:
+/// `2` (stereo, the default behaviour of every earlier variant), `6` (5.1) or `8` (7.1). The host
+/// clamps the request to what it can actually capture and echoes the resolved count via
+/// [`punktfunk_connection_audio_channels`]; the `0xC9` audio frames are Opus-(multi)stream encoded
+/// for that layout. A client that wants surround calls this; everything else inherits stereo.
+///
+/// # Safety
+/// Same as [`punktfunk_connect`].
+#[cfg(feature = "quic")]
+#[no_mangle]
+#[allow(clippy::too_many_arguments)]
+pub unsafe extern "C" fn punktfunk_connect_ex6(
+    host: *const std::os::raw::c_char,
+    port: u16,
+    width: u32,
+    height: u32,
+    refresh_hz: u32,
+    compositor: u32,
+    gamepad: u32,
+    bitrate_kbps: u32,
+    video_caps: u8,
+    audio_channels: u8,
+    launch_id: *const std::os::raw::c_char,
+    pin_sha256: *const u8,
+    observed_sha256_out: *mut u8,
+    client_cert_pem: *const std::os::raw::c_char,
+    client_key_pem: *const std::os::raw::c_char,
+    timeout_ms: u32,
 ) -> *mut PunktfunkConnection {
     let r = std::panic::catch_unwind(AssertUnwindSafe(|| {
         if host.is_null() {
@@ -1029,6 +1104,7 @@ pub unsafe extern "C" fn punktfunk_connect_ex5(
             gamepad,
             bitrate_kbps,
             video_caps,
+            crate::audio::normalize_channels(audio_channels),
             launch,
             pin,
             identity,
@@ -1045,6 +1121,7 @@ pub unsafe extern "C" fn punktfunk_connect_ex5(
                     inner: c,
                     last: std::sync::Mutex::new(None),
                     last_audio: std::sync::Mutex::new(None),
+                    audio_pcm: std::sync::Mutex::new(AudioPcmState::default()),
                 }))
             }
             Err(_) => std::ptr::null_mut(),
@@ -1250,6 +1327,121 @@ pub unsafe extern "C" fn punktfunk_connection_next_audio(
     })
 }
 
+/// Read the audio channel count the host resolved for this session (from its Welcome): `2`
+/// (stereo), `6` (5.1) or `8` (7.1). `*out` is filled when non-NULL. The `0xC9` Opus frames are
+/// (multistream-)encoded for this layout; an embedder decoding raw frames itself must build its
+/// decoder from THIS value (see [`crate::audio::layout_for`]) — or use
+/// [`punktfunk_connection_next_audio_pcm`], which decodes in-core. Available immediately after a
+/// successful connect (it doesn't change without a reconfigure).
+///
+/// # Safety
+/// `c` is a valid connection handle; `out` is NULL or writable for one `u8`.
+#[cfg(feature = "quic")]
+#[no_mangle]
+pub unsafe extern "C" fn punktfunk_connection_audio_channels(
+    c: *mut PunktfunkConnection,
+    out: *mut u8,
+) -> PunktfunkStatus {
+    guard(|| {
+        let c = match unsafe { c.as_ref() } {
+            Some(c) => c,
+            None => return PunktfunkStatus::NullPointer,
+        };
+        if !out.is_null() {
+            // SAFETY: `out` is non-null and the caller guarantees it is writable for one `u8`.
+            unsafe { *out = c.inner.audio_channels };
+        }
+        PunktfunkStatus::Ok
+    })
+}
+
+/// One decoded audio frame from [`punktfunk_connection_next_audio_pcm`]: interleaved 32-bit
+/// float PCM at 48 kHz, in the canonical wire channel order `FL FR FC LFE RL RR SL SR` (the
+/// first `channels` of it). `samples` points at `frame_count * channels` floats and borrows
+/// connection memory **until the next PCM call** on this handle.
+#[cfg(feature = "quic")]
+#[repr(C)]
+pub struct PunktfunkAudioPcm {
+    /// Interleaved f32 samples (wire channel order), `frame_count * channels` long.
+    pub samples: *const f32,
+    /// Samples per channel in this frame.
+    pub frame_count: u32,
+    /// Channel count (2/6/8) — the negotiated [`punktfunk_connection_audio_channels`].
+    pub channels: u8,
+    /// Source packet sequence number.
+    pub seq: u32,
+    /// Capture presentation timestamp (ns).
+    pub pts_ns: u64,
+}
+
+/// Pull the next audio frame and **decode it in-core** to interleaved f32 PCM — for embedders
+/// without a multistream-capable Opus decoder (e.g. Apple, whose AudioToolbox Opus path is
+/// stereo-only). The decoder is built once from the negotiated channel count and handles 2/6/8
+/// channels (a 1-coupled-stream multistream decoder is exactly a stereo decoder). Same
+/// timeout/closed semantics as [`punktfunk_connection_next_audio`]; `out->samples` borrows
+/// connection memory until the next PCM call on this handle. Use EITHER this or
+/// [`punktfunk_connection_next_audio`] on a given connection, from one dedicated audio thread —
+/// not both (they share the underlying queue).
+///
+/// # Safety
+/// `c` is a valid connection handle; `out` is writable. At most one thread pulls audio.
+#[cfg(feature = "quic")]
+#[no_mangle]
+pub unsafe extern "C" fn punktfunk_connection_next_audio_pcm(
+    c: *mut PunktfunkConnection,
+    out: *mut PunktfunkAudioPcm,
+    timeout_ms: u32,
+) -> PunktfunkStatus {
+    guard(|| {
+        let c = match unsafe { c.as_ref() } {
+            Some(c) => c,
+            None => return PunktfunkStatus::NullPointer,
+        };
+        if out.is_null() {
+            return PunktfunkStatus::NullPointer;
+        }
+        let channels = crate::audio::normalize_channels(c.inner.audio_channels);
+        let pkt = match c
+            .inner
+            .next_audio(std::time::Duration::from_millis(timeout_ms as u64))
+        {
+            Ok(pkt) => pkt,
+            Err(e) => return e.status(),
+        };
+        let mut state = c.audio_pcm.lock().unwrap();
+        if state.decoder.is_none() {
+            let layout = crate::audio::layout_for(channels, false);
+            match opus::MSDecoder::new(48_000, layout.streams, layout.coupled, layout.mapping) {
+                Ok(d) => {
+                    // Largest legal Opus frame is 120 ms = 5760 samples/ch.
+                    state.pcm = vec![0f32; 5760 * channels as usize];
+                    state.decoder = Some(d);
+                }
+                Err(_) => return PunktfunkStatus::Unsupported,
+            }
+        }
+        let AudioPcmState { decoder, pcm } = &mut *state;
+        let dec = decoder.as_mut().unwrap();
+        // `decode_float` divides the output buffer length by the channel count to get the
+        // per-channel capacity; an empty payload requests packet-loss concealment.
+        match dec.decode_float(&pkt.data, pcm, false) {
+            Ok(frame_count) => {
+                unsafe {
+                    *out = PunktfunkAudioPcm {
+                        samples: pcm.as_ptr(),
+                        frame_count: frame_count as u32,
+                        channels,
+                        seq: pkt.seq,
+                        pts_ns: pkt.pts_ns,
+                    };
+                }
+                PunktfunkStatus::Ok
+            }
+            Err(_) => PunktfunkStatus::BadPacket,
+        }
+    })
+}
+
 /// Pull the next rumble (force-feedback) update, waiting up to `timeout_ms`. Amplitudes
 /// are 0..0xFFFF (`low` = low-frequency motor, `high` = high-frequency), `(0, 0)` = stop.
 /// Same timeout/closed semantics as [`punktfunk_connection_next_audio`].
@@ -1414,6 +1606,33 @@ pub unsafe extern "C" fn punktfunk_connection_color_info(
     })
 }
 
+/// Read the session's resolved chroma subsampling (from the host's Welcome) as the HEVC
+/// `chroma_format_idc`: `1` = 4:2:0 (the default every pre-4:4:4 host produced), `3` = full-chroma
+/// 4:4:4. `*out` is filled when non-NULL. The in-band SPS is authoritative; this lets the embedder
+/// pre-size its decoder / pick a 4:4:4 pixel format up front. Available immediately after a
+/// successful connect (it doesn't change without a reconfigure).
+///
+/// # Safety
+/// `c` is a valid connection handle; `out` is NULL or writable for one `u8`.
+#[cfg(feature = "quic")]
+#[no_mangle]
+pub unsafe extern "C" fn punktfunk_connection_chroma_format(
+    c: *mut PunktfunkConnection,
+    out: *mut u8,
+) -> PunktfunkStatus {
+    guard(|| {
+        let c = match unsafe { c.as_ref() } {
+            Some(c) => c,
+            None => return PunktfunkStatus::NullPointer,
+        };
+        if !out.is_null() {
+            // SAFETY: `out` is non-null and the caller guarantees it is writable for one `u8`.
+            unsafe { *out = c.inner.chroma_format };
+        }
+        PunktfunkStatus::Ok
+    })
+}
+
 /// Send one input event to the host as a QUIC datagram (non-blocking enqueue).
 ///
 /// # Safety

crates/punktfunk-core/src/audio.rs

@@ -0,0 +1,298 @@
+//! Shared audio layout: the single source of truth for Opus (multi)stream surround across the
+//! host, the GameStream compatibility path, and every client decoder.
+//!
+//! **Canonical wire channel order** is `FL FR FC LFE RL RR SL SR` (the GameStream/Moonlight
+//! order, and the PipeWire/PulseAudio default map for 6/8 channels). Every host capturer
+//! delivers PCM in this order and every client decodes into it, so the Opus multistream
+//! `mapping` is the **identity** (`[0, 1, …, channels-1]`) on both ends — punktfunk owns the
+//! encoder and every decoder, so the GFE-style pre-rotation Moonlight needs over SDP
+//! (`gamestream::audio::surround_params`) is a GameStream-only concern and never touches the
+//! native `punktfunk/1` path.
+//!
+//! Channel counts the protocol negotiates: `2` (stereo), `6` (5.1) and `8` (7.1). Anything
+//! else clamps to stereo ([`normalize_channels`]).
+
+/// Canonical wire channel positions; the index is the channel's slot in the interleaved PCM
+/// frame. A count of N uses positions `0..N` (always a prefix of this 8-channel order).
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+#[repr(u8)]
+pub enum WirePos {
+    FrontLeft = 0,
+    FrontRight = 1,
+    FrontCenter = 2,
+    Lfe = 3,
+    RearLeft = 4,
+    RearRight = 5,
+    SideLeft = 6,
+    SideRight = 7,
+}
+
+/// The full 8-channel wire order; the N-channel order is its first N entries.
+pub const WIRE_ORDER_8: [WirePos; 8] = {
+    use WirePos::*;
+    [
+        FrontLeft,
+        FrontRight,
+        FrontCenter,
+        Lfe,
+        RearLeft,
+        RearRight,
+        SideLeft,
+        SideRight,
+    ]
+};
+
+/// One Opus (multi)stream layout. `mapping` is the libopus multistream mapping we encode AND
+/// decode with — identity, since punktfunk owns both ends. `streams`/`coupled` give the
+/// normal-quality coupling (FL,FR)+(FC,LFE) [+(RL,RR) on 7.1] with the remaining channels as
+/// mono streams; high quality is one mono stream per channel. Bitrates match Sunshine's
+/// per-config values (stereo keeps punktfunk's live-validated 128 kbps).
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct OpusLayout {
+    /// Interleaved channel count (2, 6 or 8).
+    pub channels: u8,
+    /// Number of Opus streams in the multistream packet.
+    pub streams: u8,
+    /// How many of those streams are coupled (stereo) pairs.
+    pub coupled: u8,
+    /// libopus multistream channel mapping — identity `[0, 1, …, channels-1]`.
+    pub mapping: &'static [u8],
+    /// Target Opus bitrate in bits/sec (hard CBR; constant packet size, which GameStream's
+    /// audio FEC relies on).
+    pub bitrate: i32,
+}
+
+/// Stereo: a plain coupled pair. The 128 kbps live-validated config.
+pub const LAYOUT_STEREO: OpusLayout = OpusLayout {
+    channels: 2,
+    streams: 1,
+    coupled: 1,
+    mapping: &[0, 1],
+    bitrate: 128_000,
+};
+/// 5.1 normal quality: (FL,FR)+(FC,LFE) coupled, RL+RR mono.
+pub const LAYOUT_51: OpusLayout = OpusLayout {
+    channels: 6,
+    streams: 4,
+    coupled: 2,
+    mapping: &[0, 1, 2, 3, 4, 5],
+    bitrate: 256_000,
+};
+/// 5.1 high quality: one mono stream per channel.
+pub const LAYOUT_51_HQ: OpusLayout = OpusLayout {
+    channels: 6,
+    streams: 6,
+    coupled: 0,
+    mapping: &[0, 1, 2, 3, 4, 5],
+    bitrate: 1_536_000,
+};
+/// 7.1 normal quality: (FL,FR)+(FC,LFE)+(RL,RR) coupled, SL+SR mono.
+pub const LAYOUT_71: OpusLayout = OpusLayout {
+    channels: 8,
+    streams: 5,
+    coupled: 3,
+    mapping: &[0, 1, 2, 3, 4, 5, 6, 7],
+    bitrate: 450_000,
+};
+/// 7.1 high quality: one mono stream per channel.
+pub const LAYOUT_71_HQ: OpusLayout = OpusLayout {
+    channels: 8,
+    streams: 8,
+    coupled: 0,
+    mapping: &[0, 1, 2, 3, 4, 5, 6, 7],
+    bitrate: 2_048_000,
+};
+
+/// Pick the layout for a negotiated channel count. Unknown counts fall back to stereo (clients
+/// only ever request 2/6/8). `high_quality` selects the uncoupled high-bitrate config.
+pub fn layout_for(channels: u8, high_quality: bool) -> &'static OpusLayout {
+    match (channels, high_quality) {
+        (6, false) => &LAYOUT_51,
+        (6, true) => &LAYOUT_51_HQ,
+        (8, false) => &LAYOUT_71,
+        (8, true) => &LAYOUT_71_HQ,
+        _ => &LAYOUT_STEREO,
+    }
+}
+
+/// Clamp an arbitrary (wire / requested) channel count to one the protocol negotiates. `0`,
+/// absent, or any unsupported value becomes stereo.
+pub fn normalize_channels(requested: u8) -> u8 {
+    match requested {
+        6 => 6,
+        8 => 8,
+        _ => 2,
+    }
+}
+
+// ---- per-platform channel-layout helpers (pure data; no platform deps) --------------------
+
+/// Windows `WAVEFORMATEXTENSIBLE.dwChannelMask` for the wire layout.
+///
+/// NB 7.1 == `0x63F` (FL FR FC LFE **BL BR SL SR**), NOT `0xFF` — `0xFF` selects the
+/// front-of-center pair FLC/FRC, the wrong speakers. WASAPI delivers channels in ascending
+/// mask-bit order, which equals the wire order, so the decoded PCM needs no permutation.
+pub const fn wasapi_channel_mask(channels: u8) -> u32 {
+    const FL: u32 = 0x1;
+    const FR: u32 = 0x2;
+    const FC: u32 = 0x4;
+    const LFE: u32 = 0x8;
+    const BL: u32 = 0x10; // back left  (wire RL)
+    const BR: u32 = 0x20; // back right (wire RR)
+    const SL: u32 = 0x200; // side left
+    const SR: u32 = 0x400; // side right
+    match channels {
+        6 => FL | FR | FC | LFE | BL | BR,           // 0x3F
+        8 => FL | FR | FC | LFE | BL | BR | SL | SR, // 0x63F
+        _ => FL | FR,                                // 0x3 (stereo)
+    }
+}
+
+/// PipeWire / SPA `enum spa_audio_channel` positions in wire order — identical to the host
+/// capture side (`punktfunk-host` `audio::linux::spa_positions`): FL=3 FR=4 FC=5 LFE=6 SL=7
+/// SR=8 RL=12 RR=13. Identity routing: the client sets these on its playback node so PipeWire
+/// maps each wire slot to the matching speaker (and downmixes when the sink has fewer).
+pub fn spa_positions(channels: u8) -> &'static [u32] {
+    const STEREO: [u32; 2] = [3, 4]; // FL FR
+    const C51: [u32; 6] = [3, 4, 5, 6, 12, 13]; // FL FR FC LFE RL RR
+    const C71: [u32; 8] = [3, 4, 5, 6, 12, 13, 7, 8]; // FL FR FC LFE RL RR SL SR
+    match channels {
+        6 => &C51,
+        8 => &C71,
+        _ => &STEREO,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn layout_table_is_consistent() {
+        for l in [
+            &LAYOUT_STEREO,
+            &LAYOUT_51,
+            &LAYOUT_51_HQ,
+            &LAYOUT_71,
+            &LAYOUT_71_HQ,
+        ] {
+            // Mapping is identity and exactly `channels` entries long.
+            assert_eq!(l.mapping.len(), l.channels as usize);
+            for (i, &m) in l.mapping.iter().enumerate() {
+                assert_eq!(m as usize, i, "mapping must be identity for {l:?}");
+            }
+            // libopus invariant: total channels == coupled*2 + (streams - coupled).
+            assert_eq!(
+                l.coupled * 2 + (l.streams - l.coupled),
+                l.channels,
+                "stream/coupled accounting for {l:?}"
+            );
+            assert!(l.coupled <= l.streams);
+            assert!(l.bitrate > 0);
+        }
+    }
+
+    #[test]
+    fn layout_for_picks_expected() {
+        assert_eq!(layout_for(2, false), &LAYOUT_STEREO);
+        assert_eq!(layout_for(6, false), &LAYOUT_51);
+        assert_eq!(layout_for(6, true), &LAYOUT_51_HQ);
+        assert_eq!(layout_for(8, false), &LAYOUT_71);
+        assert_eq!(layout_for(8, true), &LAYOUT_71_HQ);
+        // Unknown / 0 → stereo.
+        assert_eq!(layout_for(0, false), &LAYOUT_STEREO);
+        assert_eq!(layout_for(3, false), &LAYOUT_STEREO);
+        assert_eq!(layout_for(7, true), &LAYOUT_STEREO);
+    }
+
+    #[test]
+    fn normalize_clamps_to_negotiable() {
+        assert_eq!(normalize_channels(2), 2);
+        assert_eq!(normalize_channels(6), 6);
+        assert_eq!(normalize_channels(8), 8);
+        for bad in [0u8, 1, 3, 4, 5, 7, 9, 255] {
+            assert_eq!(normalize_channels(bad), 2, "{bad} must clamp to stereo");
+        }
+    }
+
+    #[test]
+    fn wasapi_masks_are_correct() {
+        assert_eq!(wasapi_channel_mask(2), 0x3);
+        assert_eq!(wasapi_channel_mask(6), 0x3F);
+        assert_eq!(wasapi_channel_mask(8), 0x63F); // NOT 0xFF
+                                                   // Bit count must equal the channel count.
+        assert_eq!(wasapi_channel_mask(2).count_ones(), 2);
+        assert_eq!(wasapi_channel_mask(6).count_ones(), 6);
+        assert_eq!(wasapi_channel_mask(8).count_ones(), 8);
+    }
+
+    #[test]
+    fn spa_positions_match_wire_order() {
+        assert_eq!(spa_positions(2), &[3, 4]);
+        assert_eq!(spa_positions(6), &[3, 4, 5, 6, 12, 13]);
+        assert_eq!(spa_positions(8), &[3, 4, 5, 6, 12, 13, 7, 8]);
+        assert_eq!(spa_positions(2).len(), 2);
+        assert_eq!(spa_positions(6).len(), 6);
+        assert_eq!(spa_positions(8).len(), 8);
+    }
+
+    /// Real-libopus proof that the shared layout round-trips with channel identity: a tone fed
+    /// into wire channel N (host `opus::MSEncoder`) comes back out on channel N (client
+    /// `opus::MSDecoder`), for stereo / 5.1 / 7.1. This is the single guarantee the whole
+    /// feature rests on — encoder layout == decoder layout == identity mapping — so if a layout
+    /// constant is ever wrong, this fails. Gated on `quic` (where `opus` is a dependency).
+    #[cfg(feature = "quic")]
+    #[test]
+    fn multistream_layout_roundtrips_with_channel_identity() {
+        const SR: u32 = 48_000;
+        const SAMPLES: usize = 240; // 5 ms @ 48 kHz
+        for &channels in &[2u8, 6, 8] {
+            let l = layout_for(channels, false);
+            let ch = l.channels as usize;
+            let mut enc = opus::MSEncoder::new(
+                SR,
+                l.streams,
+                l.coupled,
+                l.mapping,
+                opus::Application::LowDelay,
+            )
+            .expect("MSEncoder");
+            enc.set_bitrate(opus::Bitrate::Bits(l.bitrate)).unwrap();
+            enc.set_vbr(false).unwrap();
+            let mut dec =
+                opus::MSDecoder::new(SR, l.streams, l.coupled, l.mapping).expect("MSDecoder");
+
+            for tone_ch in 0..ch {
+                let mut out = vec![0u8; 4000];
+                let mut energy = vec![0f64; ch];
+                // A few frames to clear the codec startup transient before measuring.
+                for f in 0..8 {
+                    let mut frame = vec![0f32; SAMPLES * ch];
+                    for t in 0..SAMPLES {
+                        let phase = (f * SAMPLES + t) as f32 * 440.0 * 2.0 * std::f32::consts::PI
+                            / SR as f32;
+                        frame[t * ch + tone_ch] = 0.5 * phase.sin();
+                    }
+                    let n = enc.encode_float(&frame, &mut out).unwrap();
+                    let mut decoded = vec![0f32; SAMPLES * ch];
+                    let got = dec.decode_float(&out[..n], &mut decoded, false).unwrap();
+                    assert_eq!(got, SAMPLES, "{channels}ch frame size");
+                    if f >= 4 {
+                        for t in 0..SAMPLES {
+                            for (c, e) in energy.iter_mut().enumerate() {
+                                *e += (decoded[t * ch + c] as f64).powi(2);
+                            }
+                        }
+                    }
+                }
+                let loudest = (0..ch)
+                    .max_by(|&a, &b| energy[a].total_cmp(&energy[b]))
+                    .unwrap();
+                assert_eq!(
+                    loudest, tone_ch,
+                    "{channels}ch: tone in channel {tone_ch} must come out on {tone_ch} (energies {energy:?})"
+                );
+            }
+        }
+    }
+}

crates/punktfunk-core/src/client.rs

@@ -40,8 +40,9 @@ enum CtrlRequest {
 /// mode, the host-resolved compositor backend, the host-resolved gamepad backend, the host's
 /// certificate fingerprint, the resolved encoder bitrate (kbps), and the host↔client clock offset
 /// (ns, host minus client; 0 = no skew correction / an old host that didn't answer the handshake).
-/// The trailing `u8` is the resolved encode bit depth (8/10) and [`ColorInfo`] the resolved colour
-/// signalling, both from the [`Welcome`].
+/// The trailing `u8`s are the resolved encode bit depth (8/10), the chroma `chroma_format_idc`
+/// (1 = 4:2:0, 3 = 4:4:4), and the resolved audio channel count (2/6/8), with [`ColorInfo`] the
+/// resolved colour signalling — all from the [`Welcome`].
 type Negotiated = (
     Mode,
     CompositorPref,
@@ -51,6 +52,8 @@ type Negotiated = (
     i64,
     u8,
     ColorInfo,
+    u8,
+    u8,
 );
 
 /// Accumulated state of an in-flight / finished speed test. The data-plane pump mirrors the
@@ -202,6 +205,17 @@ pub struct NativeClient {
     /// decoder/presenter from this. [`ColorInfo::SDR_BT709`] for an older host. The static HDR
     /// mastering metadata (when [`ColorInfo::is_hdr`]) arrives via [`NativeClient::next_hdr_meta`].
     pub color: ColorInfo,
+    /// The chroma subsampling the host resolved for this session ([`Welcome::chroma_format`]), as the
+    /// HEVC `chroma_format_idc`: [`quic::CHROMA_IDC_420`] (4:2:0, the default / older host) or
+    /// [`quic::CHROMA_IDC_444`] (full-chroma 4:4:4). The in-band SPS is authoritative; this lets the
+    /// client pre-size its decoder. `CHROMA_IDC_420` for an older host that didn't report it.
+    pub chroma_format: u8,
+    /// The audio channel count the host resolved for this session ([`Welcome::audio_channels`]):
+    /// `2` (stereo), `6` (5.1) or `8` (7.1). The client MUST build its Opus (multistream) decoder
+    /// from this value (via [`crate::audio::layout_for`]) — never from its own request — so an older
+    /// host that omits it (→ `2`) yields working stereo. The `0xC9` audio frames are encoded with the
+    /// matching layout.
+    pub audio_channels: u8,
 }
 
 /// Pin the calling thread to the user-interactive QoS class on Apple targets.
@@ -246,6 +260,9 @@ impl NativeClient {
         // VIDEO_CAP_HDR) — the host upgrades to a 10-bit / HDR encode only when the matching bit is
         // set. 0 = the 8-bit BT.709 stream every client understands.
         video_caps: u8,
+        // Requested audio channel count (2 = stereo / 6 = 5.1 / 8 = 7.1); the host clamps to what it
+        // can capture and echoes the result in [`NativeClient::audio_channels`].
+        audio_channels: u8,
         launch: Option<String>,
         pin: Option<[u8; 32]>,
         identity: Option<(String, String)>,
@@ -298,6 +315,7 @@ impl NativeClient {
                     gamepad,
                     bitrate_kbps,
                     video_caps,
+                    audio_channels,
                     launch,
                     pin,
                     identity,
@@ -329,6 +347,8 @@ impl NativeClient {
             clock_offset_ns,
             bit_depth,
             color,
+            chroma_format,
+            audio_channels,
         ) = match ready_rx.recv_timeout(timeout) {
             Ok(Ok(t)) => t,
             Ok(Err(e)) => return Err(e),
@@ -360,6 +380,8 @@ impl NativeClient {
             clock_offset_ns,
             bit_depth,
             color,
+            chroma_format,
+            audio_channels,
         })
     }
 
@@ -666,6 +688,7 @@ struct WorkerArgs {
     gamepad: GamepadPref,
     bitrate_kbps: u32,
     video_caps: u8,
+    audio_channels: u8,
     launch: Option<String>,
     pin: Option<[u8; 32]>,
     identity: Option<(String, String)>,
@@ -697,6 +720,7 @@ async fn worker_main(args: WorkerArgs) {
         gamepad,
         bitrate_kbps,
         video_caps,
+        audio_channels,
         launch,
         pin,
         identity,
@@ -763,6 +787,8 @@ async fn worker_main(args: WorkerArgs) {
                 // VIDEO_CAP_10BIT | VIDEO_CAP_HDR). The host only upgrades to a 10-bit / HDR encode
                 // when the matching bit is set, so `0` stays an 8-bit BT.709 stream.
                 video_caps,
+                // Requested surround channel count; the host echoes the resolved value in Welcome.
+                audio_channels,
             }
             .encode(),
         )
@@ -834,6 +860,8 @@ async fn worker_main(args: WorkerArgs) {
             clock_offset_ns,
             welcome.bit_depth,
             welcome.color,
+            welcome.chroma_format,
+            welcome.audio_channels,
         ))
     };
 
@@ -850,6 +878,8 @@ async fn worker_main(args: WorkerArgs) {
         clock_offset_ns,
         bit_depth,
         color,
+        chroma_format,
+        audio_channels,
     ) = match setup.await {
         Ok(t) => t,
         Err(e) => {
@@ -866,6 +896,8 @@ async fn worker_main(args: WorkerArgs) {
         clock_offset_ns,
         bit_depth,
         color,
+        chroma_format,
+        audio_channels,
     )));
 
     // Input task: embedder events → QUIC datagrams.

crates/punktfunk-core/src/lib.rs

@@ -25,6 +25,7 @@
 #![forbid(unsafe_op_in_unsafe_fn)]
 
 pub mod abi;
+pub mod audio;
 #[cfg(feature = "quic")]
 pub mod client;
 pub mod config;

crates/punktfunk-core/src/quic.rs

@@ -78,12 +78,33 @@ pub struct Hello {
     /// zero-length name/launch placeholder precedes it when those are absent so the offset stays
     /// deterministic. Omitted by older clients (decodes to `0`).
     pub video_caps: u8,
+    /// Requested audio channel count: `2` (stereo, default), `6` (5.1) or `8` (7.1). The host
+    /// resolves it against what it can capture and echoes the final count in
+    /// [`Welcome::audio_channels`], which is what both ends build their Opus (multistream)
+    /// codec from. Appended after `video_caps` as a single trailing byte; when it differs from
+    /// the stereo default the name/launch/video_caps placeholders are forced (0) so it lands at a
+    /// deterministic offset. Omitted by older clients / when `2` (decodes to `2`, i.e. stereo) so
+    /// the stereo wire form stays byte-identical to the pre-surround build.
+    pub audio_channels: u8,
 }
 
 /// [`Hello::video_caps`] bit: the client can decode a 10-bit (Main10) HEVC stream.
 pub const VIDEO_CAP_10BIT: u8 = 0x01;
 /// [`Hello::video_caps`] bit: the client can present BT.2020 PQ HDR10 (implies 10-bit).
 pub const VIDEO_CAP_HDR: u8 = 0x02;
+/// [`Hello::video_caps`] bit: the client can decode a full-chroma **4:4:4** HEVC stream (HEVC
+/// Range Extensions / Rec.ITU-T H.265 `chroma_format_idc = 3`). The host emits 4:4:4 ONLY when this
+/// bit is set, the host opted in (`PUNKTFUNK_444`), the codec is HEVC, **and** the GPU/driver
+/// actually supports a 4:4:4 encode (probed) — otherwise the session stays 4:2:0 and
+/// [`Welcome::chroma_format`] reflects the real resolved value. Independent of 10-bit/HDR (4:4:4 is a
+/// chroma decision, bit depth is a depth decision; the two may combine where the hardware allows).
+pub const VIDEO_CAP_444: u8 = 0x04;
+
+/// HEVC `chroma_format_idc` for 4:2:0 — what every pre-4:4:4 build produced and the back-compat
+/// default when a peer omits [`Welcome::chroma_format`].
+pub const CHROMA_IDC_420: u8 = 1;
+/// HEVC `chroma_format_idc` for full-chroma 4:4:4 (Range Extensions).
+pub const CHROMA_IDC_444: u8 = 3;
 
 /// Per-session colour signalling (CICP / ITU-T H.273 code points) the host resolved for the
 /// encoded video, carried on [`Welcome`]. A client configures its decoder/presenter from these
@@ -198,6 +219,22 @@ pub struct Welcome {
     /// [`ColorInfo::SDR_BT709`]. The client configures its decoder/presenter from this instead of
     /// guessing from the bitstream; the mastering metadata arrives separately on [`HDR_META_MAGIC`].
     pub color: ColorInfo,
+    /// The chroma subsampling the host actually encodes at, as the HEVC `chroma_format_idc`:
+    /// [`CHROMA_IDC_420`] (4:2:0, default / older host) or [`CHROMA_IDC_444`] (full-chroma 4:4:4,
+    /// enabled only when the client advertised [`VIDEO_CAP_444`] *and* the host could open a real
+    /// 4:4:4 encode). The client sizes its decoder/surface pool from this; the in-band SPS carries
+    /// the authoritative value, so this is a hint (and the honest-downgrade channel — if the host
+    /// requested 4:4:4 but the GPU declined, this reads `CHROMA_IDC_420`). Appended after the colour
+    /// bytes as a single trailing byte; an older host that omits it decodes to [`CHROMA_IDC_420`].
+    pub chroma_format: u8,
+    /// The audio channel count the host actually resolved and **will** send on the `0xC9` plane:
+    /// `2` (stereo, default), `6` (5.1) or `8` (7.1). Echoes [`Hello::audio_channels`] clamped to
+    /// what the host can capture (Linux PipeWire always synthesizes the count; Windows WASAPI
+    /// loopback is clamped to the render endpoint's mix-format channels). The client builds its Opus
+    /// (multistream) decoder from THIS value via [`crate::audio::layout_for`] — never from its own
+    /// request — so an older host that omits the byte (→ `2`) always yields working stereo. Appended
+    /// after `chroma_format` as a single trailing byte.
+    pub audio_channels: u8,
 }
 
 /// `client → host`: data plane is bound, begin streaming.
@@ -630,10 +667,11 @@ impl Hello {
                                                                // so a Hello with neither name nor launch stays byte-identical to the bitrate-era form
                                                                // (26 bytes). When `launch` is present we must still emit name's length byte (0 for None)
                                                                // so `launch` lands at a deterministic offset.
-                                                               // `video_caps` is the last trailing field, after `launch`; when it's present (non-zero)
-                                                               // the name/launch length bytes must still be emitted (0 for absent) so it lands at a
+                                                               // `video_caps`/`audio_channels` are the trailing fields, after `launch`; when either is
+                                                               // present (video_caps non-zero / audio_channels not stereo) the name/launch length bytes
+                                                               // AND the video_caps byte must still be emitted (0 / 0) so the later byte lands at a
                                                                // deterministic offset — the same discipline `launch` already imposes on `name`.
-        let need_placeholders = self.video_caps != 0;
+        let need_placeholders = self.video_caps != 0 || self.audio_channels != 2;
         match (&self.name, &self.launch) {
             (None, None) if !need_placeholders => {}
             (name, _) => {
@@ -648,10 +686,15 @@ impl Hello {
             b.push(l.len() as u8);
             b.extend_from_slice(l.as_bytes());
         }
-        // video_caps: single trailing byte. Last field.
-        if self.video_caps != 0 {
+        // video_caps: single trailing byte. Emitted when non-zero OR when audio_channels follows
+        // (so audio_channels lands at a deterministic offset right after it).
+        if self.video_caps != 0 || self.audio_channels != 2 {
             b.push(self.video_caps);
         }
+        // audio_channels: single trailing byte. Last field; omitted when stereo (default).
+        if self.audio_channels != 2 {
+            b.push(self.audio_channels);
+        }
         b
     }
 
@@ -714,6 +757,15 @@ impl Hello {
                 let launch_len = b.get(launch_off).copied().unwrap_or(0) as usize;
                 b.get(launch_off + 1 + launch_len).copied().unwrap_or(0)
             },
+            // Optional trailing audio-channel byte, one past video_caps. Absent on an older client
+            // → stereo. Normalized so a corrupt/unsupported value can't build a bad decoder.
+            audio_channels: {
+                let name_len = b.get(26).copied().unwrap_or(0) as usize;
+                let launch_off = 27 + name_len;
+                let launch_len = b.get(launch_off).copied().unwrap_or(0) as usize;
+                let video_caps_off = launch_off + 1 + launch_len;
+                crate::audio::normalize_channels(b.get(video_caps_off + 1).copied().unwrap_or(2))
+            },
         })
     }
 }
@@ -747,6 +799,10 @@ impl Welcome {
         b.push(self.color.transfer);
         b.push(self.color.matrix);
         b.push(self.color.full_range);
+        // Chroma subsampling at offset 64 — older clients stop before this → 4:2:0 (CHROMA_IDC_420).
+        b.push(self.chroma_format);
+        // Audio channel count at offset 65 — older clients stop before this → stereo (2).
+        b.push(self.audio_channels);
         b
     }
 
@@ -755,7 +811,8 @@ impl Welcome {
         // scheme[22] pct[23] max_data[24..26] shard[26..28] encrypt[28] key[29..45]
         // salt[45..49] frames[49..53] compositor[53] gamepad[54] bitrate_kbps[55..59]
         // bit_depth[59] color.primaries[60] color.transfer[61] color.matrix[62] color.range[63]
-        // (everything from compositor on is an optional trailing byte; an older host stops earlier).
+        // chroma_format[64] audio_channels[65] (everything from compositor on is an optional
+        // trailing byte; an older host stops earlier).
         if b.len() < 53 || &b[0..4] != MAGIC {
             return Err(PunktfunkError::InvalidArg("bad Welcome"));
         }
@@ -812,6 +869,15 @@ impl Welcome {
                 matrix: b.get(62).copied().unwrap_or(ColorInfo::MC_BT709),
                 full_range: b.get(63).copied().unwrap_or(0),
             },
+            // Optional trailing chroma byte — absent on an older host (or an explicit 0 / unknown
+            // value) → 4:2:0. Only `CHROMA_IDC_444` flips the client to a 4:4:4 decode.
+            chroma_format: match b.get(64).copied() {
+                Some(CHROMA_IDC_444) => CHROMA_IDC_444,
+                _ => CHROMA_IDC_420,
+            },
+            // Optional trailing audio-channel byte — absent on an older host → stereo. Any
+            // non-{6,8} value normalizes to stereo so a corrupt byte never builds a bad decoder.
+            audio_channels: crate::audio::normalize_channels(b.get(65).copied().unwrap_or(2)),
         })
     }
 
@@ -1809,6 +1875,8 @@ mod tests {
             bitrate_kbps: 50_000,
             bit_depth: 10,
             color: ColorInfo::HDR10_BT2020_PQ,
+            chroma_format: CHROMA_IDC_444,
+            audio_channels: 2,
         };
         assert_eq!(Welcome::decode(&w.encode()).unwrap(), w);
     }
@@ -1851,6 +1919,7 @@ mod tests {
             name: Some("Test Device".into()),
             launch: Some("steam:570".into()),
             video_caps: VIDEO_CAP_10BIT,
+            audio_channels: 2,
         };
         assert_eq!(Hello::decode(&h.encode()).unwrap(), h);
         let s = Start {
@@ -1930,6 +1999,7 @@ mod tests {
             name: None,
             launch: None,
             video_caps: 0,
+            audio_channels: 2,
         };
         let enc = h.encode();
         assert_eq!(enc.len(), 26);
@@ -1969,9 +2039,11 @@ mod tests {
             bitrate_kbps: 120_000,
             bit_depth: 10,
             color: ColorInfo::HDR10_BT2020_PQ,
+            chroma_format: CHROMA_IDC_444,
+            audio_channels: 6, // 5.1 — exercises the non-default trailing byte
         };
         let wenc = w.encode();
-        assert_eq!(wenc.len(), 64); // 60 base + 4 colour bytes
+        assert_eq!(wenc.len(), 66); // 60 base + 4 colour + 1 chroma + 1 audio-channels byte
         let legacy_w = Welcome::decode(&wenc[..53]).unwrap();
         assert_eq!(legacy_w.compositor, CompositorPref::Auto);
         assert_eq!(legacy_w.gamepad, GamepadPref::Auto);
@@ -1991,13 +2063,29 @@ mod tests {
         let pre_color_w = Welcome::decode(&wenc[..60]).unwrap();
         assert_eq!(pre_color_w.bit_depth, 10);
         assert_eq!(pre_color_w.color, ColorInfo::SDR_BT709);
+        assert_eq!(pre_color_w.chroma_format, CHROMA_IDC_420); // pre-chroma host → 4:2:0
         assert_eq!(legacy_w.color, ColorInfo::SDR_BT709);
+        assert_eq!(legacy_w.chroma_format, CHROMA_IDC_420);
+        // A pre-chroma (64-byte) Welcome carries colour but no chroma/audio bytes → 4:2:0 + stereo.
+        let pre_chroma_w = Welcome::decode(&wenc[..64]).unwrap();
+        assert_eq!(pre_chroma_w.color, ColorInfo::HDR10_BT2020_PQ);
+        assert_eq!(pre_chroma_w.chroma_format, CHROMA_IDC_420);
+        assert_eq!(pre_chroma_w.audio_channels, 2); // audio byte (offset 65) absent → stereo
+                                                    // A pre-audio (65-byte) Welcome carries chroma but no audio byte → 4:4:4 + stereo.
+        let pre_audio_w = Welcome::decode(&wenc[..65]).unwrap();
+        assert_eq!(pre_audio_w.chroma_format, CHROMA_IDC_444);
+        assert_eq!(pre_audio_w.audio_channels, 2);
         assert_eq!(Welcome::decode(&wenc).unwrap().bitrate_kbps, 120_000);
         assert_eq!(Welcome::decode(&wenc).unwrap().bit_depth, 10); // full form carries it
         assert_eq!(
             Welcome::decode(&wenc).unwrap().color,
             ColorInfo::HDR10_BT2020_PQ
         );
+        assert_eq!(
+            Welcome::decode(&wenc).unwrap().chroma_format,
+            CHROMA_IDC_444
+        ); // full form carries 4:4:4
+        assert_eq!(Welcome::decode(&wenc).unwrap().audio_channels, 6); // ...and 5.1
     }
 
     #[test]
@@ -2015,6 +2103,7 @@ mod tests {
             name: Some("Enrico's MacBook".into()),
             launch: None,
             video_caps: 0,
+            audio_channels: 2,
         };
         let enc = base.encode();
         assert_eq!(
@@ -2062,6 +2151,7 @@ mod tests {
             name: None,
             launch: None,
             video_caps: 0,
+            audio_channels: 2,
         };
         // launch alone (no name): a zero-length name placeholder keeps the offset deterministic.
         let with_launch = Hello {
@@ -2268,6 +2358,7 @@ mod tests {
                 name: None,
                 launch: None,
                 video_caps: 0,
+                audio_channels: 2,
             }
             .encode();
             assert!(PairRequest::decode(&h).is_err(), "abi {abi} parsed as pair");

crates/punktfunk-core/tests/c_abi.rs

@@ -13,8 +13,10 @@ use std::process::Command;
 fn native_libs() -> &'static [&'static str] {
     if cfg!(target_os = "macos") {
         // The workspace build unifies features into the staticlib, and `quic` pulls
-        // rustls's platform verifier → Security/CoreFoundation.
+        // rustls's platform verifier → Security/CoreFoundation, plus libopus (the in-core
+        // `next_audio_pcm` decode path) which the `abi.rs` object references.
         &[
+            "-lopus",
             "-liconv",
             "-lm",
             "-framework",
@@ -23,7 +25,17 @@ fn native_libs() -> &'static [&'static str] {
             "CoreFoundation",
         ]
     } else if cfg!(target_os = "linux") {
-        &["-lgcc_s", "-lutil", "-lrt", "-lpthread", "-lm", "-ldl"]
+        // `-lopus`: the `quic` feature pulls in-core Opus decode (`next_audio_pcm`), whose
+        // symbols the linked `abi.rs` object references. Before `-lm` (opus needs libm).
+        &[
+            "-lopus",
+            "-lgcc_s",
+            "-lutil",
+            "-lrt",
+            "-lpthread",
+            "-lm",
+            "-ldl",
+        ]
     } else {
         &[]
     }

crates/punktfunk-host/Cargo.toml

@@ -61,9 +61,10 @@ utoipa-scalar = { version = "0.3", features = ["axum"] }
 tower = { version = "0.5", features = ["util"] }
 http-body-util = "0.1"
 
-# Opus stereo encode for the host->client audio plane. The `opus` crate vendors libopus via
-# `audiopus_sys` (cmake-built from source — no system lib, no vcpkg), so it builds on Windows MSVC
-# too (needs CMake + NASM, both on the box). Both platforms that have an audio-capture backend.
+# Opus encode for the host->client audio plane — stereo (`opus::Encoder`) AND 5.1/7.1 surround
+# (`opus::MSEncoder`, the safe multistream API the crate exposes; no `audiopus_sys` needed). The
+# crate vendors libopus (cmake-built from source — no system lib, no vcpkg), so it builds on Windows
+# MSVC too (needs CMake + NASM, both on the box). Both platforms that have an audio-capture backend.
 [target.'cfg(any(target_os = "linux", target_os = "windows"))'.dependencies]
 opus = "0.3"
 
@@ -99,10 +100,6 @@ serde_json = "1"
 rusqlite = { version = "0.40", features = ["bundled"] }
 # Builds/validates the xkb keymap uploaded to the virtual keyboard + tracks modifier state.
 xkbcommon = "0.8"
-# The safe `opus` crate is stereo-only; surround (5.1/7.1) needs the libopus *multistream*
-# encoder (`opus_multistream_encoder_*`). `audiopus_sys` is the sys layer `opus` already
-# vendors (same libopus link), so this adds bindings, not a second copy of the library.
-audiopus_sys = "0.2"
 # libei (EI sender) for the portable input path on KWin/GNOME (RemoteDesktop portal).
 # The `tokio` feature wires reis's event stream into tokio's reactor.
 reis = { version = "0.6.1", features = ["tokio"] }

crates/punktfunk-host/protocols/fake-input.xml

@@ -0,0 +1,73 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<protocol name="fake_input">
+  <copyright>
+    SPDX-FileCopyrightText: 2015 Martin Gräßlin
+    SPDX-License-Identifier: LGPL-2.1-or-later
+  </copyright>
+  <interface name="org_kde_kwin_fake_input" version="4">
+    <description summary="Fake input manager">
+      This interface allows other processes to provide fake input events.
+      Purpose is on the one hand side to provide testing facilities like XTest
+      on X11, but also to support use cases like remote control (a remote
+      desktop server). The compositor gates the interface: it is only exposed
+      to clients authorized through their .desktop X-KDE-Wayland-Interfaces, so
+      binding it is the authorization — no per-event confirmation dialog.
+    </description>
+    <request name="authenticate">
+      <description summary="Information about the application requesting fake input">
+        A FakeInput is required to authenticate itself by providing the
+        application name and the reason for fake input. The compositor may use
+        this information to decide whether to allow or deny the request.
+      </description>
+      <arg name="application" type="string" summary="user visible name of the application requesting fake input"/>
+      <arg name="reason" type="string" summary="reason of why fake input is requested"/>
+    </request>
+    <request name="pointer_motion">
+      <description summary="pointer motion event"/>
+      <arg name="delta_x" type="fixed" summary="X delta of the relative pointer motion"/>
+      <arg name="delta_y" type="fixed" summary="Y delta of the relative pointer motion"/>
+    </request>
+    <request name="button">
+      <description summary="pointer button event"/>
+      <arg name="button" type="uint" summary="evdev button code"/>
+      <arg name="state" type="uint" summary="button state, 0 released, 1 pressed"/>
+    </request>
+    <request name="axis">
+      <description summary="pointer axis (scroll) event"/>
+      <arg name="axis" type="uint" summary="wl_pointer.axis (0 vertical, 1 horizontal)"/>
+      <arg name="value" type="fixed" summary="axis value"/>
+    </request>
+    <request name="touch_down" since="2">
+      <description summary="touch down event"/>
+      <arg name="id" type="uint" summary="unique id of this touch point; must not be reused until up"/>
+      <arg name="x" type="fixed" summary="x coordinate in global compositor space"/>
+      <arg name="y" type="fixed" summary="y coordinate in global compositor space"/>
+    </request>
+    <request name="touch_motion" since="2">
+      <description summary="touch motion event"/>
+      <arg name="id" type="uint" summary="unique id of an existing touch point"/>
+      <arg name="x" type="fixed" summary="x coordinate in global compositor space"/>
+      <arg name="y" type="fixed" summary="y coordinate in global compositor space"/>
+    </request>
+    <request name="touch_up" since="2">
+      <description summary="touch up event"/>
+      <arg name="id" type="uint" summary="unique id of an existing touch point"/>
+    </request>
+    <request name="touch_cancel" since="2">
+      <description summary="cancel all current touch points"/>
+    </request>
+    <request name="touch_frame" since="2">
+      <description summary="end a set of touch events (atomic frame)"/>
+    </request>
+    <request name="pointer_motion_absolute" since="3">
+      <description summary="absolute pointer motion event"/>
+      <arg name="x" type="fixed" summary="x coordinate in global compositor space"/>
+      <arg name="y" type="fixed" summary="y coordinate in global compositor space"/>
+    </request>
+    <request name="keyboard_key" since="4">
+      <description summary="keyboard key event"/>
+      <arg name="button" type="uint" summary="evdev key code"/>
+      <arg name="state" type="uint" summary="key state, 0 released, 1 pressed"/>
+    </request>
+  </interface>
+</protocol>

crates/punktfunk-host/src/audio/linux/mod.rs

@@ -320,11 +320,18 @@ fn mic_pw_thread(
     .into_inner();
     let mut params = [Pod::from_bytes(&values).context("mic pod from bytes")?];
 
+    // RT_PROCESS: run the producer callback on PipeWire's realtime data loop, so the source is a
+    // *synchronous* graph node that joins its consumer's driver group and is actually driven. Without
+    // it the node is async/main-loop and, in the host's busy multi-stream graph (desktop-audio +
+    // video capture + the session), never acquires a driver — it stays suspended and its process()
+    // never fires, so every recorder hears pure silence (the long-standing "Linux host mic broken").
     stream
         .connect(
             spa::utils::Direction::Output, // we PRODUCE samples (a source)
             None,
-            pw::stream::StreamFlags::AUTOCONNECT | pw::stream::StreamFlags::MAP_BUFFERS,
+            pw::stream::StreamFlags::AUTOCONNECT
+                | pw::stream::StreamFlags::MAP_BUFFERS
+                | pw::stream::StreamFlags::RT_PROCESS,
             &mut params,
         )
         .context("pw mic stream connect")?;

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

@@ -1,7 +1,9 @@
 //! WASAPI loopback capture of the default render endpoint (system output) — the Windows analogue
-//! of the PipeWire sink-monitor backend. Delivers interleaved f32 PCM at 48 kHz stereo, ready for
-//! the existing Opus path with NO resampling (WASAPI shared-mode autoconvert does any SRC). WASAPI
-//! objects are COM-apartment-bound and not `Send`, so they live on a dedicated thread (mirrors
+//! of the PipeWire sink-monitor backend. Delivers interleaved f32 PCM at 48 kHz in the requested
+//! channel count (stereo / 5.1 / 7.1, canonical wire order FL FR FC LFE RL RR SL SR via the
+//! explicit `dwChannelMask`), ready for the Opus path with NO resampling (WASAPI shared-mode
+//! autoconvert does any SRC + up/downmix to the requested layout). WASAPI objects are
+//! COM-apartment-bound and not `Send`, so they live on a dedicated thread (mirrors
 //! `linux::PwAudioCapturer`); only the channel + stop flag + join handle are in the struct.
 
 use super::{AudioCapturer, SAMPLE_RATE};
@@ -14,9 +16,6 @@ use std::thread::{self, JoinHandle};
 use std::time::Duration;
 use wasapi::{DeviceEnumerator, Direction, SampleType, StreamMode, WaveFormat};
 
-// 48 kHz stereo 32-bit float: 2 channels * 4 bytes = 8 bytes per frame.
-const BLOCK_ALIGN: usize = 2 * 4;
-
 pub struct WasapiLoopbackCapturer {
     chunks: Receiver<Vec<f32>>,
     channels: u32,
@@ -27,8 +26,8 @@ pub struct WasapiLoopbackCapturer {
 impl WasapiLoopbackCapturer {
     pub fn open(channels: u32) -> Result<WasapiLoopbackCapturer> {
         anyhow::ensure!(
-            channels == 2,
-            "WASAPI loopback backend is stereo-only (got {channels})"
+            matches!(channels, 2 | 6 | 8),
+            "WASAPI loopback backend supports 2/6/8 channels (got {channels})"
         );
         let (tx, rx) = sync_channel::<Vec<f32>>(64);
         let stop = Arc::new(AtomicBool::new(false));
@@ -39,7 +38,7 @@ impl WasapiLoopbackCapturer {
         let join = thread::Builder::new()
             .name("punktfunk-wasapi-audio".into())
             .spawn(move || {
-                if let Err(e) = capture_thread(tx, stop_t, ready_tx) {
+                if let Err(e) = capture_thread(tx, stop_t, ready_tx, channels) {
                     tracing::error!(error = format!("{e:#}"), "wasapi loopback thread failed");
                 }
             })
@@ -47,7 +46,8 @@ impl WasapiLoopbackCapturer {
         match ready_rx.recv_timeout(Duration::from_secs(3)) {
             Ok(Ok(())) => {
                 tracing::info!(
-                    "WASAPI loopback capture: 48 kHz stereo f32 (default render endpoint)"
+                    channels,
+                    "WASAPI loopback capture: 48 kHz f32 (default render endpoint)"
                 );
                 Ok(WasapiLoopbackCapturer {
                     chunks: rx,
@@ -95,7 +95,10 @@ fn capture_thread(
     tx: SyncSender<Vec<f32>>,
     stop: Arc<AtomicBool>,
     ready: SyncSender<Result<()>>,
+    channels: u32,
 ) -> Result<()> {
+    // Interleaved f32: channels * 4 bytes per frame.
+    let block_align = channels as usize * 4;
     // COM must be initialized on THIS thread (MTA), before any device call.
     if let Err(e) = wasapi::initialize_mta()
         .ok()
@@ -106,16 +109,29 @@ fn capture_thread(
     }
     let res = (|| -> Result<()> {
         // Loopback = capture the RENDER endpoint: get the default render device, but open a CAPTURE
-        // client with loopback=true over it.
+        // client with loopback=true over it. NOTE: the virtual mic (`super::wasapi_mic`) is guarded
+        // to NEVER target this same endpoint — otherwise the client's injected mic would be captured
+        // here and streamed back to the client (infinite echo). Keep that guard in sync if this
+        // device selection ever changes.
         let device = DeviceEnumerator::new()
             .context("DeviceEnumerator")?
             .get_default_device(&Direction::Render)
             .context("default render endpoint (loopback needs a render device)")?;
         let mut audio_client = device.get_iaudioclient().context("IAudioClient")?;
-        // 48 kHz stereo f32 interleaved; autoconvert lets WASAPI's shared-mode SRC match the engine
-        // mix format to ours, so we never resample in Rust. Loopback is implied by capturing a
-        // RENDER device with Direction::Capture in shared mode (wasapi sets STREAMFLAGS_LOOPBACK).
-        let desired = WaveFormat::new(32, 32, &SampleType::Float, SAMPLE_RATE as usize, 2, None);
+        // 48 kHz f32 interleaved in the requested channel layout; autoconvert lets WASAPI's
+        // shared-mode SRC match the engine mix format to ours (incl. up/downmix to the requested
+        // channel count), so we never resample/remix in Rust. The explicit dwChannelMask pins the
+        // wire order (FL FR FC LFE RL RR SL SR; 7.1 = 0x63F, not 0xFF). Loopback is implied by
+        // capturing a RENDER device with Direction::Capture in shared mode (STREAMFLAGS_LOOPBACK).
+        let mask = punktfunk_core::audio::wasapi_channel_mask(channels as u8);
+        let desired = WaveFormat::new(
+            32,
+            32,
+            &SampleType::Float,
+            SAMPLE_RATE as usize,
+            channels as usize,
+            Some(mask),
+        );
         let (default_period, _min_period) =
             audio_client.get_device_period().context("device period")?;
         let mode = StreamMode::EventsShared {
@@ -151,7 +167,7 @@ fn capture_thread(
                     Err(e) => return Err(anyhow!("get_next_packet_size: {e}")),
                 }
             }
-            let whole = (bytes.len() / BLOCK_ALIGN) * BLOCK_ALIGN;
+            let whole = (bytes.len() / block_align) * block_align;
             if whole == 0 {
                 continue;
             }

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

@@ -5,8 +5,18 @@
 //!
 //! 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 return an
-//! error with install guidance and the host runs without mic passthrough.
+//! "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.
+//!
+//! **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
+//! captures the *mixed* output of an endpoint — i.e. everything any app renders to it, including
+//! what THIS module writes. So if the virtual-mic target is the same device the loopback captures,
+//! the client's uplinked mic is captured straight back into the host→client audio stream: an
+//! infinite echo. [`find_device`] therefore **excludes the default render endpoint** from the
+//! candidates — the mic is guaranteed to land on a different device. (Linux gets this for free: its
+//! mic is a dedicated `Audio/Source` node, structurally separate from the monitored sink.)
 //!
 //! `push` enqueues decoded interleaved-f32 PCM into a bounded ring (drop-oldest beyond ~80 ms so mic
 //! latency stays bounded); a dedicated COM-apartment thread renders it event-driven, filling silence
@@ -113,8 +123,23 @@ impl VirtualMic for WasapiVirtualMic {
     }
 }
 
-/// Resolve the virtual-mic target among render endpoints by friendly-name. Logs all candidates so a
-/// missing device is diagnosable.
+/// The endpoint ID of the device the desktop-audio loopback records (the **default render
+/// endpoint**, see [`super::wasapi_cap`]). The virtual mic must never target this device — injecting
+/// there echoes the client's mic back into the host→client audio stream. `None` if it can't be
+/// resolved (then [`find_device`] can't prove a candidate is safe and falls back to name-only
+/// matching — no worse than before the guard existed).
+fn default_render_id() -> Option<String> {
+    wasapi::DeviceEnumerator::new()
+        .ok()?
+        .get_default_device(&Direction::Render)
+        .ok()?
+        .get_id()
+        .ok()
+}
+
+/// Resolve the virtual-mic target among render endpoints by friendly-name, **excluding the endpoint
+/// the loopback captures** (the [`default_render_id`] anti-echo guard). Logs all candidates so a
+/// missing/skipped device is diagnosable.
 fn find_device() -> Result<wasapi::Device> {
     let enumerator = wasapi::DeviceEnumerator::new().context("DeviceEnumerator")?;
     let collection = enumerator
@@ -124,8 +149,11 @@ fn find_device() -> Result<wasapi::Device> {
     let want = std::env::var("PUNKTFUNK_MIC_DEVICE")
         .ok()
         .map(|s| s.to_lowercase());
+    // The device the loopback captures — a name match on it is rejected below (would echo).
+    let loopback_id = default_render_id();
     let mut names = Vec::new();
     let mut found = None;
+    let mut skipped_loopback = false;
     for i in 0..n {
         let Ok(dev) = collection.get_device_at_index(i) else {
             continue;
@@ -137,16 +165,37 @@ fn find_device() -> Result<wasapi::Device> {
             None => CANDIDATES.iter().any(|c| lname.contains(c)),
         };
         if hit && found.is_none() {
-            found = Some(dev);
+            // Anti-echo guard: never inject into the endpoint the loopback captures.
+            let is_loopback = match (dev.get_id().ok(), loopback_id.as_deref()) {
+                (Some(id), Some(lb)) => id == lb,
+                _ => false,
+            };
+            if is_loopback {
+                skipped_loopback = true;
+                tracing::warn!(device = %name,
+                    "virtual-mic candidate is the loopback (default render) endpoint — skipping; \
+                     injecting there would echo the client's mic into the desktop-audio stream");
+            } else {
+                found = Some(dev);
+            }
         }
         names.push(name);
     }
     found.ok_or_else(|| {
-        anyhow!(
-            "no virtual-mic device among render endpoints {names:?}. Install VB-Audio Virtual Cable \
-             or enable Steam Remote Play's microphone (Steam Streaming Microphone), or set \
-             PUNKTFUNK_MIC_DEVICE=<friendly-name substring>."
-        )
+        if skipped_loopback {
+            anyhow!(
+                "the only virtual-mic candidate among render endpoints {names:?} is the default \
+                 playback device the host loopback-captures — injecting there would echo the mic \
+                 back to the client. Add a SEPARATE virtual audio device for the mic (e.g. the Steam \
+                 Streaming Microphone) or set a different default playback device, then reconnect."
+            )
+        } else {
+            anyhow!(
+                "no virtual-mic device among render endpoints {names:?}. Install VB-Audio Virtual \
+                 Cable or enable Steam Remote Play's microphone (Steam Streaming Microphone), or set \
+                 PUNKTFUNK_MIC_DEVICE=<friendly-name substring>."
+            )
+        }
     })
 }
 
@@ -156,15 +205,15 @@ fn find_or_install_device() -> Result<wasapi::Device> {
     match find_device() {
         Ok(d) => Ok(d),
         Err(e) => {
-            tracing::info!("no virtual mic device present — attempting auto-install");
-            // SAFETY: `try_install_virtual_mic` is `unsafe` only because it `LoadLibraryExW`s
+            tracing::info!("no usable virtual mic device present — attempting auto-install");
+            // SAFETY: `install_steam_audio_pair` is `unsafe` only because it `LoadLibraryExW`s
             // `newdev.dll` and calls `DiInstallDriverW` through a `transmute`d function pointer;
             // calling it imposes no extra precondition here (it takes no args and aliases nothing).
             // Its internal contract holds: the `DiInstall` type matches the documented
             // `BOOL DiInstallDriverW(HWND, PCWSTR, DWORD, PBOOL)` ABI, and it passes a
             // NUL-terminated UTF-16 INF path with null/zero optional args. Invoked once on the
             // dedicated mic thread.
-            if unsafe { try_install_virtual_mic() } {
+            if unsafe { install_steam_audio_pair() } {
                 find_device()
             } else {
                 Err(e)
@@ -173,13 +222,26 @@ fn find_or_install_device() -> Result<wasapi::Device> {
     }
 }
 
-/// Best-effort: install a virtual mic device so one exists without the user installing anything.
-/// Mirrors Apollo's Steam Streaming Speakers install — Steam Remote Play ships
-/// `SteamStreamingMicrophone.inf` next to the speakers INF, so install it via `DiInstallDriverW`
-/// (loaded from `newdev.dll`, like Apollo, to avoid an extra windows-crate feature). Needs admin (the
-/// host runs as SYSTEM). Returns true on success; false (no-op) if Steam isn't installed (INF absent),
-/// the install is denied, or `PUNKTFUNK_NO_MIC_INSTALL` is set.
-unsafe fn try_install_virtual_mic() -> bool {
+/// Best-effort: install BOTH Steam Streaming audio devices (the "Steam pair") so mic passthrough
+/// works out of the box and the host has a desktop-audio sink distinct from the mic. Steam Remote
+/// Play ships `SteamStreamingMicrophone.inf` + `SteamStreamingSpeakers.inf`: the microphone gives the
+/// virtual mic a target whose **capture** endpoint apps record from, and the speakers give a
+/// **render** endpoint a headless box can loopback-capture that is NOT the mic — so the loopback and
+/// the mic land on different devices and never echo (see [`find_device`]). Returns true if either
+/// installed. No-op when Steam isn't installed (INFs absent), the install is denied (needs admin —
+/// the host runs as SYSTEM), or `PUNKTFUNK_NO_MIC_INSTALL` is set.
+unsafe fn install_steam_audio_pair() -> bool {
+    // Microphone first (the mic's actual target); speakers second (the distinct desktop-audio sink).
+    let mic = try_install_steam_audio("SteamStreamingMicrophone.inf");
+    let spk = try_install_steam_audio("SteamStreamingSpeakers.inf");
+    mic || spk
+}
+
+/// Install one Steam Streaming driver INF by filename via `DiInstallDriverW` (loaded from
+/// `newdev.dll`, like Apollo, to avoid an extra windows-crate feature). See
+/// [`install_steam_audio_pair`] for the contract; `inf_name` is a bare filename under Steam's
+/// per-arch `drivers\Windows10\{arch}\` directory.
+unsafe fn try_install_steam_audio(inf_name: &str) -> bool {
     use windows::core::{s, w, PCWSTR};
     use windows::Win32::Foundation::HWND;
     use windows::Win32::System::Environment::ExpandEnvironmentStringsW;
@@ -197,12 +259,11 @@ unsafe fn try_install_virtual_mic() -> bool {
     let subdir = "arm64";
     #[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
     let subdir = "x86";
-    let template: Vec<u16> = format!(
-        "%CommonProgramFiles(x86)%\\Steam\\drivers\\Windows10\\{subdir}\\SteamStreamingMicrophone.inf"
-    )
-    .encode_utf16()
-    .chain(std::iter::once(0))
-    .collect();
+    let template: Vec<u16> =
+        format!("%CommonProgramFiles(x86)%\\Steam\\drivers\\Windows10\\{subdir}\\{inf_name}")
+            .encode_utf16()
+            .chain(std::iter::once(0))
+            .collect();
     let mut path = vec![0u16; 1024];
     let n = ExpandEnvironmentStringsW(PCWSTR(template.as_ptr()), Some(path.as_mut_slice()));
     if n == 0 || n as usize > path.len() {
@@ -210,7 +271,7 @@ unsafe fn try_install_virtual_mic() -> bool {
     }
 
     let Ok(newdev) = LoadLibraryExW(w!("newdev.dll"), None, LOAD_LIBRARY_SEARCH_SYSTEM32) else {
-        tracing::warn!("could not load newdev.dll — virtual-mic auto-install unavailable");
+        tracing::warn!("could not load newdev.dll — Steam-audio auto-install unavailable");
         return false;
     };
     let Some(addr) = GetProcAddress(newdev, s!("DiInstallDriverW")) else {
@@ -226,13 +287,17 @@ unsafe fn try_install_virtual_mic() -> bool {
         std::ptr::null_mut(),
     ) != 0;
     if ok {
-        tracing::info!("installed the Steam Streaming Microphone virtual device");
+        tracing::info!(
+            inf = inf_name,
+            "installed a Steam Streaming virtual audio device"
+        );
         std::thread::sleep(Duration::from_secs(5)); // let the audio subsystem register the endpoint
     } else {
         let err = windows::Win32::Foundation::GetLastError();
         tracing::info!(
+            inf = inf_name,
             ?err,
-            "no virtual mic auto-installed (Steam absent / not admin) — see manual-install guidance"
+            "Steam-audio device not auto-installed (Steam absent / not admin) — see install guidance"
         );
     }
     ok

crates/punktfunk-host/src/capture.rs

@@ -62,6 +62,11 @@ pub struct OutputFormat {
     /// HDR: the capturer converts to 10-bit (IDD-push FP16 → `Rgb10a2`; the DDA secure-desktop HDR hint).
     /// `false` = 8-bit SDR.
     pub hdr: bool,
+    /// Full-chroma 4:4:4 session: the capturer must keep full chroma — deliver packed **RGB**
+    /// (`Bgra` / `Rgb10a2`), NOT the subsampled `Nv12`/`P010` the Windows video-engine path produces by
+    /// default — because 4:4:4 can only be recovered from a full-chroma source. NVENC then does the
+    /// RGB→YUV444 CSC at encode (chroma_format_idc=3). `false` on every 4:2:0 session.
+    pub chroma_444: bool,
 }
 
 impl OutputFormat {
@@ -73,6 +78,8 @@ impl OutputFormat {
         OutputFormat {
             gpu: gpu_encode(),
             hdr,
+            // The GameStream + spike paths are always 4:2:0 (4:4:4 is punktfunk/1-native only).
+            chroma_444: false,
         }
     }
 }
@@ -361,13 +368,16 @@ pub fn open_portal_monitor() -> Result<Box<dyn Capturer>> {
 #[cfg(target_os = "linux")]
 pub fn capture_virtual_output(
     vout: crate::vdisplay::VirtualOutput,
-    _want: OutputFormat,
+    want: OutputFormat,
     _capture: crate::session_plan::CaptureBackend,
 ) -> Result<Box<dyn Capturer>> {
-    // The Linux host stays 8-bit (HDR is blocked upstream) and the portal negotiates its own format, so
-    // the `OutputFormat` is unused here; the capture backend is always the portal (the `CaptureBackend`
-    // arg is a Windows-only dispatch — ignored here).
-    linux::PortalCapturer::from_virtual_output(vout).map(|c| Box::new(c) as Box<dyn Capturer>)
+    // The Linux host stays 8-bit (HDR is blocked upstream) and the portal negotiates its own pixel
+    // format, so only `want.gpu` is honored here: it gates GPU zero-copy capture (the capture backend
+    // is always the portal — the `CaptureBackend` arg is a Windows-only dispatch). `gpu = false`
+    // (a 4:4:4 NVENC session) forces the CPU mmap path so the encoder gets CPU-resident RGB to swscale
+    // into YUV444P — otherwise it would receive CUDA frames and bail.
+    linux::PortalCapturer::from_virtual_output(vout, want.gpu)
+        .map(|c| Box::new(c) as Box<dyn Capturer>)
 }
 
 /// `PUNKTFUNK_NO_WGC=1` forces the pure single-process DDA (Desktop Duplication) path everywhere: it
@@ -394,6 +404,14 @@ 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
@@ -414,8 +432,15 @@ pub fn capture_virtual_output(
                     error = %format!("{e:#}"),
                     "IDD-push open/attach failed — falling back to DDA"
                 );
-                return dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false)
-                    .map(|c| Box::new(c) as Box<dyn Capturer>);
+                return dxgi::DuplCapturer::open(
+                    target,
+                    pref,
+                    keep,
+                    want.gpu,
+                    false,
+                    want.chroma_444,
+                )
+                .map(|c| Box::new(c) as Box<dyn Capturer>);
             }
         }
     }
@@ -426,7 +451,7 @@ pub fn capture_virtual_output(
     // 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)
+        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
@@ -461,12 +486,12 @@ pub fn capture_virtual_output(
         }
         Ok(Err(e)) => {
             tracing::warn!(error = %format!("{e:#}"), "WGC open failed — falling back to DDA");
-            dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false)
+            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)
+            dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
                 .map(|c| Box::new(c) as Box<dyn Capturer>)
         }
     }

crates/punktfunk-host/src/capture/linux/mod.rs

@@ -40,6 +40,13 @@ pub struct PortalCapturer {
     /// branch to tell "format never negotiated" (modifier/format mismatch) apart from "negotiated
     /// but no buffers arrived" (compositor idle/unmapped) — the two black-screen root causes.
     negotiated: Arc<AtomicBool>,
+    /// True only while the PipeWire stream is `Streaming`. [`try_latest`](Self::try_latest) reads it
+    /// to distinguish a static desktop (alive, no new buffers) from a dead source (left `Streaming`).
+    streaming: Arc<AtomicBool>,
+    /// When the stream first dropped out of `Streaming` with no new frame; used to grace a transient
+    /// renegotiation before declaring the source lost. Cleared whenever a frame arrives or the stream
+    /// is `Streaming`.
+    stall_since: Option<std::time::Instant>,
     /// The PipeWire node this capturer consumes — surfaced in error messages for diagnosis.
     node_id: u32,
     /// Stops the PipeWire loop on teardown (sent in `Drop`). Without it a dropped or failed
@@ -82,21 +89,29 @@ impl PortalCapturer {
             node_id,
             "ScreenCast portal session started; connecting PipeWire"
         );
-        Ok(spawn_pipewire(Some(fd), node_id, None)?.into_capturer(node_id, None))
+        // This portal path (GameStream / monitor capture) is always 4:2:0, so allow zero-copy as before.
+        Ok(spawn_pipewire(Some(fd), node_id, None, true)?.into_capturer(node_id, None))
     }
 
     /// Build a capturer from an already-created virtual output ([`crate::vdisplay::VirtualOutput`]):
     /// connect PipeWire to its node (`remote_fd` selects portal-remote vs. default-daemon) and
     /// take ownership of its keepalive so the output lives exactly as long as this capturer. This
     /// is how the client's requested resolution becomes the captured resolution without scaling.
-    pub fn from_virtual_output(vout: crate::vdisplay::VirtualOutput) -> Result<PortalCapturer> {
+    /// `allow_zerocopy` mirrors [`OutputFormat::gpu`](crate::capture::OutputFormat): `false` forces the
+    /// CPU mmap path (a 4:4:4 NVENC session needs CPU-resident RGB), `true` keeps the GPU zero-copy
+    /// path subject to `PUNKTFUNK_ZEROCOPY`.
+    pub fn from_virtual_output(
+        vout: crate::vdisplay::VirtualOutput,
+        allow_zerocopy: bool,
+    ) -> Result<PortalCapturer> {
         tracing::info!(
             node_id = vout.node_id,
+            allow_zerocopy,
             "connecting PipeWire to virtual output"
         );
         let node_id = vout.node_id;
         Ok(
-            spawn_pipewire(vout.remote_fd, node_id, vout.preferred_mode)?
+            spawn_pipewire(vout.remote_fd, node_id, vout.preferred_mode, allow_zerocopy)?
                 .into_capturer(node_id, Some(vout.keepalive)),
         )
     }
@@ -109,6 +124,7 @@ struct PwHandles {
     frames: Receiver<CapturedFrame>,
     active: Arc<AtomicBool>,
     negotiated: Arc<AtomicBool>,
+    streaming: Arc<AtomicBool>,
     quit: ::pipewire::channel::Sender<()>,
     join: thread::JoinHandle<()>,
 }
@@ -121,6 +137,8 @@ impl PwHandles {
             frames: self.frames,
             active: self.active,
             negotiated: self.negotiated,
+            streaming: self.streaming,
+            stall_since: None,
             node_id,
             quit: Some(self.quit),
             join: Some(self.join),
@@ -136,6 +154,12 @@ fn spawn_pipewire(
     fd: Option<OwnedFd>,
     node_id: u32,
     preferred: Option<(u32, u32, u32)>,
+    // Allow GPU zero-copy capture (dmabuf→CUDA/VA). `false` forces the CPU mmap path even when
+    // `PUNKTFUNK_ZEROCOPY` is set — a 4:4:4 NVENC session needs CPU-resident RGB (the encoder
+    // swscales RGB→YUV444P; `hevc_nvenc` can't 4:4:4 from a CUDA RGB surface), so the session plan
+    // passes `gpu = false` for it. Without this, a 4:4:4 session under `PUNKTFUNK_ZEROCOPY=1` would
+    // get CUDA frames and the encoder would bail (`want_444 && cuda`).
+    allow_zerocopy: bool,
 ) -> Result<PwHandles> {
     // Frames flow from the pipewire thread over a small bounded channel.
     let (frame_tx, frame_rx) = sync_channel::<CapturedFrame>(8);
@@ -143,11 +167,13 @@ fn spawn_pipewire(
     let active_cb = active.clone();
     let negotiated = Arc::new(AtomicBool::new(false));
     let negotiated_cb = negotiated.clone();
+    let streaming = Arc::new(AtomicBool::new(false));
+    let streaming_cb = streaming.clone();
     // pipewire's own cross-thread channel: the receiver attaches to the loop and quits it; the
     // sender lives on the capturer and fires in its `Drop`. Absolute `::pipewire` path — the
     // inner `mod pipewire` shadows the crate name at this scope.
     let (quit_tx, quit_rx) = ::pipewire::channel::channel::<()>();
-    let zerocopy = crate::zerocopy::enabled();
+    let zerocopy = allow_zerocopy && crate::zerocopy::enabled();
     let join = thread::Builder::new()
         .name("punktfunk-pipewire".into())
         .spawn(move || {
@@ -157,6 +183,7 @@ fn spawn_pipewire(
                 frame_tx,
                 active_cb,
                 negotiated_cb,
+                streaming_cb,
                 zerocopy,
                 preferred,
                 quit_rx,
@@ -169,6 +196,7 @@ fn spawn_pipewire(
         frames: frame_rx,
         active,
         negotiated,
+        streaming,
         quit: quit_tx,
         join,
     })
@@ -219,6 +247,28 @@ impl Capturer for PortalCapturer {
                 }
             }
         }
+        if latest.is_some() || self.streaming.load(Ordering::Relaxed) {
+            // A frame arrived, or the source is alive but idle (static desktop) — normal. Clear any
+            // stall and repeat the last frame on `None`, exactly as before.
+            self.stall_since = None;
+            return Ok(latest);
+        }
+        // No new frame AND the stream has left `Streaming` (Paused/Unconnected/Error). The source
+        // went away — a compositor torn down on a Gaming↔Desktop switch, a removed virtual output.
+        // Grace a brief window (a transient mid-stream renegotiation can blip out of Streaming and
+        // back) before declaring it lost so the encode loop rebuilds in place rather than freezing
+        // on the last frame forever.
+        const STALL_GRACE: Duration = Duration::from_millis(1500);
+        let since = *self.stall_since.get_or_insert_with(std::time::Instant::now);
+        if since.elapsed() >= STALL_GRACE {
+            self.stall_since = None;
+            return Err(anyhow!(
+                "PipeWire source stalled (node {}): stream left Streaming for >{}ms with no frames \
+                 — the compositor/virtual output went away (session switch?)",
+                self.node_id,
+                STALL_GRACE.as_millis()
+            ));
+        }
         Ok(latest)
     }
 
@@ -467,6 +517,10 @@ mod pipewire {
         /// Set once a video format is agreed (`param_changed`), so a first-frame timeout can tell
         /// "format never negotiated" apart from "negotiated but no buffers arrived".
         negotiated: Arc<AtomicBool>,
+        /// True only while the PipeWire stream is in `Streaming` (the source is alive). Goes false on
+        /// `Paused`/`Unconnected`/`Error` — the source vanished (compositor torn down on a session
+        /// switch). Read by [`PortalCapturer::try_latest`] to surface a sustained drop as a loss.
+        streaming: Arc<AtomicBool>,
         /// Present when zero-copy is enabled on NVIDIA: imports a dmabuf → CUDA device buffer.
         importer: Option<crate::zerocopy::EglImporter>,
         /// VAAPI zero-copy: hand the raw dmabuf to the encoder (which imports + GPU-CSCs it) instead
@@ -1056,6 +1110,7 @@ mod pipewire {
         tx: SyncSender<CapturedFrame>,
         active: Arc<AtomicBool>,
         negotiated: Arc<AtomicBool>,
+        streaming: Arc<AtomicBool>,
         zerocopy: bool,
         preferred: Option<(u32, u32, u32)>,
         quit_rx: pw::channel::Receiver<()>,
@@ -1150,6 +1205,7 @@ mod pipewire {
             tx,
             active,
             negotiated,
+            streaming,
             importer,
             vaapi_passthrough,
             nv12: crate::zerocopy::nv12_enabled(),
@@ -1174,8 +1230,17 @@ mod pipewire {
 
         let _listener = stream
             .add_local_listener_with_user_data(data)
-            .state_changed(|_stream, _ud, old, new| {
+            .state_changed(|_stream, ud, old, new| {
                 tracing::info!(?old, ?new, "pipewire stream state");
+                // Track whether the node is actively producing. A live source sits in `Streaming`
+                // (a static desktop just sends no buffers); anything else — `Paused`/`Unconnected`/
+                // `Error` — means the source went away (compositor died, virtual output removed on a
+                // Gaming↔Desktop switch). `try_latest` turns a sustained non-Streaming state into a
+                // capture-loss so the encode loop rebuilds instead of freezing on the last frame.
+                ud.streaming.store(
+                    matches!(new, pw::stream::StreamState::Streaming),
+                    Ordering::Relaxed,
+                );
             })
             .param_changed(|_stream, ud, id, param| {
                 let Some(param) = param else { return };

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

@@ -2010,6 +2010,10 @@ pub struct DuplCapturer {
     /// first, retried (legacy DuplicateOutput can't capture HDR). Set for the secure-desktop DDA leg
     /// when the SudoVDA is in HDR; threaded into every (re)duplication incl. ACCESS_LOST recovery.
     want_hdr: bool,
+    /// Full-chroma 4:4:4 session: deliver packed RGB (`Bgra` SDR / `Rgb10a2` HDR) and SKIP the
+    /// video-engine RGB→YUV (NV12/P010) conversion — NVENC reconstructs 4:4:4 only from a full-chroma
+    /// source, so we hand it the RGB texture and it CSCs to YUV444 at encode (chroma_format_idc=3).
+    chroma_444: bool,
     /// HDR (scRGB FP16) capture state. Set when the duplication surface is `R16G16B16A16_FLOAT`
     /// (the desktop has HDR on). The frame can't be `CopyResource`d into a BGRA target, so the HDR
     /// path copies it into an FP16 SRV texture, composites the cursor, then runs [`HdrConverter`] to
@@ -2087,6 +2091,8 @@ impl DuplCapturer {
         // stage 5) so the capturer never re-derives the encode backend itself.
         gpu: bool,
         want_hdr: bool,
+        // 4:4:4 session → deliver RGB, skip the NV12/P010 video-engine conversion (see the field doc).
+        chroma_444: bool,
     ) -> Result<Self> {
         // SAFETY: runs on the capture thread that will own this `DuplCapturer`. `install_gpu_pref_hook()`
         // and the DPI-context calls take by-value handles / no args and touch only thread/process state;
@@ -2311,6 +2317,7 @@ impl DuplCapturer {
                 gpu_copy: None,
                 last_present: None,
                 want_hdr,
+                chroma_444,
                 hdr_fp16: is_hdr_init,
                 hdr_meta: hdr_meta_init,
                 fp16_src: None,
@@ -3088,7 +3095,10 @@ impl DuplCapturer {
                                                     // Video-engine path: scRGB FP16 → BT.2020 PQ P010 on the VIDEO engine (no 3D shader, and
                                                     // NVENC encodes P010 natively). Fall back to the HdrConverter pixel shader (3D) only if the
                                                     // video processor is unavailable.
-            if let Some(p010) = self.convert_to_yuv(&src, true) {
+            if let Some(p010) = (!self.chroma_444)
+                .then(|| self.convert_to_yuv(&src, true))
+                .flatten()
+            {
                 self.last_present = Some((p010.clone(), PixelFormat::P010));
                 return Ok(CapturedFrame {
                     width: self.width,
@@ -3148,7 +3158,10 @@ impl DuplCapturer {
             // conversion AND NVENC's encode stay OFF the 3D engine — the only way to keep up when a
             // game pins the 3D engine at ~100%. Fall back to handing NVENC the BGRA texture (it then
             // does RGB→YUV internally on the 3D/compute engine).
-            if let Some(nv12) = self.convert_to_yuv(&gpu, false) {
+            if let Some(nv12) = (!self.chroma_444)
+                .then(|| self.convert_to_yuv(&gpu, false))
+                .flatten()
+            {
                 self.last_present = Some((nv12.clone(), PixelFormat::Nv12));
                 return Ok(CapturedFrame {
                     width: self.width,

crates/punktfunk-host/src/config.rs

@@ -7,7 +7,7 @@
 //! **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` and the multi-site `perf`/`compositor`/
+//! plan-named `secure_dda`/`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.
 //!
@@ -63,6 +63,10 @@ pub struct HostConfig {
     pub zerocopy: bool,
     /// `PUNKTFUNK_10BIT` — host policy gate for HEVC Main10 (only honored when the client also advertised 10-bit).
     pub ten_bit: bool,
+    /// `PUNKTFUNK_444` — host policy gate for full-chroma HEVC 4:4:4 (Range Extensions). Honored only
+    /// when the client also advertised 4:4:4, the codec is HEVC, and the GPU/driver supports a 4:4:4
+    /// encode (probed) — otherwise the session stays 4:2:0. Independent of `ten_bit` (chroma vs depth).
+    pub four_four_four: bool,
     /// `PUNKTFUNK_PERF` — per-stage timing instrumentation.
     pub perf: bool,
     /// `PUNKTFUNK_VIDEO_SOURCE` — GameStream video source select (`virtual` / `portal` / unset → synthetic).
@@ -112,6 +116,7 @@ impl HostConfig {
                 .unwrap_or(2),
             zerocopy: flag("PUNKTFUNK_ZEROCOPY"),
             ten_bit: flag("PUNKTFUNK_10BIT"),
+            four_four_four: flag("PUNKTFUNK_444"),
             perf: flag("PUNKTFUNK_PERF"),
             video_source: val("PUNKTFUNK_VIDEO_SOURCE"),
             compositor: val("PUNKTFUNK_COMPOSITOR"),

crates/punktfunk-host/src/encode.rs

@@ -29,6 +29,33 @@ pub enum Codec {
     Av1,
 }
 
+/// Chroma subsampling the encoder emits, negotiated with the client (the `PUNKTFUNK_444` gate + the
+/// client's `VIDEO_CAP_444` + a GPU probe). `Yuv420` is the universal default; `Yuv444` is HEVC-only,
+/// native-protocol-only (GameStream stays 4:2:0), and the host only ever passes it after
+/// [`can_encode_444`] confirmed the active backend supports it.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Default)]
+pub enum ChromaFormat {
+    #[default]
+    Yuv420,
+    Yuv444,
+}
+
+impl ChromaFormat {
+    /// The HEVC `chroma_format_idc` this maps to: `1` (4:2:0) or `3` (4:4:4). Also the wire value
+    /// echoed in [`punktfunk_core::quic::Welcome::chroma_format`].
+    pub fn idc(self) -> u8 {
+        match self {
+            ChromaFormat::Yuv420 => punktfunk_core::quic::CHROMA_IDC_420,
+            ChromaFormat::Yuv444 => punktfunk_core::quic::CHROMA_IDC_444,
+        }
+    }
+
+    /// True for full-chroma 4:4:4.
+    pub fn is_444(self) -> bool {
+        matches!(self, ChromaFormat::Yuv444)
+    }
+}
+
 impl Codec {
     /// The FFmpeg NVENC encoder name (selected by name, not codec id — the latter would
     /// pick the software encoder).
@@ -89,6 +116,13 @@ pub struct EncoderCaps {
     /// When `false`, `set_hdr_meta` is a no-op and no in-band grade reaches the client. Only the
     /// Windows direct-NVENC path attaches it today.
     pub supports_hdr_metadata: bool,
+    /// The opened encoder is actually producing a full-chroma 4:4:4 (`chroma_format_idc = 3`) stream.
+    /// `false` on every 4:2:0 session (the default) and on a backend that declined 4:4:4. Set by the
+    /// NVENC backends (Linux + Windows). The chroma is committed to the wire (`Welcome::chroma_format`)
+    /// from the pre-open probe, so this is a *post-open cross-check*: the session glue logs loudly if
+    /// the encoder's real chroma disagrees with what was negotiated (the in-band SPS is authoritative
+    /// for the decoder either way).
+    pub chroma_444: bool,
 }
 
 /// A hardware encoder. One per session; runs on the encode thread.
@@ -193,8 +227,21 @@ pub fn open_video(
     bitrate_bps: u64,
     cuda: bool,
     bit_depth: u8,
+    chroma: ChromaFormat,
 ) -> Result<Box<dyn Encoder>> {
     validate_dimensions(codec, width, height)?;
+    // 4:4:4 is HEVC-only. The negotiator should never pass `Yuv444` for another codec (it gates on
+    // `codec == H265`), but defend the contract here so a future caller can't silently emit a stream
+    // no decoder expects: a non-HEVC 4:4:4 request degrades to 4:2:0 with a warning.
+    let chroma = if chroma.is_444() && codec != Codec::H265 {
+        tracing::warn!(
+            ?codec,
+            "4:4:4 requested for a non-HEVC codec — encoding 4:2:0"
+        );
+        ChromaFormat::Yuv420
+    } else {
+        chroma
+    };
     #[cfg(target_os = "linux")]
     {
         // Pick the GPU encode backend. NVIDIA → NVENC/CUDA (the original path, unchanged);
@@ -203,8 +250,17 @@ pub fn open_video(
         // its errors crisply instead of silently trying the other).
         let pref = crate::config::config().encoder_pref.as_str();
         let open_vaapi = || -> Result<Box<dyn Encoder>> {
-            vaapi::VaapiEncoder::open(codec, format, width, height, fps, bitrate_bps, bit_depth)
-                .map(|e| Box::new(e) as Box<dyn Encoder>)
+            vaapi::VaapiEncoder::open(
+                codec,
+                format,
+                width,
+                height,
+                fps,
+                bitrate_bps,
+                bit_depth,
+                chroma,
+            )
+            .map(|e| Box::new(e) as Box<dyn Encoder>)
         };
         match pref {
             "nvenc" | "nvidia" | "cuda" => open_nvenc_probed(
@@ -216,6 +272,7 @@ pub fn open_video(
                 bitrate_bps,
                 cuda,
                 bit_depth,
+                chroma,
             ),
             "vaapi" | "amd" | "intel" => open_vaapi(),
             "auto" | "" => {
@@ -231,6 +288,7 @@ pub fn open_video(
                         bitrate_bps,
                         cuda,
                         bit_depth,
+                        chroma,
                     )
                 } else {
                     open_vaapi()
@@ -260,6 +318,7 @@ pub fn open_video(
                         fps,
                         bitrate_bps,
                         bit_depth,
+                        chroma,
                     )
                     .map(|e| Box::new(e) as Box<dyn Encoder>)
                 }
@@ -289,6 +348,7 @@ pub fn open_video(
                         fps,
                         bitrate_bps,
                         bit_depth,
+                        chroma,
                     )
                     .map(|e| Box::new(e) as Box<dyn Encoder>)
                 }
@@ -333,6 +393,7 @@ pub fn open_video(
             bitrate_bps,
             cuda,
             bit_depth,
+            chroma,
         );
         anyhow::bail!("video encode requires Linux or Windows")
     }
@@ -355,6 +416,7 @@ fn open_nvenc_probed(
     bitrate_bps: u64,
     cuda: bool,
     bit_depth: u8,
+    chroma: ChromaFormat,
 ) -> Result<Box<dyn Encoder>> {
     const MIN_PROBE_BPS: u64 = 50_000_000;
     let mut candidates = vec![bitrate_bps];
@@ -369,7 +431,9 @@ fn open_nvenc_probed(
     }
     let mut last: Option<anyhow::Error> = None;
     for (i, &b) in candidates.iter().enumerate() {
-        match linux::NvencEncoder::open(codec, format, width, height, fps, b, cuda, bit_depth) {
+        match linux::NvencEncoder::open(
+            codec, format, width, height, fps, b, cuda, bit_depth, chroma,
+        ) {
             Ok(enc) => {
                 if i > 0 {
                     tracing::warn!(
@@ -446,6 +510,65 @@ pub fn vaapi_codec_support() -> CodecSupport {
     })
 }
 
+/// Whether the active GPU encode backend can actually produce a full-chroma **4:4:4** HEVC stream.
+/// Resolved (and cached, once) *before* the Welcome so the host advertises the chroma it will really
+/// encode — the honest-downgrade channel. 4:4:4 is HEVC-only; the probe opens a tiny encoder on the
+/// active backend (NVENC FREXT is broad on NVIDIA, but VAAPI / AMF / QSV 4:4:4 is hardware-specific,
+/// so it must be probed, never assumed). Non-HEVC codecs are always `false`.
+#[cfg(any(target_os = "linux", target_os = "windows"))]
+pub fn can_encode_444(codec: Codec) -> bool {
+    use std::sync::OnceLock;
+    if codec != Codec::H265 {
+        return false;
+    }
+    static CACHE: OnceLock<bool> = OnceLock::new();
+    *CACHE.get_or_init(|| {
+        let supported = {
+            #[cfg(target_os = "linux")]
+            {
+                // Mirror open_video's backend dispatch: VAAPI (AMD/Intel) vs NVENC (NVIDIA).
+                if linux_zero_copy_is_vaapi() {
+                    vaapi::probe_can_encode_444(codec)
+                } else {
+                    linux::probe_can_encode_444(codec)
+                }
+            }
+            #[cfg(target_os = "windows")]
+            {
+                match windows_resolved_backend() {
+                    WindowsBackend::Nvenc => {
+                        #[cfg(feature = "nvenc")]
+                        {
+                            nvenc::probe_can_encode_444(codec)
+                        }
+                        #[cfg(not(feature = "nvenc"))]
+                        {
+                            false
+                        }
+                    }
+                    WindowsBackend::Amf | WindowsBackend::Qsv => {
+                        #[cfg(feature = "amf-qsv")]
+                        {
+                            let vendor = match windows_resolved_backend() {
+                                WindowsBackend::Qsv => ffmpeg_win::WinVendor::Qsv,
+                                _ => ffmpeg_win::WinVendor::Amf,
+                            };
+                            ffmpeg_win::probe_can_encode_444(vendor, codec)
+                        }
+                        #[cfg(not(feature = "amf-qsv"))]
+                        {
+                            false
+                        }
+                    }
+                    WindowsBackend::Software => false,
+                }
+            }
+        };
+        tracing::info!(supported, "HEVC 4:4:4 encode capability probed");
+        supported
+    })
+}
+
 // ---------------------------------------------------------------------------------------------
 // Windows backend selection (the analogue of the Linux nvidia_present / linux_zero_copy_is_vaapi
 // logic). NVIDIA → NVENC, AMD → AMF, Intel → QSV; `auto` (default) reads the DXGI adapter vendor.

crates/punktfunk-host/src/encode/linux/mod.rs

@@ -11,7 +11,7 @@
 // Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
 #![deny(clippy::undocumented_unsafe_blocks)]
 
-use super::{Codec, EncodedFrame, Encoder};
+use super::{ChromaFormat, Codec, EncodedFrame, Encoder};
 use crate::capture::{CapturedFrame, FramePayload, PixelFormat};
 use anyhow::{anyhow, bail, Context, Result};
 use ffmpeg::format::Pixel;
@@ -19,9 +19,33 @@ use ffmpeg::util::frame::Video as VideoFrame;
 use ffmpeg::{codec, encoder, Dictionary, Packet, Rational};
 use ffmpeg_next as ffmpeg;
 use std::os::raw::c_int;
+use std::ptr;
 
 use ffmpeg::ffi; // = ffmpeg_sys_next
 
+/// swscale: nearest-neighbour scaler flag (`SWS_POINT`). We never rescale (src dims == dst dims), so
+/// the resampler choice only governs the colour-conversion path; POINT is the cheapest.
+const SWS_POINT: c_int = 0x10;
+/// swscale colorspace id for ITU-R BT.709 (`SWS_CS_ITU709`) — the CSC coefficients for our RGB→YUV.
+const SWS_CS_ITU709: c_int = 1;
+
+/// The swscale *source* pixel format for a captured packed RGB/BGR layout (the real byte order, not
+/// the NVENC-padded `*0` form). Used by the 4:4:4 RGB→YUV444P conversion path. Mirrors the VAAPI
+/// CPU-input mapping; YUV/10-bit inputs can't feed this path (the 4:4:4 session forces packed RGB).
+fn sws_src_pixel(format: PixelFormat) -> Result<Pixel> {
+    Ok(match format {
+        PixelFormat::Bgrx => Pixel::BGRZ, // bgr0
+        PixelFormat::Rgbx => Pixel::RGBZ, // rgb0
+        PixelFormat::Bgra => Pixel::BGRA,
+        PixelFormat::Rgba => Pixel::RGBA,
+        PixelFormat::Rgb => Pixel::RGB24,
+        PixelFormat::Bgr => Pixel::BGR24,
+        PixelFormat::Nv12 | PixelFormat::P010 | PixelFormat::Rgb10a2 => {
+            bail!("NVENC 4:4:4 CPU-input path supports packed RGB/BGR only; got {format:?}")
+        }
+    })
+}
+
 /// `AVCUDADeviceContext` (libavutil/hwcontext_cuda.h) — not in the ffmpeg-sys bindings (the
 /// crate doesn't allowlist that header), so mirror its stable 3-pointer layout. We set the
 /// first field to *our* `CUcontext` so NVENC shares the context the EGL importer maps into.
@@ -131,6 +155,10 @@ pub struct NvencEncoder {
     frame: Option<VideoFrame>,
     /// Zero-copy path: CUDA hwdevice/hwframes contexts (the encoder takes `AV_PIX_FMT_CUDA`).
     cuda: Option<CudaHw>,
+    /// 4:4:4 path only: swscale context converting the captured packed RGB/BGR → planar YUV444P
+    /// (BT.709 limited) into [`Self::frame`], because `hevc_nvenc` only emits 4:4:4 from a YUV444
+    /// *input* (RGB-in is always 4:2:0). `None` on the ordinary 4:2:0 RGB path. Freed in `Drop`.
+    sws_444: Option<*mut ffi::SwsContext>,
     src_format: PixelFormat,
     expand: bool,
     width: u32,
@@ -142,10 +170,12 @@ pub struct NvencEncoder {
     force_kf: bool,
 }
 
-// `CudaHw` holds raw `AVBufferRef`s; the encoder lives on a single thread. The CPU encoder is
-// already `Send` via ffmpeg-next; assert it for the CUDA fields too.
+// `CudaHw` holds raw `AVBufferRef`s and `sws_444` a raw `SwsContext`; the encoder lives on a single
+// thread. The CPU encoder is already `Send` via ffmpeg-next; assert it for the raw fields too.
 // SAFETY: `NvencEncoder` owns an ffmpeg-next `Encoder`/`VideoFrame` (already `Send`) plus a `CudaHw`
-// holding raw `AVBufferRef`s, which are not `Send` by default. The encoder is owned and driven by
+// holding raw `AVBufferRef`s and an optional raw `SwsContext`, none of which are `Send` by default.
+// The `SwsContext` is a self-contained swscale state object with no thread affinity, touched only
+// through `&mut self` on the one encode thread. The encoder is owned and driven by
 // exactly ONE thread — the per-session encode thread it is moved to — and is only touched through
 // `&mut self` methods, so it is never aliased or accessed concurrently. The wrapped libav contexts
 // (and the shared `CUcontext` the `CudaHw` references) have no thread affinity, so transferring
@@ -164,6 +194,7 @@ impl NvencEncoder {
         bitrate_bps: u64,
         cuda: bool,
         bit_depth: u8,
+        chroma: ChromaFormat,
     ) -> Result<Self> {
         // TODO(hdr): Linux 10-bit parity. Unlike the Windows raw-SDK path (which upconverts 8-bit
         // ARGB → Main10 via pixelBitDepthMinus8), libavcodec hevc_nvenc needs a 10-bit input pixel
@@ -175,6 +206,18 @@ impl NvencEncoder {
                 "Linux NVENC 10-bit not yet wired — encoding 8-bit"
             );
         }
+        // Full-chroma 4:4:4 (HEVC Range Extensions). `hevc_nvenc` only emits 4:4:4 from a YUV444
+        // *input* frame — feeding RGB always subsamples to 4:2:0 regardless of profile (verified on
+        // the RTX 5070 Ti). So a 4:4:4 session swscales the captured RGB → YUV444P (BT.709 limited)
+        // and feeds that with `profile=rext`. The negotiator gates this to HEVC + the single-process
+        // CPU-capture topology, so `cuda` must be false here; defend the contract.
+        let want_444 = chroma.is_444() && codec == Codec::H265;
+        if want_444 && cuda {
+            bail!(
+                "NVENC 4:4:4 needs CPU RGB frames (the session forces non-zero-copy capture for \
+                 4:4:4); got a CUDA frame — capture/encoder negotiation mismatch"
+            );
+        }
         ffmpeg::init().context("ffmpeg init")?;
         if std::env::var_os("PUNKTFUNK_FFMPEG_DEBUG").is_some() {
             // SAFETY: `av_log_set_level` sets libav's global integer log level; `48` (= AV_LOG_DEBUG)
@@ -185,7 +228,14 @@ impl NvencEncoder {
         let name = codec.nvenc_name();
         let av_codec = encoder::find_by_name(name)
             .ok_or_else(|| anyhow!("{name} not built into libavcodec"))?;
-        let (nvenc_pixel, expand) = nvenc_input(format);
+        let (rgb_pixel, rgb_expand) = nvenc_input(format);
+        // 4:4:4 feeds NVENC a planar YUV444P frame we produce by swscale; the ordinary path feeds the
+        // captured RGB straight in and lets NVENC's internal CSC subsample to 4:2:0.
+        let (nvenc_pixel, expand) = if want_444 {
+            (Pixel::YUV444P, false)
+        } else {
+            (rgb_pixel, rgb_expand)
+        };
 
         let mut video = codec::context::Context::new_with_codec(av_codec)
             .encoder()
@@ -234,12 +284,12 @@ impl NvencEncoder {
             (*video.as_mut_ptr()).gop_size = -1;
         }
 
-        // NV12 path: we did the RGB→YUV conversion ourselves as BT.709 *limited* range, so signal
-        // that in the bitstream VUI (colorspace/range/primaries/transfer) — otherwise the client
-        // decoder assumes a default and the picture comes out washed-out / wrong-contrast. The
-        // RGB-input paths leave these unset (NVENC's internal CSC writes its own VUI). Matches the
-        // Windows NV12 path's BT.709 limited-range signalling.
-        if matches!(format, PixelFormat::Nv12) {
+        // NV12 / 4:4:4 paths: we do the RGB→YUV conversion ourselves as BT.709 *limited* range
+        // (swscale), so signal that in the bitstream VUI (colorspace/range/primaries/transfer) —
+        // otherwise the client decoder assumes a default and the picture comes out washed-out /
+        // wrong-contrast. The RGB-input 4:2:0 path leaves these unset (NVENC's internal CSC writes
+        // its own VUI). Matches the Windows NV12 path's BT.709 limited-range signalling.
+        if matches!(format, PixelFormat::Nv12) || want_444 {
             // SAFETY: same `video` builder — `raw = video.as_mut_ptr()` is the non-null, properly-
             // aligned, sole-owned, not-yet-opened `AVCodecContext`. We set its four VUI colour enum
             // fields to valid `AVColorSpace`/`AVColorRange`/`AVColorPrimaries`/`AVColorTransfer-
@@ -280,6 +330,45 @@ impl NvencEncoder {
             None
         };
 
+        // 4:4:4: build the RGB→YUV444P swscale (BT.709 limited, no rescale). Mirrors the VAAPI CPU
+        // path's RGB→NV12 scaler, but the dst is full-chroma planar 4:4:4.
+        let sws_444 = if want_444 {
+            let src_av = pixel_to_av(sws_src_pixel(format)?);
+            // SAFETY: `sws_getContext` allocates a swscale context for the given src/dst dims + pixel
+            // formats. Both dims are the encoder's positive `width`/`height` as `c_int`; `src_av` is a
+            // valid `AVPixelFormat` (from the `sws_src_pixel`-validated, packed-RGB-only source), the
+            // dst is YUV444P. The trailing filter/param pointers are null = "use defaults" (documented
+            // as accepted). No Rust memory is borrowed; the returned pointer is null-checked below.
+            let sws = unsafe {
+                ffi::sws_getContext(
+                    width as c_int,
+                    height as c_int,
+                    src_av,
+                    width as c_int,
+                    height as c_int,
+                    ffi::AVPixelFormat::AV_PIX_FMT_YUV444P,
+                    SWS_POINT,
+                    ptr::null_mut(),
+                    ptr::null_mut(),
+                    ptr::null(),
+                )
+            };
+            if sws.is_null() {
+                bail!("sws_getContext(RGB→YUV444P) failed");
+            }
+            // SAFETY: `sws` is the non-null context from the call above (null-checked). The ITU-709
+            // coefficient table from `sws_getCoefficients` is a process-lifetime libswscale static,
+            // reused for src+dst matrices; `sws_setColorspaceDetails` only reads it and writes scalar
+            // CSC settings into `sws` (limited-range dst: dstRange = 0). No Rust memory is passed.
+            unsafe {
+                let cs709 = ffi::sws_getCoefficients(SWS_CS_ITU709);
+                ffi::sws_setColorspaceDetails(sws, cs709, 1, cs709, 0, 0, 1 << 16, 1 << 16);
+            }
+            Some(sws)
+        } else {
+            None
+        };
+
         // Low-latency NVENC tuning (plan §7 / linux-setup doc).
         let mut opts = Dictionary::new();
         opts.set("preset", "p1"); // fastest
@@ -288,6 +377,12 @@ impl NvencEncoder {
         opts.set("bf", "0");
         opts.set("delay", "0");
         opts.set("forced-idr", "1"); // RFI/request_keyframe → real IDR under the infinite GOP
+        if want_444 {
+            // HEVC Range Extensions — the profile that carries chroma_format_idc=3. With a YUV444P
+            // input `hevc_nvenc` auto-selects it, but pin it explicitly so the chroma is never silently
+            // dropped on a future libavcodec.
+            opts.set("profile", "rext");
+        }
 
         // Split-frame encode across both NVENC engines (GB203 has 2) when the pixel rate exceeds
         // a single engine's HEVC capacity (~1 Gpix/s); e.g. 5120x1440@240 = 1.77 Gpix/s needs it,
@@ -321,6 +416,7 @@ impl NvencEncoder {
             enc,
             frame,
             cuda: cuda_hw,
+            sws_444,
             src_format: format,
             expand,
             width,
@@ -333,6 +429,15 @@ impl NvencEncoder {
 }
 
 impl Encoder for NvencEncoder {
+    fn caps(&self) -> super::EncoderCaps {
+        super::EncoderCaps {
+            // 4:4:4 iff this session opened the RGB→YUV444P swscale path (FREXT). RFI/HDR-SEI stay
+            // unsupported on libavcodec NVENC (the trait defaults).
+            chroma_444: self.sws_444.is_some(),
+            ..super::EncoderCaps::default()
+        }
+    }
+
     fn submit(&mut self, captured: &CapturedFrame) -> Result<()> {
         anyhow::ensure!(
             captured.width == self.width && captured.height == self.height,
@@ -411,6 +516,47 @@ impl NvencEncoder {
             bytes.len(),
             src_row * h
         );
+        // 4:4:4: swscale the packed RGB straight into the planar YUV444P input frame (BT.709 limited),
+        // then send it — no byte-expand. The 4:2:0 RGB path (below) feeds NVENC packed RGB directly.
+        if let Some(sws) = self.sws_444 {
+            let frame = self
+                .frame
+                .as_mut()
+                .context("CPU frame missing (encoder opened in CUDA mode)")?;
+            // SAFETY: `format == self.src_format` and `bytes.len() >= src_row * h` (the `ensure!`s
+            // above), so `sws_scale` reads `h` rows of `src_row` bytes from `src_data[0] = bytes`
+            // (packed RGB is single-plane; the other src planes are null/0) — all in bounds. `sws` is
+            // the non-null context built in `open`. The dst is `frame`'s underlying `AVFrame`: its
+            // `data`/`linesize` in-struct arrays were sized for YUV444P by `VideoFrame::new`, and the
+            // 3 planes are each `width`×`height`. All pointers are live locals for this synchronous
+            // call; the encoder runs only on this thread (`unsafe impl Send`), so no aliasing/race.
+            unsafe {
+                let dst_av = frame.as_mut_ptr();
+                let src_data: [*const u8; 4] =
+                    [bytes.as_ptr(), ptr::null(), ptr::null(), ptr::null()];
+                let src_stride: [c_int; 4] = [src_row as c_int, 0, 0, 0];
+                let r = ffi::sws_scale(
+                    sws,
+                    src_data.as_ptr(),
+                    src_stride.as_ptr(),
+                    0,
+                    h as c_int,
+                    (*dst_av).data.as_ptr(),
+                    (*dst_av).linesize.as_ptr(),
+                );
+                if r < 0 {
+                    bail!("sws_scale(RGB→YUV444P) failed ({r})");
+                }
+            }
+            frame.set_pts(Some(pts));
+            frame.set_kind(if idr {
+                ffmpeg::picture::Type::I
+            } else {
+                ffmpeg::picture::Type::None
+            });
+            self.enc.send_frame(frame).context("send_frame(444)")?;
+            return Ok(());
+        }
         let frame = self
             .frame
             .as_mut()
@@ -526,3 +672,51 @@ impl NvencEncoder {
         Ok(())
     }
 }
+
+impl Drop for NvencEncoder {
+    fn drop(&mut self) {
+        if let Some(sws) = self.sws_444.take() {
+            // SAFETY: `sws` is the non-null `SwsContext` allocated by `sws_getContext` in `open` and
+            // owned exclusively by this encoder (taken out of the field so it can't be freed twice).
+            // `sws_freeContext` frees it; nothing else references it after this single-threaded drop.
+            unsafe { ffi::sws_freeContext(sws) };
+        }
+    }
+}
+
+/// Probe whether this NVIDIA GPU + driver + libavcodec can actually encode HEVC **4:4:4** (Range
+/// Extensions). Opens a tiny real `hevc_nvenc` 4:4:4 session — the exact path [`NvencEncoder::open`]
+/// takes for a live 4:4:4 stream — and reports whether it succeeded. HEVC-only; the result is cached
+/// by the caller ([`crate::encode::can_encode_444`]). A GPU/driver/ffmpeg without RExt 4:4:4 fails
+/// the open here, so the host resolves the session to 4:2:0 before the Welcome (honest downgrade).
+pub fn probe_can_encode_444(codec: Codec) -> bool {
+    if codec != Codec::H265 {
+        return false;
+    }
+    if ffmpeg::init().is_err() {
+        return false;
+    }
+    // Quiet ffmpeg's open error on a GPU that lacks 4:4:4 — the probe failing is an expected outcome.
+    // SAFETY: libav initialized above; `av_log_{get,set}_level` only read/write the global int level
+    // (no pointer args) and are always sound post-init.
+    let prev = unsafe {
+        let p = ffi::av_log_get_level();
+        ffi::av_log_set_level(ffi::AV_LOG_FATAL);
+        p
+    };
+    let ok = NvencEncoder::open(
+        codec,
+        PixelFormat::Bgra,
+        640,
+        480,
+        30,
+        2_000_000,
+        false, // CPU input (the 4:4:4 path never uses CUDA)
+        8,
+        ChromaFormat::Yuv444,
+    )
+    .is_ok();
+    // SAFETY: restore the saved global log level (scalar arg, no pointers).
+    unsafe { ffi::av_log_set_level(prev) };
+    ok
+}

crates/punktfunk-host/src/encode/linux/vaapi.rs

@@ -160,6 +160,18 @@ pub fn probe_can_encode(codec: Codec) -> bool {
     }
 }
 
+/// Whether the active VAAPI GPU can encode HEVC **4:4:4** (Range Extensions). **Deferred in v1 —
+/// always `false`.** VAAPI HEVC 4:4:4 encode is narrow and vendor-specific (the lab's AMD Phoenix1 /
+/// RDNA3 exposes only `VAProfileHEVCMain`/`Main10` `EncSlice`, no `Main444`), and there is no
+/// validated hardware to build + verify the 4:4:4 surface/profile path against. Returning `false`
+/// keeps the negotiation honest: a VAAPI host resolves every session to 4:2:0 before the Welcome, so
+/// the client never builds a 4:4:4 decoder it would only get 4:2:0 frames for. (Follow-up: implement
+/// + validate on an Intel Arc / RDNA4-class box that advertises a HEVC 4:4:4 encode entrypoint.)
+pub fn probe_can_encode_444(_codec: Codec) -> bool {
+    tracing::info!("VAAPI HEVC 4:4:4 encode is not implemented yet — declining (encoding 4:2:0)");
+    false
+}
+
 /// Drain the encoder for one packet (shared poll logic).
 fn poll_encoder(enc: &mut encoder::video::Encoder, fps: u32) -> Result<Option<EncodedFrame>> {
     let mut pkt = Packet::empty();
@@ -848,6 +860,7 @@ pub struct VaapiEncoder {
 unsafe impl Send for VaapiEncoder {}
 
 impl VaapiEncoder {
+    #[allow(clippy::too_many_arguments)]
     pub fn open(
         codec: Codec,
         format: PixelFormat,
@@ -856,10 +869,18 @@ impl VaapiEncoder {
         fps: u32,
         bitrate_bps: u64,
         bit_depth: u8,
+        chroma: super::ChromaFormat,
     ) -> Result<Self> {
         if bit_depth != 8 {
             tracing::warn!(bit_depth, "VAAPI 10-bit not yet wired — encoding 8-bit");
         }
+        // VAAPI 4:4:4 is deferred (see `probe_can_encode_444`): no validated AMD/Intel hardware in the
+        // lab exposes a HEVC 4:4:4 encode entrypoint, and the probe returns false so the host never
+        // negotiates 4:4:4 for a VAAPI session. If a request slips through, fall back to 4:2:0 rather
+        // than emit an unverified stream — the host signalled 4:2:0 in the Welcome anyway.
+        if chroma.is_444() {
+            tracing::warn!("VAAPI 4:4:4 encode not implemented — encoding 4:2:0");
+        }
         ffmpeg::init().context("ffmpeg init")?;
         if std::env::var_os("PUNKTFUNK_FFMPEG_DEBUG").is_some() {
             // SAFETY: `av_log_set_level` sets libav's global integer log level; `48` (= AV_LOG_DEBUG)

crates/punktfunk-host/src/encode/windows/ffmpeg_win.rs

@@ -31,7 +31,7 @@
 // Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
 #![deny(clippy::undocumented_unsafe_blocks)]
 
-use super::{Codec, EncodedFrame, Encoder};
+use super::{ChromaFormat, Codec, EncodedFrame, Encoder};
 use crate::capture::{dxgi::D3d11Frame, CapturedFrame, FramePayload, PixelFormat};
 use anyhow::{anyhow, bail, Context, Result};
 use ffmpeg::format::Pixel;
@@ -241,6 +241,18 @@ unsafe fn open_win_encoder(
 /// driver/runtime rejects codecs the video engine can't do (AV1 on pre-RDNA3 AMD / pre-Arc Intel,
 /// or HEVC on a very old part). Used to build the GameStream codec advertisement so a client never
 /// negotiates a codec the encoder can't open. Torn down immediately.
+/// Whether the active AMD (AMF) / Intel (QSV) GPU can encode HEVC **4:4:4**. **Deferred in v1 —
+/// always `false`.** AMF/QSV HEVC 4:4:4 encode is narrow (AMD RDNA3+, Intel Arc/Xe2+) and the
+/// libavcodec profile/pixel-format incantation is vendor- and driver-specific — a wrong profile
+/// `avcodec_open2` *silently* falls back to 4:2:0, so a positive probe would need a verify-by-frame,
+/// and there is no AMD/Intel Windows box in the lab to build + validate that against. Returning
+/// `false` keeps the negotiation honest: an AMF/QSV host resolves every session to 4:2:0 before the
+/// Welcome. (Follow-up: implement + validate on an RDNA3+/Arc Windows box.)
+pub fn probe_can_encode_444(_vendor: WinVendor, _codec: Codec) -> bool {
+    tracing::info!("AMF/QSV HEVC 4:4:4 encode is not implemented yet — declining (encoding 4:2:0)");
+    false
+}
+
 pub fn probe_can_encode(vendor: WinVendor, codec: Codec) -> bool {
     if ffmpeg::init().is_err() {
         return false;
@@ -1096,6 +1108,7 @@ pub struct FfmpegWinEncoder {
 unsafe impl Send for FfmpegWinEncoder {}
 
 impl FfmpegWinEncoder {
+    #[allow(clippy::too_many_arguments)]
     #[allow(clippy::too_many_arguments)]
     pub fn open(
         vendor: WinVendor,
@@ -1106,7 +1119,15 @@ impl FfmpegWinEncoder {
         fps: u32,
         bitrate_bps: u64,
         bit_depth: u8,
+        chroma: ChromaFormat,
     ) -> Result<Self> {
+        // AMF/QSV 4:4:4 is deferred (see `probe_can_encode_444`): no validated AMD/Intel Windows
+        // hardware in the lab, and the AMF/QSV HEVC 4:4:4 profile/format incantations are vendor- and
+        // driver-specific (a wrong profile silently encodes 4:2:0). The probe returns false so the host
+        // never negotiates 4:4:4 for an AMF/QSV session; if a request slips through, fall back to 4:2:0.
+        if chroma.is_444() {
+            tracing::warn!("AMF/QSV 4:4:4 encode not implemented — encoding 4:2:0");
+        }
         ffmpeg::init().context("ffmpeg init")?;
         if std::env::var_os("PUNKTFUNK_FFMPEG_DEBUG").is_some() {
             // SAFETY: `ffmpeg::init()` ran on the line above, so libav is initialised; `av_log_set_level`

crates/punktfunk-host/src/encode/windows/nvenc.rs

@@ -16,7 +16,7 @@
 // Every `unsafe` block / impl in this file carries a `// SAFETY:` proof; enforce it.
 #![deny(clippy::undocumented_unsafe_blocks)]
 
-use super::{Codec, EncodedFrame, Encoder, EncoderCaps};
+use super::{ChromaFormat, Codec, EncodedFrame, Encoder, EncoderCaps};
 use crate::capture::{CapturedFrame, FramePayload, PixelFormat};
 use anyhow::{anyhow, bail, Context, Result};
 use std::collections::{HashMap, VecDeque};
@@ -57,6 +57,15 @@ pub struct NvencD3d11Encoder {
     buffer_fmt: nv::NV_ENC_BUFFER_FORMAT,
     /// Encoded bit depth (8 or 10). 10 → HEVC Main10 (NVENC upconverts the 8-bit ARGB input).
     bit_depth: u8,
+    /// Full-chroma 4:4:4 (HEVC Range Extensions, `chroma_format_idc = 3`) requested for this session.
+    /// NVENC ingests the RGB (ARGB/ABGR10) input and CSCs it to YUV444 internally — the `FREXT` profile
+    /// + `chromaFormatIDC = 3` in the encode config carry the chroma. Gated on the GPU's
+    /// `NV_ENC_CAPS_SUPPORT_YUV444_ENCODE` (cleared in `query_caps` on a card that lacks it) and on an
+    /// RGB input format (NV12/P010 capture can't reconstruct 4:4:4). HEVC-only.
+    chroma_444: bool,
+    /// `NV_ENC_CAPS_SUPPORT_YUV444_ENCODE` from the caps probe — whether this GPU can 4:4:4 encode at
+    /// all. `chroma_444` is forced off when this is false (graceful downgrade to 4:2:0).
+    yuv444_supported: bool,
     /// HDR: the capturer is delivering BT.2020 PQ 10-bit (`PixelFormat::Rgb10a2`) frames. Sets the
     /// `ABGR10` input format + the BT.2020/PQ colour VUI. Derived per-frame from the capture format
     /// (HDR can toggle mid-session); a change re-inits the session.
@@ -103,6 +112,7 @@ pub struct NvencD3d11Encoder {
 unsafe impl Send for NvencD3d11Encoder {}
 
 impl NvencD3d11Encoder {
+    #[allow(clippy::too_many_arguments)]
     pub fn open(
         codec: Codec,
         _format: PixelFormat,
@@ -111,6 +121,7 @@ impl NvencD3d11Encoder {
         fps: u32,
         bitrate_bps: u64,
         bit_depth: u8,
+        chroma: ChromaFormat,
     ) -> Result<Self> {
         Ok(Self {
             encoder: ptr::null_mut(),
@@ -122,6 +133,9 @@ impl NvencD3d11Encoder {
             bitrate_bps,
             buffer_fmt: nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ARGB,
             bit_depth,
+            // 4:4:4 is HEVC-only; the GPU-support gate is applied in `query_caps`.
+            chroma_444: chroma.is_444() && codec == Codec::H265,
+            yuv444_supported: false,
             hdr: false,
             hdr_meta: None,
             regs: HashMap::new(),
@@ -209,6 +223,7 @@ impl NvencD3d11Encoder {
         let wmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_WIDTH_MAX);
         let hmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_HEIGHT_MAX);
         let ten_bit = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
+        let yuv444 = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_YUV444_ENCODE);
         let rfi = self.get_cap(
             enc,
             nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_REF_PIC_INVALIDATION,
@@ -235,6 +250,13 @@ impl NvencD3d11Encoder {
             self.bit_depth = 8;
             self.hdr = false;
         }
+        // Same for 4:4:4: a card without YUV444 encode falls back to 4:2:0. (The host already probed
+        // this via `probe_can_encode_444` before the Welcome, so this is a belt-and-braces guard.)
+        self.yuv444_supported = yuv444 != 0;
+        if self.chroma_444 && !self.yuv444_supported {
+            tracing::warn!("NVENC: this GPU can't 4:4:4 encode — falling back to 4:2:0");
+            self.chroma_444 = false;
+        }
         self.rfi_supported = rfi != 0;
         self.custom_vbv = custom_vbv != 0;
         tracing::info!(
@@ -313,9 +335,31 @@ impl NvencD3d11Encoder {
         cfg.encodeCodecConfig.hevcConfig.tier = 1;
         cfg.encodeCodecConfig.hevcConfig.level = 0;
 
-        // 10-bit HEVC Main10 (HDR foundation): NVENC upconverts the 8-bit input; 8-bit leaves the
-        // preset default (Main) untouched.
-        if self.bit_depth == 10 {
+        // Chroma + bit depth. Full-chroma 4:4:4 (HEVC Range Extensions) takes precedence and composes
+        // with 10-bit (Main 4:4:4 10): NVENC ingests the RGB input (ARGB / ABGR10) and CSCs it to
+        // YUV444 internally when `chromaFormatIDC = 3` under the FREXT profile. Only valid on an RGB
+        // input — a subsampled NV12/P010 source can't reconstruct full chroma (so the capturer is
+        // forced to RGB for a 4:4:4 session, and we guard on the input format here too).
+        //
+        // ON-GLASS TODO (RTX box): confirm ARGB + chromaFormatIDC=3 + FREXT yields a *true* 4:4:4
+        // stream. NVENC's RGB→YUV CSC is documented to honor chromaFormatIDC (unlike libavcodec's
+        // wrapper, which always subsamples RGB to 4:2:0 — hence the Linux path feeds planar YUV444
+        // instead). If on-glass shows 4:2:0, the follow-up is a BGRA→AYUV shader feeding the native
+        // `NV_ENC_BUFFER_FORMAT_AYUV` 4:4:4 input format.
+        let rgb_input = matches!(
+            self.buffer_fmt,
+            nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ARGB
+                | nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ABGR10
+        );
+        if self.chroma_444 && rgb_input {
+            cfg.profileGUID = nv::NV_ENC_HEVC_PROFILE_FREXT_GUID;
+            cfg.encodeCodecConfig.hevcConfig.set_chromaFormatIDC(3);
+            if self.bit_depth == 10 {
+                cfg.encodeCodecConfig.hevcConfig.set_pixelBitDepthMinus8(2); // Main 4:4:4 10
+            }
+        } else if self.bit_depth == 10 {
+            // 10-bit HEVC Main10 (HDR foundation): NVENC upconverts the 8-bit input; 8-bit leaves the
+            // preset default (Main) untouched.
             cfg.profileGUID = nv::NV_ENC_HEVC_PROFILE_MAIN10_GUID;
             cfg.encodeCodecConfig.hevcConfig.set_pixelBitDepthMinus8(2); // 10 - 8
         }
@@ -787,6 +831,9 @@ impl Encoder for NvencD3d11Encoder {
         EncoderCaps {
             supports_rfi: self.rfi_supported,
             supports_hdr_metadata: self.hdr,
+            // Reflects what the session actually configured (cleared in `query_caps` if the GPU lacks
+            // YUV444 encode), so the glue can confirm 4:4:4 vs the negotiated request.
+            chroma_444: self.chroma_444,
         }
     }
 
@@ -904,3 +951,69 @@ impl Drop for NvencD3d11Encoder {
         unsafe { self.teardown() };
     }
 }
+
+/// Probe whether the active NVIDIA GPU can encode HEVC **4:4:4** (`NV_ENC_CAPS_SUPPORT_YUV444_ENCODE`).
+/// Creates a throwaway hardware D3D11 device + NVENC session, queries the cap, and tears down. HEVC-only;
+/// the result is cached by the caller ([`crate::encode::can_encode_444`]) and read *before* the Welcome
+/// so the host advertises the chroma it can really encode (honest downgrade to 4:2:0 on a card without it).
+pub fn probe_can_encode_444(codec: Codec) -> bool {
+    use windows::Win32::Foundation::HMODULE;
+    use windows::Win32::Graphics::Direct3D::{D3D_DRIVER_TYPE_HARDWARE, D3D_FEATURE_LEVEL_11_0};
+    use windows::Win32::Graphics::Direct3D11::{
+        D3D11CreateDevice, D3D11_CREATE_DEVICE_BGRA_SUPPORT, D3D11_SDK_VERSION,
+    };
+    if codec != Codec::H265 {
+        return false;
+    }
+    // SAFETY: a self-contained probe owning every handle it creates. `D3D11CreateDevice` (HARDWARE
+    // driver, NULL adapter) fills `device` or returns Err (→ false). `open_encode_session_ex` opens an
+    // NVENC session against that device's raw pointer (valid while `device` is held) or errors (→ false,
+    // tearing nothing down). `get_encode_caps` reads one scalar cap into `val` via the loaded API table.
+    // `destroy_encoder` frees the session exactly once; `device`/its context drop with the COM wrappers.
+    // No handle escapes this call and nothing runs concurrently.
+    unsafe {
+        let mut device: Option<ID3D11Device> = None;
+        if D3D11CreateDevice(
+            None,
+            D3D_DRIVER_TYPE_HARDWARE,
+            HMODULE::default(),
+            D3D11_CREATE_DEVICE_BGRA_SUPPORT,
+            Some(&[D3D_FEATURE_LEVEL_11_0]),
+            D3D11_SDK_VERSION,
+            Some(&mut device),
+            None,
+            None,
+        )
+        .is_err()
+        {
+            return false;
+        }
+        let Some(device) = device else { return false };
+        let mut params = nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS {
+            version: nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER,
+            deviceType: nv::NV_ENC_DEVICE_TYPE::NV_ENC_DEVICE_TYPE_DIRECTX,
+            device: device.as_raw(),
+            apiVersion: nv::NVENCAPI_VERSION,
+            ..Default::default()
+        };
+        let mut enc: *mut c_void = ptr::null_mut();
+        if (API.open_encode_session_ex)(&mut params, &mut enc)
+            .result_without_string()
+            .is_err()
+        {
+            return false;
+        }
+        let mut param = nv::NV_ENC_CAPS_PARAM {
+            version: nv::NV_ENC_CAPS_PARAM_VER,
+            capsToQuery: nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_YUV444_ENCODE,
+            reserved: [0; 62],
+        };
+        let mut val: i32 = 0;
+        let ok = (API.get_encode_caps)(enc, nv::NV_ENC_CODEC_HEVC_GUID, &mut param, &mut val)
+            .result_without_string()
+            .is_ok()
+            && val != 0;
+        let _ = (API.destroy_encoder)(enc);
+        ok
+    }
+}

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

@@ -41,8 +41,6 @@ type Aes128CbcEnc = cbc::Encryptor<aes::Aes128>;
 /// `RTP_PAYLOAD_TYPE_FEC  127`).
 const AUDIO_PACKET_TYPE: u8 = 97;
 const AUDIO_FEC_PACKET_TYPE: u8 = 127;
-/// Stereo Opus bitrate (unchanged from the live-validated stereo path).
-const OPUS_BITRATE: i32 = 128_000;
 
 /// Audio FEC geometry (moonlight-common-c `RtpAudioQueue.h`: `RTPA_DATA_SHARDS 4`,
 /// `RTPA_FEC_SHARDS 2`). Blocks are aligned: the client synthesizes the block base as
@@ -82,67 +80,20 @@ impl Default for AudioParams {
     }
 }
 
-/// One Opus (multi)stream layout. Channel order is the GameStream/Moonlight order
-/// FL FR FC LFE RL RR [SL SR]; `mapping` is the libopus multistream mapping we *encode*
-/// with — identical to Sunshine's `audio.cpp stream_configs` (verified verbatim 2026-06-10):
-/// identity mapping, so normal quality couples (FL,FR) and (FC,LFE) [+ (RL,RR) on 7.1] with
-/// the remaining channels as mono streams; high quality is one mono stream per channel.
-/// Bitrates are Sunshine's per-config values (stereo keeps punktfunk's existing 128 kbps).
-pub struct OpusLayout {
-    pub channels: u8,
-    pub streams: u8,
-    pub coupled: u8,
-    pub mapping: &'static [u8],
-    pub bitrate: i32,
-}
-
-pub const LAYOUT_STEREO: OpusLayout = OpusLayout {
-    channels: 2,
-    streams: 1,
-    coupled: 1,
-    mapping: &[0, 1],
-    bitrate: OPUS_BITRATE,
-};
-pub const LAYOUT_51: OpusLayout = OpusLayout {
-    channels: 6,
-    streams: 4,
-    coupled: 2,
-    mapping: &[0, 1, 2, 3, 4, 5],
-    bitrate: 256_000,
-};
-pub const LAYOUT_51_HQ: OpusLayout = OpusLayout {
-    channels: 6,
-    streams: 6,
-    coupled: 0,
-    mapping: &[0, 1, 2, 3, 4, 5],
-    bitrate: 1_536_000,
-};
-pub const LAYOUT_71: OpusLayout = OpusLayout {
-    channels: 8,
-    streams: 5,
-    coupled: 3,
-    mapping: &[0, 1, 2, 3, 4, 5, 6, 7],
-    bitrate: 450_000,
-};
-pub const LAYOUT_71_HQ: OpusLayout = OpusLayout {
-    channels: 8,
-    streams: 8,
-    coupled: 0,
-    mapping: &[0, 1, 2, 3, 4, 5, 6, 7],
-    bitrate: 2_048_000,
+// The Opus surround layout table (channel order FL FR FC LFE RL RR [SL SR], identity mapping,
+// Sunshine's per-config bitrates) now lives in `punktfunk_core::audio`, shared with the native
+// `punktfunk/1` path and every client decoder. Re-export the pieces the GameStream module + its
+// RTSP SDP (`rtsp.rs`) reference; the GFE-specific `surround_params` SDP rotation stays below.
+pub use punktfunk_core::audio::{
+    OpusLayout, LAYOUT_51, LAYOUT_51_HQ, LAYOUT_71, LAYOUT_71_HQ, LAYOUT_STEREO,
 };
 
-/// Pick the encoder layout for the negotiated session parameters. Unknown channel counts
-/// fall back to stereo (the client can only request 2/6/8 — `AUDIO_CONFIGURATION_*` in
+/// Pick the encoder layout for the negotiated session parameters. Thin wrapper over the shared
+/// [`punktfunk_core::audio::layout_for`] keyed on this module's [`AudioParams`] (unknown channel
+/// counts fall back to stereo; the client can only request 2/6/8 — `AUDIO_CONFIGURATION_*` in
 /// Limelight.h).
 pub fn layout_for(params: &AudioParams) -> &'static OpusLayout {
-    match (params.channels, params.high_quality) {
-        (6, false) => &LAYOUT_51,
-        (6, true) => &LAYOUT_51_HQ,
-        (8, false) => &LAYOUT_71,
-        (8, true) => &LAYOUT_71_HQ,
-        _ => &LAYOUT_STEREO,
-    }
+    punktfunk_core::audio::layout_for(params.channels, params.high_quality)
 }
 
 /// The `a=fmtp:97 surround-params=` digit string for a layout: channelCount, streams,
@@ -345,21 +296,21 @@ fn run(
 }
 
 /// Opus encoder for one session: the plain stereo encoder (the live-validated path, byte
-/// identical) or a libopus multistream encoder for 5.1/7.1.
+/// identical) or the safe `opus::MSEncoder` multistream encoder for 5.1/7.1. Both are
+/// cross-platform (Linux + Windows) — surround no longer needs `audiopus_sys`.
 #[cfg(any(target_os = "linux", target_os = "windows"))]
 enum SessionEncoder {
     Stereo(opus::Encoder),
-    // Surround needs the libopus *multistream* encoder via `audiopus_sys` (Linux-only dep).
-    #[cfg(target_os = "linux")]
-    Surround(MsEncoder),
+    Surround(opus::MSEncoder),
 }
 
 #[cfg(any(target_os = "linux", target_os = "windows"))]
 impl SessionEncoder {
     fn new(layout: &'static OpusLayout) -> Result<SessionEncoder> {
+        // RESTRICTED_LOWDELAY (`opus::Application::LowDelay`) + hard CBR, matching Sunshine — CBR
+        // keeps the Opus packet size constant, which the GameStream audio FEC (equal-length shards)
+        // relies on, and the client asserts a constant per-stream TOC.
         if layout.channels == 2 {
-            // RESTRICTED_LOWDELAY + CBR, matching Sunshine — CBR keeps the Opus TOC byte
-            // constant, which the client asserts per stream.
             let mut enc = opus::Encoder::new(
                 SAMPLE_RATE,
                 opus::Channels::Stereo,
@@ -370,138 +321,32 @@ impl SessionEncoder {
             enc.set_vbr(false).ok();
             Ok(SessionEncoder::Stereo(enc))
         } else {
-            #[cfg(target_os = "linux")]
-            {
-                Ok(SessionEncoder::Surround(MsEncoder::new(layout)?))
-            }
-            #[cfg(not(target_os = "linux"))]
-            {
-                anyhow::bail!(
-                    "surround audio ({} ch) needs the libopus multistream encoder (Linux only) — \
-                     use a stereo session",
-                    layout.channels
-                )
-            }
+            let mut enc = opus::MSEncoder::new(
+                SAMPLE_RATE,
+                layout.streams,
+                layout.coupled,
+                layout.mapping,
+                opus::Application::LowDelay,
+            )
+            .map_err(|e| anyhow::anyhow!("create Opus multistream encoder: {e}"))?;
+            enc.set_bitrate(opus::Bitrate::Bits(layout.bitrate)).ok();
+            enc.set_vbr(false).ok();
+            Ok(SessionEncoder::Surround(enc))
         }
     }
 
-    /// Encode one interleaved frame (`samples_per_channel * channels` f32s) into `out`,
-    /// returning the packet length.
-    fn encode_float(
-        &mut self,
-        frame: &[f32],
-        samples_per_channel: usize,
-        out: &mut [u8],
-    ) -> Result<usize> {
-        // `samples_per_channel` only feeds the multistream (surround) encoder; stereo infers it.
-        #[cfg(not(target_os = "linux"))]
-        let _ = samples_per_channel;
+    /// Encode one interleaved frame into `out`, returning the packet length. Both encoders infer
+    /// the per-channel sample count from `frame.len()` and their channel count.
+    fn encode_float(&mut self, frame: &[f32], out: &mut [u8]) -> Result<usize> {
         match self {
             SessionEncoder::Stereo(enc) => enc.encode_float(frame, out).context("opus encode"),
-            #[cfg(target_os = "linux")]
-            SessionEncoder::Surround(enc) => enc.encode_float(frame, samples_per_channel, out),
+            SessionEncoder::Surround(enc) => enc
+                .encode_float(frame, out)
+                .context("opus multistream encode"),
         }
     }
 }
 
-/// RAII wrapper for `OpusMSEncoder` (the safe `opus` crate is stereo-only; the multistream
-/// API comes from `audiopus_sys`, the same libopus the crate already links). Configured like
-/// the stereo path: RESTRICTED_LOWDELAY, hard CBR, per-layout bitrate.
-#[cfg(target_os = "linux")]
-struct MsEncoder {
-    st: std::ptr::NonNull<audiopus_sys::OpusMSEncoder>,
-}
-
-// SAFETY: `MsEncoder` owns a unique `OpusMSEncoder` via `NonNull` (it is neither `Clone` nor
-// `Sync`, so the pointer is never aliased). libopus's multistream encoder state is a self-contained
-// heap allocation with no thread-local or thread-affine state, so moving ownership to another thread
-// is sound; every method takes `&mut self`, keeping access single-threaded at any instant.
-#[cfg(target_os = "linux")]
-unsafe impl Send for MsEncoder {}
-
-#[cfg(target_os = "linux")]
-impl MsEncoder {
-    fn new(layout: &OpusLayout) -> Result<MsEncoder> {
-        use std::os::raw::c_int;
-        let mut err: c_int = 0;
-        // SAFETY: every scalar arg is a valid libopus input (sample rate, channel/stream/coupled
-        // counts, the RESTRICTED_LOWDELAY application constant). `layout.mapping.as_ptr()` addresses
-        // a 'static slice of exactly `layout.channels` bytes (every `OpusLayout` constant upholds
-        // that), which is the element count `opus_multistream_encoder_create` reads through it, and
-        // `&mut err` is a live local the call writes its status into. libopus copies the mapping into
-        // its own allocation, so the pointer need only be valid for the call; the returned pointer is
-        // null/`OPUS_OK`-checked below before any use.
-        let st = unsafe {
-            audiopus_sys::opus_multistream_encoder_create(
-                SAMPLE_RATE as i32,
-                layout.channels as c_int,
-                layout.streams as c_int,
-                layout.coupled as c_int,
-                layout.mapping.as_ptr(),
-                audiopus_sys::OPUS_APPLICATION_RESTRICTED_LOWDELAY,
-                &mut err,
-            )
-        };
-        let st = std::ptr::NonNull::new(st)
-            .filter(|_| err == audiopus_sys::OPUS_OK)
-            .ok_or_else(|| anyhow::anyhow!("opus_multistream_encoder_create failed ({err})"))?;
-        // SAFETY: `st` is the non-null encoder `opus_multistream_encoder_create` just returned, owned
-        // exclusively here. Each `opus_multistream_encoder_ctl` call passes a valid request constant
-        // with the single by-value `c_int` argument that request's variadic ABI expects
-        // (`OPUS_SET_BITRATE_REQUEST` → bitrate, `OPUS_SET_VBR_REQUEST` → 0). No pointer escapes the
-        // call and the encoder outlives it.
-        unsafe {
-            audiopus_sys::opus_multistream_encoder_ctl(
-                st.as_ptr(),
-                audiopus_sys::OPUS_SET_BITRATE_REQUEST,
-                layout.bitrate as c_int,
-            );
-            audiopus_sys::opus_multistream_encoder_ctl(
-                st.as_ptr(),
-                audiopus_sys::OPUS_SET_VBR_REQUEST,
-                0 as c_int, // hard CBR (constant packet size — also what audio FEC relies on)
-            );
-        }
-        Ok(MsEncoder { st })
-    }
-
-    fn encode_float(
-        &mut self,
-        frame: &[f32],
-        samples_per_channel: usize,
-        out: &mut [u8],
-    ) -> Result<usize> {
-        // SAFETY: `self.st` is the live encoder from `new`. libopus reads `samples_per_channel *
-        // channels` f32s through `frame.as_ptr()`; every caller passes a `frame` of exactly that
-        // length together with the matching `samples_per_channel` (`audio_body`'s `frame_len =
-        // samples_per_channel * layout.channels`; the round-trip tests size identically), so the read
-        // stays in bounds. `out.as_mut_ptr()` is written for at most `out.len()` bytes, which is
-        // passed as the capacity bound. Both buffers are live locals outliving this synchronous call;
-        // the return value is range-checked before being used as a length.
-        let n = unsafe {
-            audiopus_sys::opus_multistream_encode_float(
-                self.st.as_ptr(),
-                frame.as_ptr(),
-                samples_per_channel as std::os::raw::c_int,
-                out.as_mut_ptr(),
-                out.len() as i32,
-            )
-        };
-        anyhow::ensure!(n > 0, "opus_multistream_encode_float failed ({n})");
-        Ok(n as usize)
-    }
-}
-
-#[cfg(target_os = "linux")]
-impl Drop for MsEncoder {
-    fn drop(&mut self) {
-        // SAFETY: `self.st` is the encoder `opus_multistream_encoder_create` returned; this
-        // `MsEncoder` owns it uniquely and `drop` runs exactly once, so the destroy frees it once
-        // with no subsequent use.
-        unsafe { audiopus_sys::opus_multistream_encoder_destroy(self.st.as_ptr()) }
-    }
-}
-
 #[cfg(any(target_os = "linux", target_os = "windows"))]
 fn audio_body(
     cap: &mut dyn AudioCapturer,
@@ -565,7 +410,7 @@ fn audio_body(
                     *s = (*s * gain).clamp(-1.0, 1.0);
                 }
             }
-            let n = enc.encode_float(&frame, samples_per_channel, &mut out)?;
+            let n = enc.encode_float(&frame, &mut out)?;
             // AES-128-CBC the Opus payload (RTP header stays plaintext). Per-packet IV =
             // BE32(rikeyid + seq) in [0..4], zero elsewhere; PKCS7 padding.
             let iv_seq = (rikeyid as u32).wrapping_add(seq as u32);
@@ -775,41 +620,33 @@ mod tests {
     /// Real-codec proof of the 5.1 mapping math: encode with our encoder layout, decode with
     /// the mapping a stock Moonlight client derives from our advertised surround-params
     /// (parse → GFE swap), and verify a tone fed into each input channel comes out on the
-    /// same output channel.
-    #[cfg(target_os = "linux")]
+    /// same output channel. Cross-platform via the safe `opus` crate — this also guards the
+    /// (now un-gated) Windows GameStream surround build.
     #[test]
     fn multistream_51_roundtrip_channel_identity() {
         let layout = &LAYOUT_51;
         let samples = 240; // 5 ms
         let ch = layout.channels as usize;
 
-        // Client-side decoder mapping derived exactly as moonlight-common-c does.
+        // Client-side decoder mapping derived exactly as moonlight-common-c does (GFE swap).
         let s = surround_params(layout, false);
         let digits: Vec<u8> = s.bytes().map(|b| b - b'0').collect();
         let client_mapping = client_swap(&digits[3..]);
 
-        let mut err = 0i32;
-        // SAFETY: scalar args are valid libopus inputs. `client_mapping.as_ptr()` addresses a
-        // `Vec<u8>` of exactly `ch` entries (derived from the advertised surround-params), which is
-        // the element count the decoder reads through it, and `&mut err` is a live local the call
-        // writes. The returned pointer is `OPUS_OK`/non-null-checked immediately below before use.
-        let dec = unsafe {
-            audiopus_sys::opus_multistream_decoder_create(
-                SAMPLE_RATE as i32,
-                ch as i32,
-                layout.streams as i32,
-                layout.coupled as i32,
-                client_mapping.as_ptr(),
-                &mut err,
-            )
-        };
-        assert_eq!(err, audiopus_sys::OPUS_OK);
-        assert!(!dec.is_null());
+        let mut dec =
+            opus::MSDecoder::new(SAMPLE_RATE, layout.streams, layout.coupled, &client_mapping)
+                .expect("multistream decoder");
 
         for tone_ch in 0..ch {
-            let mut enc = MsEncoder::new(layout).unwrap();
+            let mut enc = opus::MSEncoder::new(
+                SAMPLE_RATE,
+                layout.streams,
+                layout.coupled,
+                layout.mapping,
+                opus::Application::LowDelay,
+            )
+            .expect("multistream encoder");
             let mut out = vec![0u8; 1400];
-            let mut decoded = vec![0f32; samples * ch];
             let mut energy = vec![0f64; ch];
             // A few frames so the codec converges past its startup transient.
             for f in 0..8 {
@@ -819,28 +656,15 @@ mod tests {
                         / SAMPLE_RATE as f32;
                     frame[t * ch + tone_ch] = 0.5 * phase.sin();
                 }
-                let n = enc.encode_float(&frame, samples, &mut out).unwrap();
+                let n = enc.encode_float(&frame, &mut out).unwrap();
                 assert!(n > 0);
-                // SAFETY: `dec` is the non-null decoder asserted above. `out.as_ptr()` is read for
-                // the `n` encoded bytes just produced by `encode_float`; `decoded.as_mut_ptr()` is
-                // written for up to `samples * ch` f32s and `decoded` is exactly that long; `samples`
-                // is the per-channel frame size. All buffers are live locals outliving the call; the
-                // return is checked to equal `samples`.
-                let got = unsafe {
-                    audiopus_sys::opus_multistream_decode_float(
-                        dec,
-                        out.as_ptr(),
-                        n as i32,
-                        decoded.as_mut_ptr(),
-                        samples as i32,
-                        0,
-                    )
-                };
-                assert_eq!(got as usize, samples);
+                let mut decoded = vec![0f32; samples * ch];
+                let got = dec.decode_float(&out[..n], &mut decoded, false).unwrap();
+                assert_eq!(got, samples);
                 if f >= 4 {
                     for t in 0..samples {
-                        for c in 0..ch {
-                            energy[c] += (decoded[t * ch + c] as f64).powi(2);
+                        for (c, e) in energy.iter_mut().enumerate() {
+                            *e += (decoded[t * ch + c] as f64).powi(2);
                         }
                     }
                 }
@@ -854,9 +678,6 @@ mod tests {
                  (energies: {energy:?})"
             );
         }
-        // SAFETY: `dec` is the decoder `opus_multistream_decoder_create` returned; the test owns it
-        // and destroys it exactly once here, after the final decode — no later use, no double free.
-        unsafe { audiopus_sys::opus_multistream_decoder_destroy(dec) };
     }
 
     /// Live 5.1 capture → multistream encode → decode, against a real PipeWire session.
@@ -869,7 +690,15 @@ mod tests {
     fn surround_capture_live() {
         let mut cap = crate::audio::open_audio_capture(6).expect("open 6ch capture");
         let layout = &LAYOUT_51;
-        let mut enc = MsEncoder::new(layout).unwrap();
+        let mut enc = opus::MSEncoder::new(
+            SAMPLE_RATE,
+            layout.streams,
+            layout.coupled,
+            layout.mapping,
+            opus::Application::LowDelay,
+        )
+        .unwrap();
+        enc.set_vbr(false).ok(); // hard CBR so packet sizes are constant (audio FEC relies on it)
         let mut out = vec![0u8; 1400];
         let mut acc: Vec<f32> = Vec::new();
         let frame_len = 240 * 6;
@@ -880,49 +709,24 @@ mod tests {
             acc.extend_from_slice(&chunk);
             while acc.len() >= frame_len && packets < 100 {
                 let frame: Vec<f32> = acc.drain(..frame_len).collect();
-                let n = enc.encode_float(&frame, 240, &mut out).unwrap();
+                let n = enc.encode_float(&frame, &mut out).unwrap();
                 sizes.insert(n);
                 packets += 1;
             }
         }
         // Hard CBR: every multistream packet must be the same size (audio FEC relies on it).
         assert_eq!(sizes.len(), 1, "CBR sizes: {sizes:?}");
-        // And a stock client's decoder must accept them.
+        // And a stock client's GFE-derived decoder must accept them.
         let s = surround_params(layout, false);
         let digits: Vec<u8> = s.bytes().map(|b| b - b'0').collect();
         let client_mapping = client_swap(&digits[3..]);
-        let mut err = 0i32;
-        // SAFETY: scalar args are valid; `client_mapping.as_ptr()` addresses a 6-entry `Vec<u8>`
-        // (matches the 6-channel layout the decoder reads through it), alive past the call, and
-        // `&mut err` is a live local. The pointer is `OPUS_OK`-checked before use.
-        let dec = unsafe {
-            audiopus_sys::opus_multistream_decoder_create(
-                48000,
-                6,
-                layout.streams as i32,
-                layout.coupled as i32,
-                client_mapping.as_ptr(),
-                &mut err,
-            )
-        };
-        assert_eq!(err, audiopus_sys::OPUS_OK);
+        let mut dec =
+            opus::MSDecoder::new(SAMPLE_RATE, layout.streams, layout.coupled, &client_mapping)
+                .unwrap();
         let mut pcm = vec![0f32; 240 * 6];
-        // SAFETY: `dec` is the non-null decoder from create. `out.as_ptr()` is read for the CBR
-        // packet length passed in (`*sizes.first()`, a real encoded packet size in `out`);
-        // `pcm.as_mut_ptr()` is written for up to `240 * 6` f32s and `pcm` is exactly that long;
-        // `240` is the per-channel frame size. All buffers are live locals outliving the call.
-        let got = unsafe {
-            audiopus_sys::opus_multistream_decode_float(
-                dec,
-                out.as_ptr(),
-                *sizes.first().unwrap() as i32,
-                pcm.as_mut_ptr(),
-                240,
-                0,
-            )
-        };
-        // SAFETY: `dec` is owned by the test; destroyed exactly once here after the final decode.
-        unsafe { audiopus_sys::opus_multistream_decoder_destroy(dec) };
+        let got = dec
+            .decode_float(&out[..*sizes.first().unwrap()], &mut pcm, false)
+            .unwrap();
         assert_eq!(got, 240);
     }
 }

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

@@ -114,12 +114,12 @@ fn run(
     // `video_cap`, since a reconnect at a different resolution needs a freshly-sized output; the
     // output is released when this capturer drops at stream end (RAII via its keepalive).
     if crate::config::config().video_source.as_deref() == Some("virtual") {
-        // The launched app picks the compositor (e.g. gamescope for game entries) and the
-        // nested command.
-        let compositor = app
-            .and_then(|a| a.compositor)
-            .map(Ok)
-            .unwrap_or_else(|| crate::vdisplay::detect().context("detect compositor"))?;
+        // Open the virtual-display source: pick the live compositor, normalize the session env
+        // (apply_session_env/apply_input_env — gamescope ATTACH/resize + KWin/Mutter retargeting,
+        // exactly like the native plane), create a virtual output at the client mode, and capture it.
+        // Re-runnable: the encode loop calls it again on a mid-stream capture loss to FOLLOW a
+        // Desktop<->Game switch.
+        let (mut capturer, compositor) = open_gs_virtual_source(cfg, app)?;
         tracing::info!(
             ?compositor,
             app = ?app.map(|a| &a.title),
@@ -127,31 +127,6 @@ fn run(
             h = cfg.height,
             "video source: virtual display (native client resolution)"
         );
-        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
-        // process-global env var, so concurrent sessions can't stomp each other's launch target.
-        vd.set_launch_command(app.and_then(|a| a.cmd.clone()));
-        let vout = vd
-            .create(punktfunk_core::Mode {
-                width: cfg.width,
-                height: cfg.height,
-                refresh_hz: cfg.fps,
-            })
-            .context("create virtual output at client resolution")?;
-        // `want_hdr=false`: the IDD-push backend (opt-in PUNKTFUNK_IDD_PUSH) has no monitor-HDR
-        // auto-detection — it converts its always-FP16 ring per this flag — and GameStream HDR is not
-        // negotiated into StreamConfig here, so an IDD-push GameStream session streams SDR even on an
-        // HDR desktop. (The default WGC backend DOES auto-detect HDR from the output colorspace, but
-        // IDD-push bypasses WGC.) Acceptable for the experimental IDD-push A/B path; HDR over IDD-push
-        // is wired only for punktfunk/1 (want_hdr = negotiated bit_depth >= 10). TODO: derive want_hdr
-        // from a GameStream HDR flag once StreamConfig carries one.
-        let mut capturer = capture::capture_virtual_output(
-            vout,
-            capture::OutputFormat::resolve(false),
-            crate::session_plan::CaptureBackend::resolve(),
-        )
-        .context("capture virtual output")?;
-        capturer.set_active(true);
         // Launch the app's command now that capture is live, for the backends that DON'T nest it via
         // set_launch_command above: Windows (no gamescope) and Linux kwin/mutter/wlroots (which stream
         // the existing desktop, so the app must be spawned into the session to land on the streamed
@@ -171,8 +146,14 @@ fn run(
                 }
             }
         }
+        // Rebuild closure: re-open the source on a mid-stream capture loss, RE-DETECTING the live
+        // compositor — so a Desktop<->Game switch (at the client's fixed mode) is FOLLOWED in place
+        // without a Moonlight reconnect. (A resolution change can't be followed mid-stream on
+        // GameStream — WxH is locked at ANNOUNCE — but a session toggle keeps the negotiated mode.)
+        let rebuild = || open_gs_virtual_source(cfg, app).map(|(c, _)| c);
         return stream_body(
-            &mut *capturer,
+            &mut capturer,
+            Some(&rebuild),
             &sock,
             cfg,
             running,
@@ -200,8 +181,10 @@ fn run(
         }
     };
     capturer.set_active(true);
+    // Portal/synthetic source: no compositor virtual output to re-detect, so no rebuild closure.
     let result = stream_body(
-        &mut *capturer,
+        &mut capturer,
+        None,
         &sock,
         cfg,
         running,
@@ -215,6 +198,53 @@ fn run(
     result
 }
 
+/// Open the virtual-display video source for a GameStream session: pick the LIVE compositor + normalize
+/// the session env (apply_session_env/apply_input_env — gamescope ATTACH/resize, KWin/Mutter
+/// retargeting) exactly like the native plane (punktfunk1.rs resolve_compositor), create a virtual
+/// output at the client's mode, and capture it. Returns the capturer (it owns the output's keepalive;
+/// the stateless VirtualDisplay factory is dropped here) plus the resolved compositor. An apps.json
+/// entry can PIN a compositor (skips the live detect/retarget). Re-run on a mid-stream capture loss to
+/// FOLLOW a Desktop<->Game switch: it re-detects the now-live compositor and re-targets at it. Does NOT
+/// launch the app (that happens once at stream start; a rebuild must not re-spawn it).
+fn open_gs_virtual_source(
+    cfg: StreamConfig,
+    app: Option<&super::apps::AppEntry>,
+) -> Result<(Box<dyn Capturer>, crate::vdisplay::Compositor)> {
+    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
+    };
+    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
+    // process-global env var, so concurrent sessions can't stomp each other's launch target.
+    vd.set_launch_command(app.and_then(|a| a.cmd.clone()));
+    let vout = vd
+        .create(punktfunk_core::Mode {
+            width: cfg.width,
+            height: cfg.height,
+            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).
+    let capturer = capture::capture_virtual_output(
+        vout,
+        capture::OutputFormat::resolve(false),
+        crate::session_plan::CaptureBackend::resolve(),
+    )
+    .context("capture virtual output")?;
+    capturer.set_active(true);
+    Ok((capturer, compositor))
+}
+
 /// One frame's packets, handed from the encode thread to the send thread.
 type PacketBatch = Vec<Vec<u8>>;
 
@@ -367,7 +397,11 @@ fn percentile(v: &mut [u32], q: f64) -> u32 {
 /// (see [`spawn_sender`]) so a send spike can never stall capture/encode.
 #[allow(clippy::too_many_arguments)]
 fn stream_body(
-    capturer: &mut dyn Capturer,
+    // `&mut Box` (not `&mut dyn`) so a mid-stream capture-loss rebuild can SWAP the capturer in place.
+    capturer: &mut Box<dyn Capturer>,
+    // Re-open the video source on capture loss (virtual-display path → follow a Desktop<->Game switch);
+    // `None` for the portal/synthetic source, which has nothing to re-detect (propagate the error).
+    rebuild: Option<&dyn Fn() -> Result<Box<dyn Capturer>>>,
     sock: &UdpSocket,
     cfg: StreamConfig,
     running: &Arc<AtomicBool>,
@@ -397,6 +431,9 @@ fn stream_body(
         cfg.bitrate_kbps as u64 * 1000,
         frame.is_cuda(),
         8, // GameStream/Moonlight path: 8-bit (its own codec negotiation)
+        // 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.
+        encode::ChromaFormat::Yuv420,
     )
     .context("open video encoder for stream")?;
     // FEC overhead percent (Sunshine default 20). Override with PUNKTFUNK_FEC_PCT (0 = data-only).
@@ -459,7 +496,12 @@ fn stream_body(
     // RFI capability is fixed for the session (probed at encoder open). Query it once so the
     // recovery path skips the always-`false` invalidate call on encoders without NVENC RFI and
     // forces a keyframe directly instead.
-    let supports_rfi = enc.caps().supports_rfi;
+    let mut supports_rfi = enc.caps().supports_rfi;
+
+    // Bound consecutive capture-loss rebuilds (a delivered frame clears the counter) so a permanently
+    // dead source can't loop forever — it ends the stream after the cap, falling back to a reconnect.
+    const MAX_REBUILDS: u32 = 5;
+    let mut rebuilds: u32 = 0;
 
     while running.load(Ordering::SeqCst) {
         let tick = Instant::now();
@@ -467,9 +509,69 @@ fn stream_body(
         // armed (cheap Relaxed atomic, re-read each frame).
         let measure = perf || stats.is_armed();
         // Advance to the freshest captured frame if one arrived; otherwise reuse the last.
-        if let Some(f) = capturer.try_latest().context("capture frame")? {
-            frame = f;
-            uniq += 1;
+        match capturer.try_latest() {
+            Ok(Some(f)) => {
+                frame = f;
+                uniq += 1;
+                rebuilds = 0; // a delivered frame clears the consecutive-loss counter
+            }
+            Ok(None) => {} // no new frame — reuse the last (static/idle desktop)
+            Err(e) => {
+                // The capture source went away — the compositor was torn down on a Desktop<->Game
+                // switch, or the virtual output was removed. On the virtual-display path, re-detect the
+                // now-live compositor and re-attach IN PLACE (the send thread + packetizer + socket +
+                // RTP clock all survive), then force an IDR so Moonlight resyncs — so the stream FOLLOWS
+                // the switch with no client reconnect. Build the new source BEFORE dropping the old.
+                // Bounded by a counter + a ~40s budget; on exhaustion, end the stream (Moonlight
+                // reconnect). The portal/synthetic path has no rebuild closure → propagate as before.
+                let Some(rebuild) = rebuild else {
+                    return Err(e).context("capture frame");
+                };
+                rebuilds += 1;
+                if rebuilds > MAX_REBUILDS {
+                    return Err(e).context("capture lost — rebuild attempts exhausted");
+                }
+                tracing::warn!(error = %format!("{e:#}"), rebuild = rebuilds,
+                    "gamestream: capture lost — rebuilding source in place (following a session switch)");
+                let rebuild_deadline = Instant::now() + Duration::from_secs(40);
+                let new_cap = loop {
+                    match rebuild() {
+                        Ok(c) => break c,
+                        Err(e2) => {
+                            if !running.load(Ordering::SeqCst) || Instant::now() >= rebuild_deadline
+                            {
+                                return Err(e2)
+                                    .context("capture lost — no source within the rebuild budget");
+                            }
+                            tracing::warn!(error = %format!("{e2:#}"),
+                                "gamestream: source not up yet — retrying");
+                            std::thread::sleep(Duration::from_millis(500));
+                        }
+                    }
+                };
+                *capturer = new_cap;
+                capturer.set_active(true);
+                frame = capturer.next_frame().context("first frame after rebuild")?;
+                // Re-open the encoder for the new source (same negotiated WxH → same SPS profile) and
+                // force an IDR so Moonlight resyncs on the first emitted AU.
+                enc = encode::open_video(
+                    cfg.codec,
+                    frame.format,
+                    frame.width,
+                    frame.height,
+                    cfg.fps,
+                    cfg.bitrate_kbps as u64 * 1000,
+                    frame.is_cuda(),
+                    8,
+                    encode::ChromaFormat::Yuv420, // GameStream stays 4:2:0
+                )
+                .context("reopen encoder after rebuild")?;
+                supports_rfi = enc.caps().supports_rfi;
+                enc.request_keyframe();
+                next_frame = Instant::now();
+                tracing::info!("gamestream: source rebuilt — stream continues");
+                continue;
+            }
         }
         let t_cap = tick.elapsed();
         // Honor a client recovery request. Prefer reference-frame invalidation (the encoder

crates/punktfunk-host/src/inject.rs

@@ -24,6 +24,9 @@ pub trait InputInjector {
 pub enum Backend {
     /// wlroots virtual pointer + keyboard Wayland protocols — the headless-Sway path.
     WlrVirtual,
+    /// KWin `org_kde_kwin_fake_input` — direct injection, no RemoteDesktop portal / approval dialog
+    /// (authorized by the host's `.desktop`). The headless KDE-Desktop path; what krdpserver uses.
+    KwinFakeInput,
     /// libei via `reis` — Wayland-native (RemoteDesktop portal). Not yet implemented.
     Libei,
     /// libei directly against gamescope's own EIS socket (no portal): input lands in the
@@ -47,6 +50,16 @@ pub fn open(backend: Backend) -> Result<Box<dyn InputInjector>> {
                 anyhow::bail!("wlroots virtual input requires Linux + a Wayland compositor")
             }
         }
+        Backend::KwinFakeInput => {
+            #[cfg(target_os = "linux")]
+            {
+                Ok(Box::new(kwin_fake_input::KwinFakeInjector::open()?))
+            }
+            #[cfg(not(target_os = "linux"))]
+            {
+                anyhow::bail!("KWin fake_input requires Linux + a KWin Wayland session")
+            }
+        }
         Backend::Libei => {
             #[cfg(target_os = "linux")]
             {
@@ -90,12 +103,18 @@ pub fn open(backend: Backend) -> Result<Box<dyn InputInjector>> {
 /// Pick the injection backend for the current session. gamescope hosts its own EIS server (no
 /// portal), so a gamescope session injects directly into it. wlroots/Sway only implements the
 /// ScreenCast portal (no RemoteDesktop), so libei can't run there — use the wlr virtual-input
-/// protocols. KWin and GNOME implement RemoteDesktop but not the wlr protocols, so use libei.
-/// `PUNKTFUNK_INPUT_BACKEND=wlr|libei|gamescope|uinput` overrides the auto-detection.
+/// protocols. **KWin** exposes `org_kde_kwin_fake_input` (direct injection, no portal / approval
+/// dialog — the only headless-capable path; what krdpserver uses), so prefer it there. **GNOME**
+/// has neither fake_input nor the wlr protocols, so it uses libei via the RemoteDesktop portal
+/// (which needs a user to approve, or a pre-seeded grant — not truly headless).
+/// `PUNKTFUNK_INPUT_BACKEND=wlr|kwin|libei|gamescope|uinput` overrides the auto-detection.
 pub fn default_backend() -> Backend {
     if let Ok(v) = std::env::var("PUNKTFUNK_INPUT_BACKEND") {
         match v.trim().to_ascii_lowercase().as_str() {
             "wlr" | "wlroots" | "wlrvirtual" => return Backend::WlrVirtual,
+            "kwin" | "fakeinput" | "fake_input" | "kwin-fake-input" => {
+                return Backend::KwinFakeInput
+            }
             "libei" | "ei" | "portal" => return Backend::Libei,
             "gamescope" | "gamescope-ei" => return Backend::GamescopeEi,
             "uinput" => return Backend::Uinput,
@@ -112,16 +131,26 @@ pub fn default_backend() -> Backend {
     }
     #[cfg(not(target_os = "windows"))]
     {
-        if crate::config::config()
-            .compositor
-            .as_deref()
-            .is_some_and(|v| v.trim().eq_ignore_ascii_case("gamescope"))
-        {
-            return Backend::GamescopeEi;
+        // An explicit compositor pick (set per connect / mid-stream) is the strongest signal.
+        let compositor = crate::config::config().compositor.clone();
+        if let Some(c) = compositor.as_deref() {
+            let c = c.trim();
+            if c.eq_ignore_ascii_case("gamescope") {
+                return Backend::GamescopeEi;
+            }
+            if c.eq_ignore_ascii_case("kwin") {
+                return Backend::KwinFakeInput;
+            }
+            if c.eq_ignore_ascii_case("wlroots") || c.eq_ignore_ascii_case("sway") {
+                return Backend::WlrVirtual;
+            }
+            // mutter (GNOME) falls through to the XDG_CURRENT_DESKTOP check below.
         }
         let desktop = std::env::var("XDG_CURRENT_DESKTOP").unwrap_or_default();
         let d = desktop.to_ascii_uppercase();
-        if d.contains("KDE") || d.contains("GNOME") {
+        if d.contains("KDE") {
+            Backend::KwinFakeInput
+        } else if d.contains("GNOME") {
             Backend::Libei
         } else {
             Backend::WlrVirtual
@@ -478,6 +507,9 @@ pub mod gamepad {
     }
 }
 #[cfg(target_os = "linux")]
+#[path = "inject/linux/kwin_fake_input.rs"]
+mod kwin_fake_input;
+#[cfg(target_os = "linux")]
 #[path = "inject/linux/libei.rs"]
 mod libei;
 #[cfg(target_os = "windows")]

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

@@ -0,0 +1,209 @@
+//! Headless input injection on KWin via the privileged `org_kde_kwin_fake_input` protocol — the
+//! exact path KDE's own headless RDP server (`krdpserver`) uses. KWin advertises this restricted
+//! global only to a client authorized through its installed `.desktop` `X-KDE-Wayland-Interfaces`
+//! (we ship `io.unom.Punktfunk.Host.desktop`, which lists `org_kde_kwin_fake_input` alongside
+//! `zkde_screencast_unstable_v1`). Binding the global IS the authorization, so injection needs **no
+//! RemoteDesktop portal and no "Allow remote control?" dialog** — it works with no user present,
+//! 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.
+
+#![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).
+#![deny(clippy::undocumented_unsafe_blocks)]
+
+use super::{gs_button_to_evdev, vk_to_evdev, InputEvent, InputInjector};
+use anyhow::{Context, Result};
+use punktfunk_core::input::InputKind;
+use wayland_client::protocol::wl_registry::{self, WlRegistry};
+use wayland_client::{Connection, Dispatch, EventQueue, Proxy, QueueHandle};
+
+// 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.
+#[allow(clippy::all, dead_code, non_camel_case_types, non_snake_case, unused)]
+pub mod fake {
+    use wayland_client;
+    use wayland_client::protocol::*;
+
+    pub mod __interfaces {
+        use wayland_client::protocol::__interfaces::*;
+        wayland_scanner::generate_interfaces!("protocols/fake-input.xml");
+    }
+    use self::__interfaces::*;
+
+    wayland_scanner::generate_client_code!("protocols/fake-input.xml");
+}
+
+use fake::org_kde_kwin_fake_input::OrgKdeKwinFakeInput as FakeInput;
+
+/// Highest interface version we drive. `keyboard_key` arrived at v4; KWin advertises ≥4.
+const MAX_VERSION: u32 = 4;
+
+/// `wl_pointer.axis` values used by `axis`.
+const AXIS_VERTICAL: u32 = 0;
+const AXIS_HORIZONTAL: u32 = 1;
+/// `code` value marking a horizontal scroll event (mirrors `gamestream::input` / the wlr backend).
+const SCROLL_HORIZONTAL: u32 = 1;
+
+/// Registry-bound globals (the Wayland dispatch state).
+#[derive(Default)]
+struct State {
+    fake: Option<FakeInput>,
+}
+
+impl Dispatch<WlRegistry, ()> for State {
+    fn event(
+        state: &mut Self,
+        registry: &WlRegistry,
+        event: wl_registry::Event,
+        _: &(),
+        _: &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, ()));
+            }
+        }
+    }
+}
+
+// fake_input emits no events.
+impl Dispatch<FakeInput, ()> for State {
+    fn event(
+        _: &mut Self,
+        _: &FakeInput,
+        _: <FakeInput as Proxy>::Event,
+        _: &(),
+        _: &Connection,
+        _: &QueueHandle<Self>,
+    ) {
+    }
+}
+
+pub struct KwinFakeInjector {
+    conn: Connection,
+    queue: EventQueue<State>,
+    state: State,
+    fake: FakeInput,
+}
+
+impl KwinFakeInjector {
+    pub fn open() -> Result<Self> {
+        let conn = Connection::connect_to_env()
+            .context("connect to KWin Wayland (is WAYLAND_DISPLAY set to the KWin socket?)")?;
+        let mut queue = conn.new_event_queue();
+        let qh = queue.handle();
+        let _registry = conn.display().get_registry(&qh, ());
+        let mut state = State::default();
+        queue
+            .roundtrip(&mut state)
+            .context("Wayland registry roundtrip")?;
+
+        let fake = state.fake.clone().context(
+            "KWin does not expose org_kde_kwin_fake_input to this client — install the host's \
+             .desktop (io.unom.Punktfunk.Host.desktop, X-KDE-Wayland-Interfaces) and re-login so \
+             KWin authorizes it (the grant is cached per-exe on first connect), or this is not a \
+             KWin session",
+        )?;
+        // Authenticate (the legacy handshake; for an interface-authorized client KWin accepts it
+        // without a dialog — same as krdpserver/krfb headless).
+        fake.authenticate("punktfunk".into(), "remote streaming input".into());
+        queue
+            .roundtrip(&mut state)
+            .context("fake_input authenticate roundtrip")?;
+        conn.flush().ok();
+
+        tracing::info!("KWin fake_input ready (headless keyboard/mouse/touch — no portal)");
+        Ok(Self {
+            conn,
+            queue,
+            state,
+            fake,
+        })
+    }
+}
+
+impl InputInjector for KwinFakeInjector {
+    fn inject(&mut self, event: &InputEvent) -> Result<()> {
+        match event.kind {
+            InputKind::MouseMove => {
+                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;
+                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);
+                }
+            }
+            InputKind::MouseButtonDown | InputKind::MouseButtonUp => {
+                if let Some(btn) = gs_button_to_evdev(event.code) {
+                    let st = u32::from(event.kind == InputKind::MouseButtonDown);
+                    self.fake.button(btn, st);
+                }
+            }
+            InputKind::MouseScroll => {
+                // GameStream sends WHEEL_DELTA(120)-scaled units; a notch ≈ 15px. Vertical flips
+                // sign on the Wayland axis, horizontal passes through — same as the wlr backend.
+                let horizontal = event.code == SCROLL_HORIZONTAL;
+                let axis = if horizontal {
+                    AXIS_HORIZONTAL
+                } else {
+                    AXIS_VERTICAL
+                };
+                let notches = event.x as f64 / 120.0;
+                let sign = if horizontal { 1.0 } else { -1.0 };
+                self.fake.axis(axis, sign * notches * 15.0);
+            }
+            InputKind::KeyDown | InputKind::KeyUp => {
+                // Raw evdev keycode; KWin resolves it through the session's own keymap (and tracks
+                // modifier state itself, so no separate modifiers request is needed).
+                if let Some(evdev) = vk_to_evdev(event.code as u8) {
+                    let st = u32::from(event.kind == InputKind::KeyDown);
+                    self.fake.keyboard_key(evdev as u32, st);
+                } else {
+                    tracing::debug!(vk = event.code, "unmapped VK keycode — dropped");
+                }
+            }
+            // 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;
+                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;
+                    if event.kind == InputKind::TouchDown {
+                        self.fake.touch_down(event.code, x, y);
+                    } else {
+                        self.fake.touch_motion(event.code, x, y);
+                    }
+                    self.fake.touch_frame();
+                }
+            }
+            InputKind::TouchUp => {
+                self.fake.touch_up(event.code);
+                self.fake.touch_frame();
+            }
+            // Gamepads are injected through uinput, not the compositor.
+            InputKind::GamepadButton | InputKind::GamepadAxis => {}
+        }
+        // Surface protocol errors / disconnects, then push the batch to the compositor.
+        self.queue
+            .dispatch_pending(&mut self.state)
+            .context("wayland dispatch")?;
+        self.conn.flush().context("wayland flush")?;
+        Ok(())
+    }
+}

crates/punktfunk-host/src/punktfunk1.rs

@@ -355,6 +355,15 @@ fn resolve_bitrate_kbps(requested: u32) -> u32 {
     }
 }
 
+/// Resolve the audio channel count the session will capture + encode from the client's request.
+/// Normalizes to one of 2 (stereo) / 6 (5.1) / 8 (7.1); anything else (older client, garbage)
+/// becomes stereo. Both backends can produce the requested count (PipeWire pads/upmixes positions,
+/// WASAPI loopback up/downmixes via AUTOCONVERTPCM), so no capability clamp is needed here — the
+/// surround channels just carry up/downmixed content when the host's sink has fewer real channels.
+fn resolve_audio_channels(requested: u8) -> u8 {
+    punktfunk_core::audio::normalize_channels(requested)
+}
+
 /// Static FEC override: `PUNKTFUNK_FEC_PCT`, when set, PINS the recovery percent and DISABLES
 /// adaptive FEC — so a speed test / measurement keeps a fixed, known overhead. `None` ⇒ adaptive
 /// FEC (the host sizes recovery to the loss the client reports). `0` disables FEC entirely.
@@ -623,6 +632,17 @@ async fn serve_session(
             "encoder bitrate"
         );
 
+        // Resolve the audio channel count (client request → stereo / 5.1 / 7.1). The capturer opens
+        // at this count: PipeWire synthesizes the requested positions (padding with silence when the
+        // sink has fewer), WASAPI loopback up/downmixes via AUTOCONVERTPCM — so a client always gets
+        // the channels it asked for, and the Welcome echoes the value the audio thread will encode.
+        let audio_channels = resolve_audio_channels(hello.audio_channels);
+        tracing::info!(
+            requested = hello.audio_channels,
+            resolved = audio_channels,
+            "audio channels"
+        );
+
         // Resolve the encode bit depth: HEVC Main10 only when the client advertised it AND the host
         // opted in (PUNKTFUNK_10BIT). A client that can't decode 10-bit (caps bit clear, or an older
         // client) always gets the 8-bit stream. PUNKTFUNK_10BIT is the host policy gate until a
@@ -642,6 +662,44 @@ async fn serve_session(
             "encode bit depth"
         );
 
+        // Resolve the chroma subsampling: full-chroma HEVC 4:4:4 only when ALL of — the host opted in
+        // (PUNKTFUNK_444), the client advertised VIDEO_CAP_444, the session is single-process (the
+        // two-process WGC relay encodes 4:2:0 in v1), and the active GPU/driver actually supports a
+        // 4:4:4 encode (probed, cached). The native path always encodes HEVC. We resolve this BEFORE
+        // the Welcome so `chroma_format` reflects what we'll really emit — the honest-downgrade
+        // channel: if any gate fails the client is told 4:2:0 before it builds its decoder. The probe
+        // opens a tiny encoder; it runs only when both opt-ins are set and is cached after the first.
+        let host_wants_444 = crate::config::config().four_four_four;
+        let client_supports_444 = hello.video_caps & punktfunk_core::quic::VIDEO_CAP_444 != 0;
+        let single_process = crate::session_plan::resolve_topology()
+            == crate::session_plan::SessionTopology::SingleProcess;
+        // The GPU probe opens a real (tiny) encoder on first use, so run it off the reactor like the
+        // compositor probe above (blocking probes → spawn_blocking). Short-circuit so it only runs when
+        // the cheap gates already pass. The result is cached process-wide (a negative latches until
+        // restart — acceptable: a GPU either supports HEVC 4:4:4 or it doesn't, and a transient open
+        // failure here is rare since the session's own encoder isn't open yet).
+        let gpu_supports_444 = if host_wants_444 && client_supports_444 && single_process {
+            tokio::task::spawn_blocking(|| {
+                crate::encode::can_encode_444(crate::encode::Codec::H265)
+            })
+            .await
+            .context("4:4:4 capability probe task")?
+        } else {
+            false
+        };
+        let chroma = if gpu_supports_444 {
+            crate::encode::ChromaFormat::Yuv444
+        } else {
+            crate::encode::ChromaFormat::Yuv420
+        };
+        tracing::info!(
+            chroma = ?chroma,
+            host_wants_444,
+            client_supports_444,
+            single_process,
+            "encode chroma"
+        );
+
         // Reserve a UDP port for the data plane (bind, read it back, rebind in UdpTransport).
         let probe = std::net::UdpSocket::bind("0.0.0.0:0")?;
         let udp_port = probe.local_addr()?.port();
@@ -691,6 +749,12 @@ async fn serve_session(
             } else {
                 ColorInfo::SDR_BT709
             },
+            // The chroma the encoder will actually emit (resolved + GPU-probed above) — 4:4:4 only
+            // when every gate passed, else 4:2:0. The client sizes its decoder from this.
+            chroma_format: chroma.idc(),
+            // The resolved audio channel count the audio thread will capture + Opus-(multi)stream
+            // encode (2/6/8). The client builds its decoder from this echoed value.
+            audio_channels,
         };
         io::write_msg(&mut send, &welcome.encode()).await?;
 
@@ -884,9 +948,10 @@ async fn serve_session(
         let conn = conn.clone();
         let stop = stop.clone();
         let cap = audio_cap.clone();
+        let channels = welcome.audio_channels;
         std::thread::Builder::new()
             .name("punktfunk1-audio".into())
-            .spawn(move || audio_thread(conn, stop, cap))
+            .spawn(move || audio_thread(conn, stop, cap, channels))
             .map_err(|e| tracing::error!(error = %e, "audio thread spawn failed — session continues without audio"))
             .ok()
     } else {
@@ -946,6 +1011,13 @@ async fn serve_session(
     let launch_for_dp = hello.launch.clone();
     let bitrate_kbps = welcome.bitrate_kbps; // resolved encoder bitrate (Hello clamped, or default)
     let bit_depth = welcome.bit_depth; // resolved encode bit depth (8, or 10 when negotiated)
+                                       // Resolved chroma — derive the typed value back from the wire byte the Welcome carried (so the
+                                       // session uses exactly what the client was told). `Yuv444` only when the handshake gate passed.
+    let chroma = if welcome.chroma_format == punktfunk_core::quic::CHROMA_IDC_444 {
+        crate::encode::ChromaFormat::Yuv444
+    } else {
+        crate::encode::ChromaFormat::Yuv420
+    };
     let stop_stream = stop.clone();
     let fec_target_dp = fec_target.clone(); // data-plane handle to the adaptive-FEC target
     let conn_stream = conn.clone(); // for sending the source's real HDR metadata (0xCE) mid-stream
@@ -1005,6 +1077,7 @@ async fn serve_session(
                         compositor,
                         bitrate_kbps,
                         bit_depth,
+                        chroma,
                         probe_rx,
                         probe_result_tx,
                         fec_target: fec_target_dp,
@@ -1493,33 +1566,88 @@ fn input_thread(
     }
 }
 
-/// The audio thread: desktop capture → Opus (48 kHz stereo, 5 ms, CBR — same tuning as the
-/// GameStream path) → `AUDIO_MAGIC` datagrams. QUIC already encrypts; no extra layer.
-/// The capturer comes from (and returns to) the persistent slot — see [`AudioCapSlot`].
+/// Opus encoder for the native audio plane: a plain stereo encoder (the live-validated,
+/// byte-identical path) or a libopus *multistream* encoder for 5.1/7.1, both behind one
+/// `encode_float`. Surround uses the safe `opus::MSEncoder` (no `audiopus_sys`).
 #[cfg(any(target_os = "linux", target_os = "windows"))]
-fn audio_thread(conn: quinn::Connection, stop: Arc<AtomicBool>, audio_cap: AudioCapSlot) {
-    use crate::audio::{CHANNELS, SAMPLE_RATE};
+enum NativeAudioEnc {
+    Stereo(opus::Encoder),
+    Surround(opus::MSEncoder),
+}
+
+#[cfg(any(target_os = "linux", target_os = "windows"))]
+impl NativeAudioEnc {
+    /// Build the encoder for `channels` (2/6/8), hard-CBR + RESTRICTED_LOWDELAY like the
+    /// GameStream path; bitrate from the shared layout table (stereo keeps the validated 128 kbps).
+    fn new(channels: u8) -> Result<NativeAudioEnc, opus::Error> {
+        if channels == 2 {
+            let mut e = opus::Encoder::new(
+                crate::audio::SAMPLE_RATE,
+                opus::Channels::Stereo,
+                opus::Application::LowDelay,
+            )?;
+            e.set_bitrate(opus::Bitrate::Bits(128_000)).ok();
+            e.set_vbr(false).ok();
+            Ok(NativeAudioEnc::Stereo(e))
+        } else {
+            let l = punktfunk_core::audio::layout_for(channels, false);
+            let mut e = opus::MSEncoder::new(
+                crate::audio::SAMPLE_RATE,
+                l.streams,
+                l.coupled,
+                l.mapping,
+                opus::Application::LowDelay,
+            )?;
+            e.set_bitrate(opus::Bitrate::Bits(l.bitrate)).ok();
+            e.set_vbr(false).ok();
+            Ok(NativeAudioEnc::Surround(e))
+        }
+    }
+
+    fn encode_float(&mut self, frame: &[f32], out: &mut [u8]) -> Result<usize, opus::Error> {
+        match self {
+            NativeAudioEnc::Stereo(e) => e.encode_float(frame, out),
+            NativeAudioEnc::Surround(e) => e.encode_float(frame, out),
+        }
+    }
+}
+
+/// The audio thread: desktop capture → Opus (48 kHz, 5 ms, CBR — same tuning as the GameStream
+/// path) → `AUDIO_MAGIC` datagrams, at the negotiated `channels` (2 stereo / 6 = 5.1 / 8 = 7.1,
+/// canonical wire order FL FR FC LFE RL RR SL SR). QUIC already encrypts; no extra layer. The
+/// capturer comes from (and returns to) the persistent slot — see [`AudioCapSlot`].
+#[cfg(any(target_os = "linux", target_os = "windows"))]
+fn audio_thread(
+    conn: quinn::Connection,
+    stop: Arc<AtomicBool>,
+    audio_cap: AudioCapSlot,
+    channels: u8,
+) {
+    use crate::audio::SAMPLE_RATE;
     const FRAME_MS: usize = 5;
     const SAMPLES_PER_FRAME: usize = SAMPLE_RATE as usize * FRAME_MS / 1000; // 240
+    let want = punktfunk_core::audio::normalize_channels(channels);
 
+    // Reuse the cached capturer ONLY when its channel count matches this session's; a stereo
+    // capturer left by a prior session must not feed a 5.1/7.1 session (the encoder + the client's
+    // decoder are sized for `want`, so a mismatched capturer would garble/desync the audio).
     let capturer = match audio_cap.lock().unwrap().take() {
-        Some(mut c) => {
+        Some(mut c) if c.channels() == want as u32 => {
             c.drain(); // discard audio captured between sessions
             c
         }
-        None => match crate::audio::open_audio_capture(CHANNELS as u32) {
-            Ok(c) => c,
-            Err(e) => {
-                tracing::warn!(error = %format!("{e:#}"), "punktfunk/1 audio unavailable — session continues without it");
-                return;
+        prev => {
+            drop(prev); // wrong channel count (or none): clean teardown, open fresh at `want`
+            match crate::audio::open_audio_capture(want as u32) {
+                Ok(c) => c,
+                Err(e) => {
+                    tracing::warn!(error = %format!("{e:#}"), "punktfunk/1 audio unavailable — session continues without it");
+                    return;
+                }
             }
-        },
+        }
     };
-    let mut enc = match opus::Encoder::new(
-        SAMPLE_RATE,
-        opus::Channels::Stereo,
-        opus::Application::LowDelay,
-    ) {
+    let mut enc = match NativeAudioEnc::new(want) {
         Ok(e) => e,
         Err(e) => {
             tracing::error!(error = %e, "opus encoder");
@@ -1527,12 +1655,11 @@ fn audio_thread(conn: quinn::Connection, stop: Arc<AtomicBool>, audio_cap: Audio
             return;
         }
     };
-    enc.set_bitrate(opus::Bitrate::Bits(128_000)).ok();
-    enc.set_vbr(false).ok();
 
-    let frame_len = SAMPLES_PER_FRAME * CHANNELS;
+    let frame_len = SAMPLES_PER_FRAME * want as usize;
     let mut acc: Vec<f32> = Vec::with_capacity(frame_len * 4);
-    let mut opus_buf = vec![0u8; 1500];
+    // Sized for the largest surround frame (7.1 HQ ≈ 1.3 KB at 5 ms); ample for normal quality.
+    let mut opus_buf = vec![0u8; 4096];
     let mut seq: u32 = 0;
     // Reopen-with-backoff: hold the capturer in an Option so a mid-session capture-thread death
     // (device unplug, daemon restart) reopens instead of muting the rest of a multi-hour session.
@@ -1542,14 +1669,17 @@ fn audio_thread(conn: quinn::Connection, stop: Arc<AtomicBool>, audio_cap: Audio
     // restart). The first open already happened above; failing THAT still ends the session quietly.
     let mut capturer = Some(capturer);
     let mut last_failed: Option<std::time::Instant> = None;
-    tracing::info!("punktfunk/1 audio streaming (Opus 48 kHz stereo, 5 ms datagrams)");
+    tracing::info!(
+        channels = want,
+        "punktfunk/1 audio streaming (Opus 48 kHz, 5 ms datagrams)"
+    );
     'session: while !stop.load(Ordering::SeqCst) {
         if capturer.is_none() {
             if last_failed.is_some_and(|t| t.elapsed() < INJECTOR_REOPEN_BACKOFF) {
                 std::thread::sleep(std::time::Duration::from_millis(200));
                 continue;
             }
-            match crate::audio::open_audio_capture(CHANNELS as u32) {
+            match crate::audio::open_audio_capture(want as u32) {
                 Ok(c) => {
                     tracing::info!("punktfunk/1 audio capture reopened");
                     capturer = Some(c);
@@ -1599,7 +1729,12 @@ fn audio_thread(conn: quinn::Connection, stop: Arc<AtomicBool>, audio_cap: Audio
 /// Stub — punktfunk/1 audio needs Linux (PipeWire capture + libopus); non-Linux dev builds
 /// run sessions without it, same as when the capturer fails to open.
 #[cfg(not(any(target_os = "linux", target_os = "windows")))]
-fn audio_thread(_conn: quinn::Connection, _stop: Arc<AtomicBool>, _audio_cap: AudioCapSlot) {
+fn audio_thread(
+    _conn: quinn::Connection,
+    _stop: Arc<AtomicBool>,
+    _audio_cap: AudioCapSlot,
+    _channels: u8,
+) {
     tracing::warn!("punktfunk/1 audio requires Linux or Windows — session continues without it");
 }
 
@@ -2256,6 +2391,45 @@ struct SessionSwitch {
 /// read (so no handshake plumbing). Opt-in via `PUNKTFUNK_SESSION_WATCH`; readiness of the new
 /// backend is left to the encode thread's `build_pipeline_with_retry` (the watcher never writes
 /// env). Exits when `stop` is set or the channel closes.
+/// Whether to run the mid-stream session-switch watcher. An explicit `PUNKTFUNK_SESSION_WATCH` wins
+/// (truthy → on; `0`/`false`/`no`/`off`/empty → off). When unset it defaults **on** for Steam HTPC
+/// platforms (Bazzite / SteamOS) — which flip Gaming↔Desktop and need the host to follow the switch
+/// mid-stream — and **off** elsewhere, preserving the opt-in default for plain desktop hosts.
+fn session_watch_enabled() -> bool {
+    match std::env::var("PUNKTFUNK_SESSION_WATCH") {
+        Ok(v) => {
+            let v = v.trim();
+            !(v.is_empty()
+                || v == "0"
+                || v.eq_ignore_ascii_case("false")
+                || v.eq_ignore_ascii_case("no")
+                || v.eq_ignore_ascii_case("off"))
+        }
+        Err(_) => is_steam_htpc_platform(),
+    }
+}
+
+/// True on Bazzite or SteamOS (matched against os-release `ID`/`ID_LIKE`) — the platforms that flip
+/// between Steam Gaming Mode and a Desktop session, where following a mid-stream switch is the
+/// sensible default. Anything else (incl. non-Linux, where the file is absent) → false.
+fn is_steam_htpc_platform() -> bool {
+    let Ok(os) = std::fs::read_to_string("/etc/os-release") else {
+        return false;
+    };
+    os.lines().any(|line| {
+        let line = line.trim();
+        let Some(val) = line
+            .strip_prefix("ID=")
+            .or_else(|| line.strip_prefix("ID_LIKE="))
+        else {
+            return false;
+        };
+        val.trim_matches('"')
+            .split_whitespace()
+            .any(|tok| tok.eq_ignore_ascii_case("bazzite") || tok.eq_ignore_ascii_case("steamos"))
+    })
+}
+
 fn session_watcher_loop(tx: std::sync::mpsc::Sender<SessionSwitch>, stop: Arc<AtomicBool>) {
     use crate::vdisplay;
     const DEBOUNCE: std::time::Duration = std::time::Duration::from_secs(3);
@@ -2329,6 +2503,8 @@ struct SessionContext {
     bitrate_kbps: u32,
     /// Negotiated encode bit depth (8, or 10 = HEVC Main10).
     bit_depth: u8,
+    /// Negotiated chroma subsampling (4:2:0, or 4:4:4 when the client + host + GPU all support it).
+    chroma: crate::encode::ChromaFormat,
     /// Speed-test burst requests (see [`service_probes`]).
     probe_rx: std::sync::mpsc::Receiver<ProbeRequest>,
     /// Speed-test results back to the control task.
@@ -2359,7 +2535,7 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
     // path now reads this typed `SessionPlan` instead of re-deriving from config at each dispatch site
     // (the latent "capture and encode disagree on the backend" hazard, plan §2.4). `bit_depth` is the
     // only per-session input — capture/topology/encoder are otherwise pure functions of `HostConfig`.
-    let plan = crate::session_plan::SessionPlan::resolve(ctx.bit_depth);
+    let plan = crate::session_plan::SessionPlan::resolve(ctx.bit_depth, ctx.chroma);
     tracing::info!(?plan, "resolved session plan");
     // Windows two-process secure-desktop path: when the host runs as SYSTEM (required for the secure
     // desktop + SendInput), WGC can't activate in-process, so we capture the normal desktop via a
@@ -2381,6 +2557,8 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
         compositor,
         bitrate_kbps,
         bit_depth,
+        // The resolved chroma is already captured in `plan` (above); ignore the duplicate here.
+        chroma: _,
         probe_rx,
         probe_result_tx,
         fec_target,
@@ -2491,9 +2669,9 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
     // place when the box flips Gaming↔Desktop. When not spawned, session_rx just stays empty.
     let mut compositor = compositor;
     let (session_tx, session_rx) = std::sync::mpsc::channel::<SessionSwitch>();
-    let watch = std::env::var_os("PUNKTFUNK_SESSION_WATCH").is_some()
-        && crate::config::config().compositor.is_none();
+    let watch = session_watch_enabled() && crate::config::config().compositor.is_none();
     let _watcher = if watch {
+        tracing::info!("session watcher on — following a mid-stream Gaming↔Desktop switch");
         let stop = stop.clone();
         std::thread::Builder::new()
             .name("punktfunk1-watcher".into())
@@ -2675,15 +2853,76 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
                 }
                 tracing::warn!(error = %format!("{e:#}"), rebuild = capture_rebuilds,
                     "capture lost — rebuilding pipeline in place");
-                let (new_cap, new_enc, new_frame, new_interval) =
-                    build_pipeline_with_retry(&mut vd, cur_mode, bitrate_kbps, bit_depth, plan)
-                        .context("rebuild after capture loss")?;
+                // A Bazzite/SteamOS Gaming↔Desktop switch tears the old compositor down and can take
+                // 15s+ to bring the new one up. Don't fail the session over that (the client would
+                // have to cold-reconnect, surfacing a "session failed") — keep retrying within a
+                // generous budget while the QUIC keepalive (its own thread) holds the connection,
+                // RE-DETECTING the live compositor each attempt so we follow the box to whatever
+                // session comes up: a fresh instance of the same compositor, OR a different one
+                // (the kind-change case the session watcher also handles). The client stays
+                // connected, frozen on the last frame, and the stream resumes when the new output
+                // appears — no reconnect.
+                const REBUILD_BUDGET: std::time::Duration = std::time::Duration::from_secs(40);
+                let rebuild_deadline = std::time::Instant::now() + REBUILD_BUDGET;
+                let (new_cap, new_enc, new_frame, new_interval) = loop {
+                    // Follow the active session unless an explicit PUNKTFUNK_COMPOSITOR pin forbids
+                    // retargeting (then we stick to the pinned backend and just rebuild it).
+                    if crate::config::config().compositor.is_none() {
+                        let active = crate::vdisplay::detect_active_session();
+                        if let Some(c) = crate::vdisplay::compositor_for_kind(active.kind) {
+                            crate::vdisplay::apply_session_env(&active);
+                            crate::vdisplay::apply_input_env(c);
+                            if c != compositor {
+                                if matches!(
+                                    c,
+                                    crate::vdisplay::Compositor::Kwin
+                                        | crate::vdisplay::Compositor::Mutter
+                                ) {
+                                    crate::vdisplay::settle_desktop_portal(c);
+                                }
+                                match crate::vdisplay::open(c) {
+                                    Ok(v) => {
+                                        tracing::info!(from = compositor.id(), to = c.id(),
+                                            "capture loss: active session switched compositor — retargeting");
+                                        vd = v;
+                                        compositor = c;
+                                    }
+                                    Err(e2) => tracing::warn!(error = %format!("{e2:#}"),
+                                        "capture loss: opening the newly-detected compositor failed — retrying"),
+                                }
+                            }
+                        }
+                    }
+                    match build_pipeline_with_retry(
+                        &mut vd,
+                        cur_mode,
+                        bitrate_kbps,
+                        bit_depth,
+                        plan,
+                    ) {
+                        Ok(p) => break p,
+                        Err(e2) => {
+                            if stop.load(Ordering::SeqCst)
+                                || std::time::Instant::now() >= rebuild_deadline
+                            {
+                                return Err(e2)
+                                    .context("capture lost — no compositor came up within the rebuild budget");
+                            }
+                            tracing::warn!(error = %format!("{e2:#}"),
+                                "capture lost — new session not up yet, retrying");
+                        }
+                    }
+                };
                 capturer = new_cap;
                 enc = new_enc;
                 frame = new_frame;
                 interval = new_interval;
                 enc.request_keyframe(); // belt-and-suspenders; a fresh encoder opens on an IDR anyway
                 next = std::time::Instant::now();
+                tracing::info!(
+                    compositor = compositor.id(),
+                    "capture loss: pipeline rebuilt — stream resumes"
+                );
             }
         }
         if perf && diag_at.elapsed() >= std::time::Duration::from_secs(2) {
@@ -2869,6 +3108,9 @@ fn virtual_stream_relay(ctx: SessionContext) -> Result<()> {
         compositor,
         bitrate_kbps,
         bit_depth,
+        // The two-process WGC relay encodes 4:2:0 in v1 — the handshake's `single_process` gate already
+        // forced `chroma` to Yuv420 for this topology, so the helper + secure-desktop DDA stay 4:2:0.
+        chroma: _,
         probe_rx,
         probe_result_tx,
         fec_target,
@@ -2979,6 +3221,7 @@ fn virtual_stream_relay(ctx: SessionContext) -> Result<()> {
                 // stage 5) so the DDA capturer doesn't re-derive it.
                 crate::capture::gpu_encode(),
                 hdr,
+                false, // the two-process relay path is 4:2:0 in v1
             )
             .context("open DDA for secure desktop")?;
             cap.set_active(true);
@@ -2992,6 +3235,8 @@ fn virtual_stream_relay(ctx: SessionContext) -> Result<()> {
                 bitrate_kbps as u64 * 1000,
                 frame.is_cuda(),
                 bit_depth,
+                // Secure-desktop DDA on the two-process relay path: 4:2:0 in v1 (matches the helper).
+                crate::encode::ChromaFormat::Yuv420,
             )
             .context("open video encoder for DDA")?;
             Ok(DdaPipe {
@@ -3391,6 +3636,9 @@ fn is_permanent_build_error(chain: &str) -> bool {
         "could not find output", // KWin < 6.5.6: createVirtualOutput unsupported
         "must be a node id",     // PUNKTFUNK_GAMESCOPE_NODE not an integer
         "is it installed",       // gamescope / kscreen-doctor not on PATH
+        // 4:4:4 NVENC got a CUDA frame — should never happen now the Linux capturer honors gpu=false,
+        // but fail fast instead of 8× retry (~90 s) rather than wedge the session if it ever recurs.
+        "capture/encoder negotiation mismatch",
     ];
     let lower = chain.to_ascii_lowercase();
     PERMANENT.iter().any(|p| lower.contains(p))
@@ -3440,8 +3688,20 @@ fn build_pipeline(
         bitrate_kbps as u64 * 1000,
         frame.is_cuda(),
         bit_depth,
+        plan.chroma,
     )
     .context("open video encoder")?;
+    // Post-open cross-check: the Welcome already committed `chroma_format` from the pre-open probe, so
+    // warn loudly if the encoder actually opened a different chroma than negotiated (the in-band SPS is
+    // authoritative for the decoder, but a mismatch means the probe and the live open disagreed).
+    let opened_444 = enc.caps().chroma_444;
+    if opened_444 != plan.chroma.is_444() {
+        tracing::warn!(
+            negotiated_444 = plan.chroma.is_444(),
+            opened_444,
+            "encoder chroma disagrees with the negotiated Welcome — the client was told the other value"
+        );
+    }
     let interval = std::time::Duration::from_secs_f64(1.0 / effective_hz.max(1) as f64);
     Ok((capturer, enc, frame, interval))
 }
@@ -3880,6 +4140,7 @@ mod tests {
                 GamepadPref::Auto,
                 0,
                 0,    // video_caps
+                2,    // audio_channels (stereo)
                 None, // launch
                 None,
                 Some((cert.clone(), key.clone())),
@@ -3912,6 +4173,7 @@ mod tests {
             GamepadPref::Auto,
             0,
             0,    // video_caps
+            2,    // audio_channels (stereo)
             None, // launch
             None,
             Some((cert, key)),
@@ -3965,6 +4227,7 @@ mod tests {
                 GamepadPref::Auto,
                 0,
                 0,    // video_caps
+                2,    // audio_channels (stereo)
                 None, // launch
                 None,
                 None,
@@ -3990,6 +4253,7 @@ mod tests {
             GamepadPref::Auto,
             0,
             0,    // video_caps
+            2,    // audio_channels (stereo)
             None, // launch
             Some(host_fp),
             Some((cert.clone(), key.clone())),

crates/punktfunk-host/src/session_plan.rs

@@ -106,17 +106,22 @@ pub struct SessionPlan {
     /// The IDD-push HDR hint (`bit_depth >= 10`) — the want-HDR flag the capturer was passed before.
     /// Non-IDD-push Windows backends ignore it and auto-detect HDR from the monitor; Linux is 8-bit.
     pub hdr: bool,
+    /// Handshake-negotiated chroma subsampling (4:2:0, or full-chroma 4:4:4 when the client + host +
+    /// GPU all support it). Resolved before the Welcome; `Yuv420` on every backend that declined it.
+    pub chroma: crate::encode::ChromaFormat,
 }
 
 impl SessionPlan {
-    /// Resolve the whole plan once from [`config`](crate::config) + the negotiated `bit_depth`.
-    pub fn resolve(bit_depth: u8) -> Self {
+    /// Resolve the whole plan once from [`config`](crate::config) + the negotiated `bit_depth` and
+    /// `chroma`.
+    pub fn resolve(bit_depth: u8, chroma: crate::encode::ChromaFormat) -> Self {
         SessionPlan {
             capture: CaptureBackend::resolve(),
             topology: resolve_topology(),
             encoder: resolve_encoder(),
             bit_depth,
             hdr: bit_depth >= 10,
+            chroma,
         }
     }
 
@@ -124,9 +129,24 @@ impl SessionPlan {
     /// (no second backend probe), `hdr` from the plan. Handed into `capture::capture_virtual_output` so the
     /// capturer never re-derives the encode backend.
     pub fn output_format(&self) -> crate::capture::OutputFormat {
+        let gpu = self.encoder.is_gpu();
+        // Linux NVENC 4:4:4: libavcodec `hevc_nvenc` only emits 4:4:4 from a YUV444 *input* frame —
+        // RGB-in is always subsampled to 4:2:0 (verified on the RTX 5070 Ti). So the encoder does an
+        // RGB→YUV444P swscale and needs CPU-resident RGB frames; force the zero-copy GPU capture off
+        // for a 4:4:4 NVENC session. (VAAPI 4:4:4, where the hardware supports it, keeps its dmabuf
+        // path via `scale_vaapi`; Windows NVENC ingests ARGB directly and stays GPU.)
+        #[cfg(target_os = "linux")]
+        let gpu = {
+            let force_cpu_for_nvenc_444 =
+                self.chroma.is_444() && !crate::encode::linux_zero_copy_is_vaapi();
+            gpu && !force_cpu_for_nvenc_444
+        };
         crate::capture::OutputFormat {
-            gpu: self.encoder.is_gpu(),
+            gpu,
             hdr: self.hdr,
+            // 4:4:4 needs a full-chroma source: on Windows this keeps the capturer on RGB (not the
+            // default NV12/P010 video-engine output) so NVENC can CSC to 4:4:4.
+            chroma_444: self.chroma.is_444(),
         }
     }
 }
@@ -134,7 +154,7 @@ impl SessionPlan {
 /// Process topology. On Windows this is the former `punktfunk1::should_use_helper` logic verbatim; on
 /// every other platform the session is always single-process.
 #[cfg(target_os = "windows")]
-fn resolve_topology() -> SessionTopology {
+pub(crate) fn resolve_topology() -> SessionTopology {
     let cfg = crate::config::config();
     // `NO_HELPER`/`NO_WGC` force single-process; IDD-push captures in-process in Session 0 (no helper);
     // otherwise the helper runs when forced or when we're SYSTEM (in-process WGC can't activate there).
@@ -151,7 +171,7 @@ fn resolve_topology() -> SessionTopology {
 }
 
 #[cfg(not(target_os = "windows"))]
-fn resolve_topology() -> SessionTopology {
+pub(crate) fn resolve_topology() -> SessionTopology {
     SessionTopology::SingleProcess
 }
 

crates/punktfunk-host/src/spike.rs

@@ -109,7 +109,8 @@ pub fn run(opts: Options) -> Result<()> {
         opts.fps,
         opts.bitrate_bps,
         first.is_cuda(),
-        8, // spike synthetic harness: 8-bit
+        8,                            // spike synthetic harness: 8-bit
+        encode::ChromaFormat::Yuv420, // ...and 4:2:0
     )
     .context("open encoder")?;
 

crates/punktfunk-host/src/vdisplay.rs

@@ -457,7 +457,11 @@ pub fn settle_desktop_portal(_chosen: Compositor) {}
 pub fn apply_input_env(chosen: Compositor) {
     let backend = match chosen {
         Compositor::Gamescope => "gamescope",
-        Compositor::Kwin | Compositor::Mutter => "libei",
+        // KWin: org_kde_kwin_fake_input — direct injection, no RemoteDesktop portal / approval
+        // dialog (headless, the krdpserver path), authorized by the host's shipped .desktop.
+        Compositor::Kwin => "kwin",
+        // GNOME has neither fake_input nor the wlr protocols → RemoteDesktop portal via libei.
+        Compositor::Mutter => "libei",
         Compositor::Wlroots => "wlr",
     };
     std::env::set_var("PUNKTFUNK_INPUT_BACKEND", backend);

crates/punktfunk-host/src/vdisplay/linux/gamescope.rs

@@ -15,7 +15,7 @@
 //! `inject/libei.rs`) — wired and live-validated.
 
 use super::{Mode, VirtualDisplay, VirtualOutput};
-use anyhow::{anyhow, Context, Result};
+use anyhow::{anyhow, bail, Context, Result};
 use std::process::{Child, Command, Stdio};
 use std::time::{Duration, Instant};
 
@@ -110,12 +110,11 @@ impl VirtualDisplay for GamescopeDisplay {
         // PUNKTFUNK_GAMESCOPE_NODE=<id|auto>; "auto" discovers the gamescope `Video/Source` node.
         if let Ok(id) = std::env::var("PUNKTFUNK_GAMESCOPE_NODE") {
             let node_id: u32 = if id.trim().eq_ignore_ascii_case("auto") {
-                find_gamescope_node().ok_or_else(|| {
-                    anyhow!(
-                        "PUNKTFUNK_GAMESCOPE_NODE=auto but no running gamescope Video/Source node \
-                         was found — is the headless gamescope/Steam session up?"
-                    )
-                })?
+                // Attach to the box-owned game-mode session, but FIRST make it run at the connecting
+                // client's resolution (the box is headless, so its game-mode mode is ours to set).
+                // Reuse if it already matches (fast, no restart); otherwise relaunch the box's own
+                // session at the client mode. Without this the client gets the box's default mode.
+                ensure_box_gamescope_mode(mode)?
             } else {
                 id.parse()
                     .context("PUNKTFUNK_GAMESCOPE_NODE must be a node id or 'auto'")?
@@ -368,6 +367,150 @@ fn create_managed_session_steamos(mode: Mode) -> Result<VirtualOutput> {
     })
 }
 
+/// ATTACH at the CLIENT's resolution: ensure the box's own game-mode session is running at `mode`'s
+/// output size, then return its capture node. Reuses the running session if it already matches (no
+/// restart — the rock-solid fast path a stable client always hits); otherwise reconfigures + restarts
+/// the box's OWN autologin `gamescope-session-plus@<client>` unit at the client mode. Restarting the
+/// box's own unit (rather than spawning a competing one) avoids the autologin-respawn fight the old
+/// MANAGED path hit. A headless box has no physical panel, so its game-mode resolution is ours to set;
+/// Steam restarts only on an actual resolution CHANGE.
+fn ensure_box_gamescope_mode(mode: Mode) -> Result<u32> {
+    let target = (mode.width, mode.height);
+    // Fast path: already at the client's resolution — just attach to the live node.
+    if current_gamescope_output_size() == Some(target) {
+        if let Some(node) = find_gamescope_node() {
+            tracing::info!(
+                w = mode.width,
+                h = mode.height,
+                node,
+                "gamescope: box game-mode session already at the client's resolution — reusing"
+            );
+            return Ok(node);
+        }
+    }
+    let Some(unit) = running_autologin_gamescope_unit() else {
+        // No box-owned autologin session to reconfigure (a bare/foreign gamescope): attach to
+        // whatever node exists, accepting its resolution.
+        return find_gamescope_node().ok_or_else(|| {
+            anyhow!(
+                "no running gamescope Video/Source node — is the headless game mode up? \
+                 (put the box into Steam Game Mode)"
+            )
+        });
+    };
+    tracing::info!(
+        from = ?current_gamescope_output_size(),
+        to_w = mode.width,
+        to_h = mode.height,
+        hz = mode.refresh_hz,
+        %unit,
+        "gamescope: relaunching the box game-mode session at the client's resolution"
+    );
+    // The session reads SCREEN_WIDTH/HEIGHT (+ CUSTOM_REFRESH_RATES) from the user-manager
+    // environment; set them and restart the box's own unit.
+    systemctl_user(&[
+        "set-environment",
+        &format!("SCREEN_WIDTH={}", mode.width),
+        &format!("SCREEN_HEIGHT={}", mode.height),
+        &format!("CUSTOM_REFRESH_RATES={}", mode.refresh_hz.max(1)),
+    ]);
+    systemctl_user(&["restart", &unit]);
+    // Wait for the relaunched session to come up at the new size and publish its capture node. The
+    // node appears when gamescope is up (well before Steam finishes booting); the caller's
+    // first-frame retry absorbs Steam's cold start.
+    let deadline = Instant::now() + Duration::from_secs(45);
+    loop {
+        if current_gamescope_output_size() == Some(target) {
+            if let Some(node) = find_gamescope_node() {
+                tracing::info!(
+                    node,
+                    w = mode.width,
+                    h = mode.height,
+                    "gamescope: box game-mode session relaunched at the client's resolution"
+                );
+                return Ok(node);
+            }
+        }
+        if Instant::now() >= deadline {
+            bail!(
+                "box game-mode session did not come up at {}x{} within 45s after relaunch \
+                 (Steam may still be booting)",
+                mode.width,
+                mode.height
+            );
+        }
+        std::thread::sleep(Duration::from_millis(500));
+    }
+}
+
+/// Output (capture) resolution `-W <w> -H <h>` of the running `gamescope` binary, parsed from its
+/// `/proc/<pid>/cmdline`. `None` if no gamescope is running or the flags aren't present.
+fn current_gamescope_output_size() -> Option<(u32, u32)> {
+    for entry in std::fs::read_dir("/proc").ok()?.flatten() {
+        let name = entry.file_name();
+        let Some(pid) = name.to_str() else { continue };
+        if !pid.bytes().all(|b| b.is_ascii_digit()) {
+            continue;
+        }
+        let Ok(raw) = std::fs::read(format!("/proc/{pid}/cmdline")) else {
+            continue;
+        };
+        let args: Vec<String> = raw
+            .split(|&b| b == 0)
+            .filter(|s| !s.is_empty())
+            .map(|s| String::from_utf8_lossy(s).into_owned())
+            .collect();
+        // Match the gamescope BINARY by argv[0]'s basename — NOT /proc/<pid>/exe, which is commonly
+        // unreadable for the gamescope process (returns empty). The session wrapper scripts run as
+        // bash/sh (argv[0] != gamescope), so they're excluded; the -W/-H presence check below is the
+        // final filter.
+        let is_gamescope = args
+            .first()
+            .map(|a0| a0.rsplit('/').next().unwrap_or(a0) == "gamescope")
+            .unwrap_or(false);
+        if !is_gamescope {
+            continue;
+        }
+        let flag = |names: &[&str]| -> Option<u32> {
+            args.iter().enumerate().find_map(|(i, a)| {
+                names
+                    .contains(&a.as_str())
+                    .then(|| args.get(i + 1).and_then(|v| v.parse().ok()))
+                    .flatten()
+            })
+        };
+        if let (Some(w), Some(h)) = (
+            flag(&["-W", "--output-width"]),
+            flag(&["-H", "--output-height"]),
+        ) {
+            return Some((w, h));
+        }
+    }
+    None
+}
+
+/// The running autologin gaming-mode unit (`gamescope-session-plus@<client>.service`), if any — the
+/// box's own game-mode session, which [`ensure_box_gamescope_mode`] reconfigures + restarts.
+fn running_autologin_gamescope_unit() -> Option<String> {
+    let out = Command::new("systemctl")
+        .args([
+            "--user",
+            "list-units",
+            "--type=service",
+            "--state=running",
+            "--no-legend",
+            "--plain",
+            "gamescope-session-plus@*.service",
+        ])
+        .output()
+        .ok()?;
+    String::from_utf8_lossy(&out.stdout)
+        .lines()
+        .filter_map(|l| l.split_whitespace().next())
+        .find(|u| u.starts_with("gamescope-session-plus@") && u.ends_with(".service"))
+        .map(|u| u.to_string())
+}
+
 /// Stop every running autologin gaming-mode session (`gamescope-session-plus@*.service`) so its
 /// single-instance Steam is free for our own host-managed session. Records the units so
 /// [`schedule_restore_tv_session`] can restart them on disconnect. Our own session is the transient

crates/punktfunk-host/src/vdisplay/linux/kwin.rs

@@ -6,8 +6,14 @@
 //! node for it. The node lives on the user's default PipeWire daemon, so [`VirtualOutput::remote_fd`]
 //! is `None` and capture connects to that daemon directly.
 //!
-//! Requirements: KWin must expose the privileged `zkde_screencast` global — a real Plasma session
-//! authorizes it for its own clients; the headless test exposes it to bare clients via
+//! Requirements: KWin must expose the privileged `zkde_screencast` global. It is a *restricted*
+//! protocol — KWin advertises it only to a client whose installed `.desktop` lists it under
+//! `X-KDE-Wayland-Interfaces` (KWin maps the connecting client to a `.desktop` by resolving
+//! `/proc/<pid>/exe` against `Exec=`, then caches the grant per-executable for the session's life).
+//! So an interactive Plasma session does NOT hand it to a bare client — the host packages ship
+//! `io.unom.Punktfunk.Host.desktop` (`Exec=/usr/bin/punktfunk-host`,
+//! `X-KDE-Wayland-Interfaces=zkde_screencast_unstable_v1,…`) so it is present before the host first
+//! connects. The headless test path instead exposes it to bare clients via
 //! `KWIN_WAYLAND_NO_PERMISSION_CHECKS=1`. The compositor backend must implement
 //! `createVirtualOutput`: the **DRM backend** (any version) or the **VirtualBackend since KWin
 //! 6.5.6** (`kwin_wayland --virtual`); on `--virtual` < 6.5.6 the request fails with
@@ -406,9 +412,11 @@ pub fn probe() -> Result<()> {
     queue.roundtrip(&mut state).context("registry roundtrip")?;
     if state.screencast.is_none() {
         bail!(
-            "KWin is up but does not (yet) expose zkde_screencast_unstable_v1 — needs a real \
-             KDE session (or KWIN_WAYLAND_NO_PERMISSION_CHECKS=1), and KWin ≥ 6.5.6 for the \
-             headless virtual output"
+            "KWin is up but does not expose zkde_screencast_unstable_v1 to this client — KWin gates \
+             it on the host's .desktop X-KDE-Wayland-Interfaces (install \
+             io.unom.Punktfunk.Host.desktop with Exec=/usr/bin/punktfunk-host, then re-login so KWin \
+             re-reads it — the grant is cached per-exe on first connect), or set \
+             KWIN_WAYLAND_NO_PERMISSION_CHECKS=1 for the headless test; needs KWin ≥ 6.5.6"
         );
     }
     Ok(())
@@ -437,8 +445,9 @@ fn run(
 
     let screencast = state.screencast.clone().ok_or_else(|| {
         anyhow!(
-            "KWin does not expose zkde_screencast_unstable_v1 (need a real KDE session, or run \
-             KWin with KWIN_WAYLAND_NO_PERMISSION_CHECKS=1 for the headless test)"
+            "KWin does not expose zkde_screencast_unstable_v1 to this client — install the host's \
+             .desktop (io.unom.Punktfunk.Host.desktop, X-KDE-Wayland-Interfaces) and re-login so \
+             KWin authorizes it, or run KWin with KWIN_WAYLAND_NO_PERMISSION_CHECKS=1 (headless test)"
         )
     })?;
 

crates/punktfunk-host/src/windows/wgc_helper.rs

@@ -98,6 +98,9 @@ pub fn run(opts: HelperOptions) -> Result<()> {
         opts.bitrate_kbps as u64 * 1000,
         false,          // not cuda
         opts.bit_depth, // 8, or 10 = Main10 (HDR auto-upgrades from the Rgb10a2 frame regardless)
+        // The two-process WGC relay helper encodes 4:2:0 in v1 (4:4:4 over the relay is a follow-up);
+        // the host gates 4:4:4 to the single-process topology.
+        encode::ChromaFormat::Yuv420,
     )
     .context("open NVENC")?;
 

docs-site/bun.lock

@@ -9,6 +9,7 @@
         "@scalar/api-reference-react": "^0.9.47",
         "@tanstack/react-router": "^1.121.0",
         "@tanstack/react-start": "^1.121.0",
+        "@unom/app-ui": "^0.1.0",
         "@unom/style": "^0.4.4",
         "@unom/ui": "^0.8.16",
         "fumadocs-core": "^16.10.5",
@@ -231,6 +232,8 @@
 
     "@headlessui/vue": ["@headlessui/vue@1.7.23", "", { "dependencies": { "@tanstack/vue-virtual": "^3.0.0-beta.60" }, "peerDependencies": { "vue": "^3.2.0" } }, "sha512-JzdCNqurrtuu0YW6QaDtR2PIYCKPUWq28csDyMvN4zmGccmE7lz40Is6hc3LA4HFeCI7sekZ/PQMTNmn9I/4Wg=="],
 
+    "@icons-pack/react-simple-icons": ["@icons-pack/react-simple-icons@13.13.0", "", { "peerDependencies": { "react": "^16.13 || ^17 || ^18 || ^19" } }, "sha512-B5HhQMIpcSH4z8IZ8HFhD59CboHceKYMpPC9kAwGyKntvPdyJJv26DLu4Z1wAjcCLyrJhf11tMhiQGom9Rxb9g=="],
+
     "@img/colour": ["@img/colour@1.1.0", "", {}, "sha512-Td76q7j57o/tLVdgS746cYARfSyxk8iEfRxewL9h4OMzYhbW4TAcppl0mT4eyqXddh6L/jwoM75mo7ixa/pCeQ=="],
 
     "@img/sharp-darwin-arm64": ["@img/sharp-darwin-arm64@0.34.5", "", { "optionalDependencies": { "@img/sharp-libvips-darwin-arm64": "1.2.4" }, "os": "darwin", "cpu": "arm64" }, "sha512-imtQ3WMJXbMY4fxb/Ndp6HBTNVtWCUI0WdobyheGf5+ad6xX8VIDO8u2xE4qc/fr08CKG/7dDseFtn6M6g/r3w=="],
@@ -909,6 +912,8 @@
 
     "@unhead/vue": ["@unhead/vue@2.1.15", "", { "dependencies": { "hookable": "^6.0.1", "unhead": "2.1.15" }, "peerDependencies": { "vue": ">=3.5.18" } }, "sha512-SSByXfEjhzPn8gXdEdgpYqpLMPSkLUH2HVE0GxZfOtNsJ0GgOHQs0g9T67ZZ1z0kTELLKdtOtYrzrbv9+ffF7g=="],
 
+    "@unom/app-ui": ["@unom/app-ui@0.1.0", "https://git.unom.io/api/packages/unom/npm/%40unom%2Fapp-ui/-/0.1.0/app-ui-0.1.0.tgz", { "dependencies": { "@icons-pack/react-simple-icons": "^13.13.0" }, "peerDependencies": { "@unom/style": "^0.4.4", "react": "^19.0.0" } }, "sha512-znHZOIRWyJDj4va2X/E4GwvxWZsVeWEYpvu7iHTBIa0UXjkX9aoiujJcMyfPpc2Vof53iafl9hIszgSgjQwzhg=="],
+
     "@unom/style": ["@unom/style@0.4.4", "https://git.unom.io/api/packages/unom/npm/%40unom%2Fstyle/-/0.4.4/style-0.4.4.tgz", { "peerDependencies": { "motion": "^12" } }, "sha512-M45nihK+LGyxwy2mmHYRKggaocTt+EKNVFNaMpTvTaIUpozi7bmKIkbM2/enMYS0/UYTaZrBSZs/a0nPXqkAKw=="],
 
     "@unom/ui": ["@unom/ui@0.8.16", "https://git.unom.io/api/packages/unom/npm/%40unom%2Fui/-/0.8.16/ui-0.8.16.tgz", { "dependencies": { "@tanstack/react-router": "^1.170.11", "@tsdown/css": "^0.22.1", "clsx": "^2.1.1", "howler": "^2.2.4", "sonner": "^2.0.7", "tailwind-merge": "^3.6.0" }, "peerDependencies": { "@payloadcms/richtext-lexical": "^3.85.0", "@tanstack/react-virtual": "^3.14.2", "@unom/style": "^0.4.4", "class-variance-authority": "^0.7.1", "embla-carousel-react": "^8.6.0", "lucide-react": "^1.17.0", "motion": "^12.40.0", "radix-ui": "^1.4.3", "react": "^19.2.7", "react-dom": "^19.2.7", "typescript": "^6.0.3", "zod": "^4.4.3" } }, "sha512-ZH7VOyaRDT81VY8nm1hmx8a4CeObykP8egZbnV4Nju6kE8rQ28wdpBo0X+Zsdu8WvTEmHZGwPR53NHWJULyciw=="],

docs-site/content/docs/channels.md

@@ -44,9 +44,10 @@ one-line edit of `/etc/apt/sources.list.d/punktfunk.list` (`stable` ↔ `canary`
 
 1. Make sure `main` is green.
 2. (Optional) bump any user-facing version that isn't derived from the tag — the Android
-   `versionName` fallback (`clients/android/app/build.gradle.kts`) and the Decky `plugin.json`
-   `version` are cosmetic self-reported strings; everything else (binaries via
-   `PUNKTFUNK_BUILD_VERSION`, MSIX, apt/rpm, the `.dmg`) derives from the tag automatically.
+   `versionName` fallback (`clients/android/app/build.gradle.kts`) is a cosmetic self-reported
+   string; everything else (binaries via `PUNKTFUNK_BUILD_VERSION`, MSIX, apt/rpm, the `.dmg`, and
+   the **Decky** plugin version — CI stamps it into `package.json`, where it drives the plugin's own
+   [self-update check](/docs/steam-deck#updating)) derives from the tag automatically.
 3. Tag and push — **one** tag releases every platform:
    ```sh
    git tag v0.2.0

docs-site/content/docs/clients.md

@@ -44,7 +44,7 @@ It ships as a real package, not just a source build — full steps in
 
 - **Any Flatpak distro (recommended)** — `flatpak install https://flatpak.unom.io/io.unom.Punktfunk.flatpakref`
   from the hosted [`flatpak.unom.io`](/docs/install-client#linux-desktop-flatpak) repo, then
-  `flatpak update`; this is also what the Decky plugin launches.
+  `flatpak update`; this is also what the [Decky plugin](/docs/steam-deck) launches.
 - **Ubuntu / Debian** — `apt install punktfunk-client` from the punktfunk apt registry.
 - **Fedora / Bazzite** — `rpm-ostree install punktfunk-client` from the Gitea RPM registry.
 - **Arch / SteamOS** — the `punktfunk-client` split package from the `PKGBUILD`.
@@ -108,7 +108,8 @@ punktfunk-probe --connect <host>:9777 --pin <fp>  # connect to one
 | You're streaming to… | Use |
 |---|---|
 | A Mac, iPhone, iPad, or Apple TV | The **Apple app** |
-| A Linux desktop or laptop, or a Steam Deck | **`punktfunk-client`** (GTK4) |
+| A Linux desktop or laptop | **`punktfunk-client`** (GTK4) |
+| A **Steam Deck** | The **[Decky plugin](/docs/steam-deck)** in Gaming Mode, or the GTK4 client in Desktop Mode |
 | An Android phone or TV | The **Android app** |
 | Windows | The native **`punktfunk-client`** (signed MSIX) or **Moonlight** |
 | A browser, a smart TV, or any other device | **Moonlight** |

docs-site/content/docs/install-client.md

@@ -16,7 +16,7 @@ Whichever client you install, the first connection needs a one-time [pairing](/d
 | Device | Install |
 |--------|---------|
 | **Linux** desktop / laptop | [Flatpak](#linux-desktop-flatpak) (any distro) or native apt/rpm/Arch packages |
-| **Steam Deck** | [Flatpak in Desktop Mode](#steam-deck) (or the Decky plugin) |
+| **Steam Deck** | [Decky plugin](/docs/steam-deck) for Gaming Mode, or [Flatpak in Desktop Mode](#steam-deck) |
 | **Windows** | [Signed MSIX](#windows) from the package registry |
 | **macOS** | [Notarized `.dmg`](#macos) from the releases page |
 | **iPhone / iPad / Apple TV** | [TestFlight beta](#ios-ipados-apple-tv) |
@@ -57,16 +57,23 @@ punktfunk-client --connect <host>:9777
 
 ## Steam Deck
 
-In **Desktop Mode**, install the Flatpak exactly as [above](#linux-desktop-flatpak) — it carries
-its own libadwaita + SDL3 and survives SteamOS updates:
+Most Deck users want **Gaming Mode**: install the **[Decky plugin](/docs/steam-deck)** and a
+**punktfunk** panel lands in the Quick Access Menu, so you can discover hosts, pair with a PIN, and
+stream **without dropping to the desktop**. Follow the **[Steam Deck (Decky) guide](/docs/steam-deck)**
+— it walks through Decky Loader, the plugin, and the one-time client install.
+
+> The plugin doesn't decode video itself — it launches the Flatpak client below. The Decky guide
+> covers installing both, so start there: a Flatpak on its own won't add the Gaming Mode panel.
+
+For **Desktop Mode** (or to add the client to Game Mode as a non-Steam app yourself), install the
+Flatpak exactly as [above](#linux-desktop-flatpak) — it carries its own libadwaita + SDL3 and
+survives SteamOS updates:
 
 ```sh
 flatpak install --user https://flatpak.unom.io/io.unom.Punktfunk.flatpakref
 ```
 
-Add it to Game Mode as a non-Steam app, or use the **Decky plugin**, which launches this same
-Flatpak (`flatpak run io.unom.Punktfunk --connect …`). See
-[packaging/flatpak](https://git.unom.io/unom/punktfunk/src/branch/main/packaging/flatpak/README.md).
+See [packaging/flatpak](https://git.unom.io/unom/punktfunk/src/branch/main/packaging/flatpak/README.md).
 
 ## Windows
 

docs-site/content/docs/install.md

@@ -43,12 +43,12 @@ signed installer — see [Windows Host](/docs/windows-host) for what it includes
    ```
 
 3. Run `punktfunk-host-setup-<ver>.exe` (elevated). It installs to `C:\Program Files\punktfunk`,
-   optionally installs the bundled **SudoVDA** virtual-display driver, and registers + starts the
+   installs the bundled **pf-vdisplay** virtual-display driver, and registers + starts the
    `LocalSystem` service (`/VERYSILENT` for an unattended install). Upgrades and uninstall go through
    Add/Remove Programs.
 
-You need an NVIDIA GPU + driver (the host is NVENC-only on Windows). More detail — including the CLI
-`punktfunk-host service install` path — is in
+For hardware encode you need a GPU — NVIDIA (NVENC), AMD (AMF), or Intel (QSV); there's a software
+fallback without one. More detail — including the CLI `punktfunk-host service install` path — is in
 [Running as a Service → Windows](/docs/running-as-a-service#windows).
 
 ## What the packages are

docs-site/content/docs/meta.json

@@ -17,6 +17,7 @@
     "---Connecting---",
     "clients",
     "install-client",
+    "steam-deck",
     "moonlight",
     "pairing",
     "---Configuration---",

docs-site/content/docs/requirements.md

@@ -19,9 +19,10 @@ environments it supports today, each with its own guide:
 Other wlroots compositors (Sway/Hyprland) also work but aren't a primary target. If your desktop isn't
 listed, the host still needs one of these compositor backends to create a virtual display.
 
-> **Windows host:** punktfunk also runs as a native host on **Windows 10/11 (x64) with an NVIDIA GPU**
-> — a signed installer that registers a service and bundles a virtual-display driver. It's NVIDIA-only
-> and newer than the Linux host; see [Windows Host](/docs/windows-host).
+> **Windows host:** punktfunk also runs as a native host on **Windows 10/11 (x64)** — a signed
+> installer that registers a service and bundles a virtual-display driver. It encodes on NVIDIA
+> (NVENC), AMD (AMF), or Intel (QSV), with a software fallback, and is newer than the Linux host; see
+> [Windows Host](/docs/windows-host).
 
 ## GPU and driver
 

docs-site/content/docs/roadmap.md

@@ -35,7 +35,7 @@ see [Status & Progress](/docs/status).
   from one process.
 - **Native-resolution virtual displays** on Linux across KWin, GNOME/Mutter, gamescope, and
   Sway/wlroots, with a fully zero-copy GPU path to NVENC (stable 240 fps at 5120×1440).
-- **A native Windows host** (NVIDIA, x64) — a signed installer with secure-desktop capture and a
+- **A native Windows host** (x64; NVIDIA/AMD/Intel encode) — a signed installer with secure-desktop capture and a
   bundled virtual-display driver, and the only host that can stream **HDR** (10-bit BT.2020 PQ,
   captured from an HDR Windows desktop and encoded as HEVC Main10). See
   [Windows Host](/docs/windows-host). *(Beta — newer than the Linux host.)*
@@ -55,8 +55,8 @@ see [Status & Progress](/docs/status).
 - **Apple stage-2 presenter as the default.** The lower-latency `VTDecompressionSession` →
   `CAMetalLayer` path is live behind an opt-in flag and graduating to the default.
 - **Web console parity.** Surfacing the speed test and bitrate picker the apps already have.
-- **Windows host hardening.** Broader real-world testing, AMD/Intel encode (NVIDIA-only today), and
-  bundling the ViGEm gamepad driver.
+- **Windows host hardening.** Broader real-world testing — especially on-glass validation of the
+  AMD (AMF) and Intel (QSV) encode paths, which are CI-green but newer than NVENC.
 
 ## 🔭 Planned
 

docs-site/content/docs/running-as-a-service.md

@@ -95,13 +95,14 @@ model Sunshine/Apollo use.
 
 The easy path is the **signed installer**: download `punktfunk-host-setup-<ver>.exe` from the package
 registry ([`punktfunk-host-windows`](https://git.unom.io/unom/-/packages)) and run it. It drops the host
-into `C:\Program Files\punktfunk`, optionally installs the bundled **SudoVDA** virtual-display driver,
-and registers + starts the service for you (`/VERYSILENT` for unattended). Upgrades and uninstall are
+into `C:\Program Files\punktfunk`, installs the bundled **pf-vdisplay** virtual-display driver, and
+registers + starts the service for you (`/VERYSILENT` for unattended). Upgrades and uninstall are
 handled through Add/Remove Programs.
 
 Prefer the CLI? Run `punktfunk-host service install` from an elevated prompt — see
-[Windows service](https://git.unom.io/unom/punktfunk/src/branch/main/docs/windows-service.md). Either
-way you need an NVIDIA GPU + driver (the host is NVENC-only on Windows).
+[Windows service](https://git.unom.io/unom/punktfunk/src/branch/main/docs/windows-service.md). For
+hardware encode you need a GPU — NVIDIA (NVENC), AMD (AMF), or Intel (QSV); the host falls back to
+software H.264 without one.
 
 ## Verifying
 

docs-site/content/docs/status.md

@@ -14,7 +14,7 @@ A high-level view of where punktfunk stands. The ordered plan of work is on the
 | **Core** — `punktfunk-core` + C ABI (protocol · FEC · crypto) | ✅ complete & hardened |
 | **GameStream host** (Moonlight-compatible) | ✅ working end-to-end; HDR/surround-audio polish open |
 | **Native protocol** — `punktfunk/1` (QUIC control + UDP data, GF(2¹⁶) Leopard FEC + AES-GCM) | ✅ full session planes, validated live |
-| **Windows host** (NVIDIA, x64) | 🟡 implemented & shipping as a signed installer; NVIDIA-only, newer than the Linux host |
+| **Windows host** (x64) | 🟡 implemented & shipping as a signed installer; NVIDIA/AMD/Intel encode, newer than the Linux host |
 | **macOS / iOS / iPadOS / tvOS client** | ✅ full client; on-glass stage-2 presenter behind an opt-in flag, becoming the default |
 | **Linux client** (`punktfunk-client`, GTK4/libadwaita) | ✅ full client; VAAPI zero-copy decode + software fallback |
 | **Windows client** (`punktfunk-client`, WinUI 3) | ✅ stage 1 complete; ships as signed MSIX; on-glass hardware validation pending |

docs-site/content/docs/steam-deck.md

@@ -0,0 +1,100 @@
+---
+title: Steam Deck (Decky)
+description: Install the punktfunk Decky plugin to discover, pair, and stream from the Steam Deck's Gaming Mode — no drop to Desktop.
+---
+
+The **Decky plugin** adds a **punktfunk** panel to the Steam Deck's Quick Access Menu (the `…`
+button), so you can find a host, pair, and start streaming **without leaving Gaming Mode**. It's the
+couch-friendly front end for the Steam Deck — built from real Steam UI, gamepad-navigable end to end.
+
+Under the hood the plugin doesn't decode video itself: it discovers hosts, runs the PIN pairing, and
+**launches the regular [Linux client](/docs/clients#linux-desktop-client-gtk4)** (the
+`io.unom.Punktfunk` Flatpak) the way gamescope needs so it fullscreens correctly. So the Deck has two
+ways to stream, and they share one client + one paired identity:
+
+- **Gaming Mode** → the **Decky plugin** (this page).
+- **Desktop Mode** → run the [Flatpak](/docs/install-client#steam-deck) directly, like any Linux app.
+
+## Before you start
+
+You need three things on the Deck:
+
+1. **Decky Loader** — the plugin loader. Install it from [decky.xyz](https://decky.xyz/) if you
+   haven't already.
+2. **The punktfunk client Flatpak** — the plugin launches it, so install it once in **Desktop Mode**:
+
+   ```sh
+   flatpak install --user https://flatpak.unom.io/io.unom.Punktfunk.flatpakref
+   ```
+
+   (Full options: [Install a Client → Steam Deck](/docs/install-client#steam-deck). Without it, the
+   panel's **Stream** button reports `client-not-found`.)
+3. **A punktfunk host** running on your LAN — see [Install the Host](/docs/install). The Deck finds
+   it automatically over mDNS, so nothing to configure here.
+
+## Install the plugin
+
+The plugin is published as a ready-to-install zip on every build. You don't need the Decky CLI or a
+developer toolchain — just paste a URL into Decky:
+
+1. On the Deck, open the **Quick Access Menu** (`…`) → the **plug** icon (Decky) → the **gear**
+   (Settings) → enable **Developer Mode**.
+2. Open the new **Developer** tab and choose **Install Plugin from URL**.
+3. Paste the **stable** link and confirm:
+
+   ```
+   https://git.unom.io/api/packages/unom/generic/punktfunk-decky/latest/punktfunk.zip
+   ```
+
+The **punktfunk** panel appears in the Quick Access Menu right away — no Deck restart needed.
+
+> **Channels.** The link above is the **stable** channel (moves on `vX.Y.Z` releases). For the latest
+> `main` build use the **canary** zip — `…/generic/punktfunk-decky/canary/punktfunk.zip` — or pin an
+> exact version with `…/punktfunk-decky/<version>/punktfunk.zip`. See [Release Channels](/docs/channels).
+
+## Use it
+
+Open the **punktfunk** panel from the Quick Access Menu, or **Open punktfunk** for the full-screen
+page (host list + stream settings).
+
+- **Discover** — hosts on your network appear automatically (mDNS). Tap **Refresh** to rescan. A
+  lock icon means the host requires [pairing](/docs/pairing).
+- **Pair** — for a locked host, [arm pairing on the host](/docs/pairing) (its console or web
+  console shows a 4-digit PIN), then enter that PIN on the Deck's keypad. Pairing persists, so the
+  next connection is silent.
+- **Stream** — pick a host and the stream launches fullscreen in Gaming Mode (as a hidden Steam
+  shortcut, so gamescope focuses it).
+- **Settings** — resolution, refresh, bitrate, gamepad type, and mic, written to the client the
+  plugin launches. Leave **Resolution** / **Refresh** on *Native* to get the Deck's own mode.
+
+To **leave a stream**: the in-client controller chord **L1 + R1 + Start + Select**, or close the
+"game" from the Steam overlay. Exiting the client ends the Steam game and drops you back to Gaming
+Mode.
+
+## Updating
+
+The plugin **checks for updates itself** — no Decky store needed. When a newer build is available it
+shows an **Update to vX** button (in the Quick Access Menu panel and on the full page). Tap it,
+confirm Decky's prompt, and the plugin downloads, verifies, replaces itself, and reloads — without
+leaving Gaming Mode.
+
+The check follows the [channel](/docs/channels) you installed from: a plugin installed from the
+**stable** link tracks stable releases; one installed from the **canary** link tracks `main` builds.
+
+> If the **Update** button never appears (an older Decky Loader, or no network), update manually:
+> Decky → **Developer** → **Install Plugin from URL**, and paste the same channel link again. Decky
+> replaces the installed copy in place.
+
+## Troubleshooting
+
+| Symptom | Fix |
+|---|---|
+| **Stream** shows `client-not-found` | Install the client Flatpak in Desktop Mode (see [Before you start](#before-you-start)). |
+| No hosts listed | Make sure the host is running and on the **same LAN**; the Deck needs `avahi` (shipped on SteamOS). Tap **Refresh**. |
+| Pairing fails / "not armed" | The PIN is shown only after you **arm pairing on the host**. Arm it, then enter the PIN within the window. |
+| Stream launches but doesn't focus | Start it from the panel (not by launching the Flatpak by hand) so Steam/gamescope focuses it. |
+
+The plugin source lives in
+[`clients/decky`](https://git.unom.io/unom/punktfunk/src/branch/main/clients/decky/README.md).
+</content>
+</invoke>

docs-site/content/docs/windows-host.md

@@ -1,45 +1,78 @@
 ---
 title: "Windows Host"
-description: "Run the punktfunk streaming host on a Windows PC — a first-class, virtual-display host."
+description: "Run the punktfunk streaming host on a Windows PC — a first-class, all-vendor, virtual-display host."
 ---
 
+Set up a punktfunk host on a **Windows 10/11 PC** and stream its desktop or games to any punktfunk or
+[Moonlight](/docs/moonlight) client. A signed installer registers a Windows service that streams at the
+client's **exact resolution and refresh** via punktfunk's own **virtual display** — including
+**HDR10** (10-bit BT.2020 PQ) when your Windows desktop is in HDR mode. The virtual display is created
+on the fly, so you need **no second monitor and no dummy HDMI plug**, and capture keeps working even on
+the secure desktop (UAC prompts, the lock screen).
 
-**Status: implemented and shipping — x64-only.** Alongside the Linux host, punktfunk runs as a
-first-class native **Windows host**: a signed installer registers a `LocalSystem` service that streams
-your Windows desktop or games to any punktfunk or Moonlight client, at the client's exact resolution
-via a **virtual display** — including **HDR10** (10-bit BT.2020 PQ) when your Windows desktop is in HDR
-mode. punktfunk has its own **indirect display driver (IDD)** that the host pushes finished frames
-straight into, so you get a real on-the-fly virtual display with no physical monitor or dummy HDMI
-plug — even on the secure desktop (UAC / lock screen). The Windows host is newer and less
-battle-tested than the Linux host. (The Linux host is 8-bit only — HDR there is blocked upstream.)
+> New to this? Skim [Requirements](/docs/requirements) first.
 
-> This page is about the Windows **host** (streaming *from* a Windows PC). To stream *to* a Windows
-> PC, see the [Windows client](/docs/clients#windows-desktop-client).
+> This page is about the Windows **host** — streaming *from* a Windows PC. To stream *to* a Windows PC,
+> see the [Windows client](/docs/clients#windows-desktop-client).
 
 ## Requirements
 
-- **Windows 10/11, x64.** ARM64 is not supported — both NVENC and the virtual-display driver are
-  x64-only.
-- **An NVIDIA GPU + driver.** The host encodes with NVENC (`nvEncodeAPI64.dll`); there is no other
-  encoder backend on Windows.
-- **(Optional) ViGEmBus** for virtual gamepads — a manual prerequisite for now
-  ([releases](https://github.com/nefarius/ViGEmBus/releases)).
+- **Windows 10 or 11, x64.** ARM64 is not built (no ARM64 NVIDIA driver, and the virtual-display
+  driver is x64-only).
+- **A GPU for hardware encode** — the host auto-detects the vendor:
+  - **NVIDIA** → NVENC
+  - **AMD** → AMF
+  - **Intel** → QSV
+
+  No discrete GPU? The host falls back to a **software H.264** encoder (higher CPU use, lower quality —
+  fine for light desktop use).
+- **No gamepad prerequisite.** The virtual gamepad drivers are bundled in the installer — there is
+  nothing else to download. (Earlier builds needed ViGEmBus; it is no longer used.)
 
 ## Install
 
-Download the signed `punktfunk-host-setup-<ver>.exe` from the package registry and run it — it
-installs the host into `C:\Program Files\punktfunk`, optionally installs the bundled **SudoVDA**
-virtual-display driver, and registers + starts the service. Full steps (including the silent install
-and the CLI `punktfunk-host service install` path) are in
-[Running as a Service → Windows](/docs/running-as-a-service#windows); packaging internals live in
+Download the signed `punktfunk-host-setup-<ver>.exe` from the
+[package registry](https://git.unom.io/unom/-/packages) and run it. The installer:
+
+- drops the host into `C:\Program Files\punktfunk` and registers + starts the **`PunktfunkHost`**
+  service,
+- installs the bundled **virtual-display driver** (`pf-vdisplay`) so the host can create per-client
+  displays,
+- installs the bundled **virtual gamepad drivers** (DualSense, DualShock 4, Xbox 360),
+- registers the bundled **HDR Vulkan layer** so Vulkan games can enable HDR over the virtual display,
+- sets up the **web management console** (see below).
+
+For an unattended install, append `/VERYSILENT`. Upgrades and uninstall go through **Add/Remove
+Programs**; your config and pairings are kept across upgrades. Prefer the CLI, or want the full
+service/firewall details? See [Running as a Service → Windows](/docs/running-as-a-service#windows).
+Packaging internals live in
 [`packaging/windows`](https://git.unom.io/unom/punktfunk/src/branch/main/packaging/windows/README.md).
 
+### Web console & pairing
+
 The installer also sets up the **web management console** (status, paired devices, the PIN pairing
-flow): it bundles the console plus its own bun runtime and runs it as the **`PunktfunkWeb`** service
-on **`http://<this-PC>:3000`**, starting at boot. During setup you choose the console **login
-password** (pre-filled with a secure random default and shown again on the final page); change it
-later in `%ProgramData%\punktfunk\web-password`. Open the console from any browser on the LAN and log
-in — no extra install, and the host's management API stays loopback-only behind it.
+flow): it bundles the console plus its own runtime and runs it as the **`PunktfunkWeb`** service on
+**`http://<this-PC>:3000`**, starting at boot. During setup you choose the console **login password**
+(pre-filled with a secure random default and shown again on the final page); change it later in
+`%ProgramData%\punktfunk\web-password`.
+
+The host **requires PIN pairing** by default (secure on a LAN). To connect the first time, open the
+console from any browser on the LAN, log in, go to **Devices → arm pairing**, and enter the PIN on
+your [client](/docs/clients). The host's own management API stays loopback-only behind the console.
+
+### Configure
+
+The service reads `%ProgramData%\punktfunk\host.env`. The defaults work out of the box; common knobs:
+
+- `PUNKTFUNK_ENCODER=auto` — `auto` picks NVENC/AMF/QSV by GPU vendor. Force one with `nvenc`, `amf`,
+  `qsv`, or `sw` (software).
+- `PUNKTFUNK_HOST_CMD` — the service runs `serve --gamestream` by default (native punktfunk/1 **plus**
+  the GameStream/Moonlight-compat planes). Set it to `serve` for a **secure native-only** host with no
+  GameStream surface (GameStream pairs over plain HTTP and uses weaker legacy encryption — trusted LAN
+  only).
+
+Edit the file, then restart: `punktfunk-host service stop` / `punktfunk-host service start`. See the
+[Configuration reference](/docs/configuration) for every option.
 
 ## How it works
 
@@ -58,23 +91,36 @@ pipeline orchestration are all shared with the Linux host. The Windows host is a
 
 | Subsystem | Linux backend | Windows backend |
 |---|---|---|
-| **Capture** | xdg ScreenCast portal → PipeWire (dmabuf) | **Windows.Graphics.Capture** (+ Desktop Duplication for the secure desktop) → D3D11 texture; FP16/10-bit when the desktop is HDR |
-| **Virtual display** | KWin / Mutter / Sway / gamescope | **SudoVDA** signed IDD — create a `WxH@Hz` monitor per session, capture it, tear it down |
-| **Encode** | `ffmpeg-next` NVENC (CUDA hwframes) | **NVENC** with a D3D11 device (`--features nvenc`); HEVC Main10 / BT.2020 PQ for HDR |
+| **Capture** | xdg ScreenCast portal → PipeWire (dmabuf) | **Windows.Graphics.Capture** + **Desktop Duplication** (secure desktop), with a zero-copy path straight from the virtual-display driver; FP16/10-bit when the desktop is HDR |
+| **Virtual display** | KWin / Mutter / Sway / gamescope | **pf-vdisplay** signed IDD — create a `WxH@Hz` monitor per session, capture it, tear it down |
+| **Encode** | NVENC (CUDA) / VAAPI (AMD·Intel) / software | **NVENC** (NVIDIA) · **AMF** (AMD) · **QSV** (Intel) · software H.264; HEVC Main10 / BT.2020 PQ for HDR |
 | **Input — mouse/keyboard** | libei / wlr protocols | **SendInput** (Win32 VK + absolute mouse) |
-| **Input — gamepads** | uinput Xbox 360 pad + rumble | **ViGEm** virtual pad + rumble back-channel |
+| **Input — gamepads** | uinput Xbox 360 + UHID DualSense/DS4 | **UMDF** virtual pads — DualSense, DualShock 4, Xbox 360 (XUSB) + rumble |
 | **Audio capture** | PipeWire sink-monitor | **WASAPI loopback** |
 | **Virtual mic** | PipeWire `Audio/Source` | WASAPI virtual mic |
 
-The virtual display uses **[SudoVDA](https://github.com/VirtualDrivers)** (the Sunshine Virtual
-Display Adapter) — a pre-built, signed Indirect Display Driver — so there is **no kernel driver to
-author or WHQL-sign**. The installer bundles and stages it; if it's absent, the host falls back to
-capturing an existing monitor (losing the per-client native-resolution output).
+The virtual display uses **pf-vdisplay**, punktfunk's own all-Rust **Indirect Display Driver (IDD)** —
+the host pushes finished frames straight into it, so you get a real virtual display with no physical
+monitor or dummy plug. The installer bundles and stages the (self-signed) driver; if it isn't
+installed, the host falls back to capturing an existing monitor, losing the per-client native-resolution
+output.
 
-## Limitations
+### HDR
 
-- **NVIDIA-only.** NVENC is the only encoder backend — there is no AMD / Intel / software encode path
-  on Windows.
-- **x64-only.** No ARM64 build (no ARM64 NVIDIA driver, and SudoVDA is x64-only).
+When your Windows desktop is in **HDR** mode, the host captures it as 10-bit, encodes **HEVC Main10 /
+BT.2020 PQ**, and the client auto-detects HDR from the stream. A small always-on **Vulkan layer**
+(bundled and registered by the installer) also lets **Vulkan games** enable HDR over the virtual
+display — something the NVIDIA/AMD drivers otherwise refuse on an indirect display. The layer is
+self-gating: it's a no-op on SDR and on real monitors. HDR is **Windows-only** (the Linux host is
+8-bit, blocked upstream).
+
+## Notes & limits
+
+- **AMD / Intel encode is newer.** The NVENC path is the most exercised; AMF (AMD) and QSV (Intel) are
+  built and tested in CI but less battle-tested on real hardware. Software H.264 is the GPU-less
+  fallback.
+- **x64-only.** No ARM64 build — no ARM64 NVIDIA driver, and the virtual-display driver is x64-only.
 - **Newer than the Linux host.** The Linux host is the most battle-tested path; the Windows host is
-  more recent, with the virtual-mic and gamepad backends the youngest pieces.
+  more recent, with the virtual-mic and AMD/Intel encode backends the youngest pieces.
+
+Trouble? See [Troubleshooting](/docs/troubleshooting) and [Pairing](/docs/pairing).

docs-site/package.json

@@ -14,6 +14,7 @@
     "@scalar/api-reference-react": "^0.9.47",
     "@tanstack/react-router": "^1.121.0",
     "@tanstack/react-start": "^1.121.0",
+    "@unom/app-ui": "^0.1.0",
     "@unom/style": "^0.4.4",
     "@unom/ui": "^0.8.16",
     "fumadocs-core": "^16.10.5",

docs-site/src/components/Footer.tsx

@@ -1,51 +1,27 @@
 import { getRouteApi } from '@tanstack/react-router'
-import type { NavigationLink, NavigationSection } from '@/lib/cms'
+import { FooterView } from '@unom/app-ui/footer'
 
 const rootApi = getRouteApi('__root__')
 
-// The docs share the marketing site's footer (same CMS global). Root-relative
-// links target the website, so resolve them against its origin — the docs don't
-// host /legal/* etc. themselves. Mirrors the website Footer, themed for docs.
+// Footer markup is shared with the marketing site via @unom/app-ui so the two
+// stay in sync. It themes itself through @unom/style tokens, which the docs map
+// onto their Fumadocs surfaces. Root-relative links target the website (the
+// docs don't host /legal/* etc.), so rebase them onto its origin.
 const SITE_URL = 'https://punktfunk.unom.io'
-const resolve = (to?: string | null) =>
-  to ? (to.startsWith('/') ? `${SITE_URL}${to}` : to) : '#'
+const resolveHref = (to: string) =>
+  to.startsWith('/') ? `${SITE_URL}${to}` : to
 
 export default function Footer() {
   const { footer } = rootApi.useLoaderData()
-  const sections: NavigationSection[] = footer?.sections ?? []
-  const tagline = footer?.tagline?.trim()
-
-  if (!sections.length && !tagline) return null
 
   return (
-    <footer className="border-t border-fd-border bg-fd-card">
-      <div className="mx-auto flex w-full max-w-6xl flex-row flex-wrap gap-12 px-4 py-12 sm:px-6">
-        {sections.map((group, gi) => (
-          <div key={group.id ?? gi}>
-            {group.title && (
-              <h3 className="mb-2 text-sm font-semibold text-fd-foreground">
-                {group.title}
-              </h3>
-            )}
-            <div className="flex flex-col gap-1">
-              {(group.entries ?? []).map((item: NavigationLink, i) => (
-                <a
-                  key={item.id ?? `${item.to}-${i}`}
-                  href={resolve(item.to)}
-                  className="text-sm text-fd-muted-foreground transition-colors hover:text-fd-foreground"
-                >
-                  {item.label}
-                </a>
-              ))}
-            </div>
-          </div>
-        ))}
-        {tagline && (
-          <p className="ml-auto self-end text-sm text-fd-muted-foreground">
-            {tagline}
-          </p>
-        )}
-      </div>
-    </footer>
+    <FooterView
+      sections={footer?.sections}
+      tagline={footer?.tagline}
+      socials={footer?.socials}
+      socialsLabel="Socials"
+      resolveHref={resolveHref}
+      className="border-t border-fd-border"
+    />
   )
 }

docs-site/src/lib/cms.ts

@@ -1,33 +1,21 @@
 // The docs reuse the punktfunk footer from the shared unom CMS (cms.unom.io).
-// The CMS is multi-tenant: footer is a per-tenant collection, so scope the read
-// to this project's tenant. Read-only GET, so a plain typed fetch rather than
-// pulling in the Payload SDK + generated types.
+// The footer shape comes from @unom/app-ui/footer so the docs and the marketing
+// site share one type. The CMS is multi-tenant: footer is a per-tenant
+// collection, so scope the read to this project's tenant. Read-only GET, so a
+// plain typed fetch rather than pulling in the Payload SDK + generated types.
+import type { FooterData } from '@unom/app-ui/footer'
+
 const CMS_URL = 'https://cms.unom.io'
 
 // This project's tenant in the shared CMS.
 const TENANT = 'punktfunk'
 
-export interface NavigationLink {
-  id?: string | null
-  label?: string | null
-  to?: string | null
-}
+export type { FooterData as Footer } from '@unom/app-ui/footer'
 
-export interface NavigationSection {
-  id?: string | null
-  title?: string | null
-  entries?: NavigationLink[] | null
-}
-
-export interface Footer {
-  tagline?: string | null
-  sections?: NavigationSection[] | null
-}
-
-export async function findFooter(): Promise<Footer | null> {
+export async function findFooter(): Promise<FooterData | null> {
   const query = `where%5Btenant.slug%5D%5Bequals%5D=${TENANT}&locale=en&depth=1&limit=1`
   const res = await fetch(`${CMS_URL}/api/footers?${query}`)
   if (!res.ok) throw new Error(`CMS footer request failed: ${res.status}`)
-  const data = (await res.json()) as { docs?: Footer[] }
+  const data = (await res.json()) as { docs?: FooterData[] }
   return data.docs?.[0] ?? null
 }

docs-site/src/styles/app.css

@@ -7,6 +7,7 @@
    design-token system the punktfunk marketing site also builds on. */
 @source '../../node_modules/fumadocs-ui/dist/**/*.js';
 @source '../../node_modules/@unom/ui/dist/**/*.{js,mjs}';
+@source '../../node_modules/@unom/app-ui/dist/**/*.{js,mjs}';
 
 /* ── punktfunk brand ────────────────────────────────────────────────────────
    The brand colour is the violet lens mark. (The marketing site's blue is just

include/punktfunk_core.h

@@ -94,6 +94,12 @@
 // BT.2020 PQ HDR10 (implies 10-bit). (Mirrors `quic::VIDEO_CAP_HDR`.)
 #define PUNKTFUNK_VIDEO_CAP_HDR 2
 
+// Video-capability bit for [`punktfunk_connect_ex5`] (`video_caps`): the client can decode a
+// full-chroma 4:4:4 HEVC stream (Range Extensions). The host emits 4:4:4 only when this is set,
+// the host opted in, the codec is HEVC, and the GPU supports it — else the stream stays 4:2:0 and
+// [`punktfunk_connection_chroma_format`] reports the real value. (Mirrors `quic::VIDEO_CAP_444`.)
+#define PUNKTFUNK_VIDEO_CAP_444 4
+
 // 16-byte AEAD authentication tag appended by GCM.
 #define TAG_LEN 16
 
@@ -180,6 +186,27 @@
 #define VIDEO_CAP_HDR 2
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`Hello::video_caps`] bit: the client can decode a full-chroma **4:4:4** HEVC stream (HEVC
+// Range Extensions / Rec.ITU-T H.265 `chroma_format_idc = 3`). The host emits 4:4:4 ONLY when this
+// bit is set, the host opted in (`PUNKTFUNK_444`), the codec is HEVC, **and** the GPU/driver
+// actually supports a 4:4:4 encode (probed) — otherwise the session stays 4:2:0 and
+// [`Welcome::chroma_format`] reflects the real resolved value. Independent of 10-bit/HDR (4:4:4 is a
+// chroma decision, bit depth is a depth decision; the two may combine where the hardware allows).
+#define VIDEO_CAP_444 4
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// HEVC `chroma_format_idc` for 4:2:0 — what every pre-4:4:4 build produced and the back-compat
+// default when a peer omits [`Welcome::chroma_format`].
+#define CHROMA_IDC_420 1
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// HEVC `chroma_format_idc` for full-chroma 4:4:4 (Range Extensions).
+#define CHROMA_IDC_444 3
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // Longest device name carried in a [`Hello`] (bytes of UTF-8; longer names are truncated on
 // encode, rejected on decode — a one-byte length prefix caps it at 255 anyway).
@@ -498,6 +525,25 @@ typedef struct {
 } PunktfunkAudioPacket;
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// One decoded audio frame from [`punktfunk_connection_next_audio_pcm`]: interleaved 32-bit
+// float PCM at 48 kHz, in the canonical wire channel order `FL FR FC LFE RL RR SL SR` (the
+// first `channels` of it). `samples` points at `frame_count * channels` floats and borrows
+// connection memory **until the next PCM call** on this handle.
+typedef struct {
+    // Interleaved f32 samples (wire channel order), `frame_count * channels` long.
+    const float *samples;
+    // Samples per channel in this frame.
+    uint32_t frame_count;
+    // Channel count (2/6/8) — the negotiated [`punktfunk_connection_audio_channels`].
+    uint8_t channels;
+    // Source packet sequence number.
+    uint32_t seq;
+    // Capture presentation timestamp (ns).
+    uint64_t pts_ns;
+} PunktfunkAudioPcm;
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // One DualSense HID-output feedback event a game wrote to the host's virtual pad
 // ([`punktfunk_connection_next_hidout`]). `kind` selects which fields are meaningful — replay it
@@ -832,6 +878,33 @@ PunktfunkConnection *punktfunk_connect_ex5(const char *host,
                                            uint32_t timeout_ms);
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Like [`punktfunk_connect_ex5`], but additionally requests the audio channel count:
+// `2` (stereo, the default behaviour of every earlier variant), `6` (5.1) or `8` (7.1). The host
+// clamps the request to what it can actually capture and echoes the resolved count via
+// [`punktfunk_connection_audio_channels`]; the `0xC9` audio frames are Opus-(multi)stream encoded
+// for that layout. A client that wants surround calls this; everything else inherits stereo.
+//
+// # Safety
+// Same as [`punktfunk_connect`].
+PunktfunkConnection *punktfunk_connect_ex6(const char *host,
+                                           uint16_t port,
+                                           uint32_t width,
+                                           uint32_t height,
+                                           uint32_t refresh_hz,
+                                           uint32_t compositor,
+                                           uint32_t gamepad,
+                                           uint32_t bitrate_kbps,
+                                           uint8_t video_caps,
+                                           uint8_t audio_channels,
+                                           const char *launch_id,
+                                           const uint8_t *pin_sha256,
+                                           uint8_t *observed_sha256_out,
+                                           const char *client_cert_pem,
+                                           const char *client_key_pem,
+                                           uint32_t timeout_ms);
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // Generate a persistent client identity: a self-signed certificate + private key, both
 // PEM, NUL-terminated, written into the caller's buffers. Generate ONCE, store both
@@ -897,6 +970,36 @@ PunktfunkStatus punktfunk_connection_next_audio(PunktfunkConnection *c,
                                                 uint32_t timeout_ms);
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Read the audio channel count the host resolved for this session (from its Welcome): `2`
+// (stereo), `6` (5.1) or `8` (7.1). `*out` is filled when non-NULL. The `0xC9` Opus frames are
+// (multistream-)encoded for this layout; an embedder decoding raw frames itself must build its
+// decoder from THIS value (see [`crate::audio::layout_for`]) — or use
+// [`punktfunk_connection_next_audio_pcm`], which decodes in-core. Available immediately after a
+// successful connect (it doesn't change without a reconfigure).
+//
+// # Safety
+// `c` is a valid connection handle; `out` is NULL or writable for one `u8`.
+PunktfunkStatus punktfunk_connection_audio_channels(PunktfunkConnection *c, uint8_t *out);
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Pull the next audio frame and **decode it in-core** to interleaved f32 PCM — for embedders
+// without a multistream-capable Opus decoder (e.g. Apple, whose AudioToolbox Opus path is
+// stereo-only). The decoder is built once from the negotiated channel count and handles 2/6/8
+// channels (a 1-coupled-stream multistream decoder is exactly a stereo decoder). Same
+// timeout/closed semantics as [`punktfunk_connection_next_audio`]; `out->samples` borrows
+// connection memory until the next PCM call on this handle. Use EITHER this or
+// [`punktfunk_connection_next_audio`] on a given connection, from one dedicated audio thread —
+// not both (they share the underlying queue).
+//
+// # Safety
+// `c` is a valid connection handle; `out` is writable. At most one thread pulls audio.
+PunktfunkStatus punktfunk_connection_next_audio_pcm(PunktfunkConnection *c,
+                                                    PunktfunkAudioPcm *out,
+                                                    uint32_t timeout_ms);
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // Pull the next rumble (force-feedback) update, waiting up to `timeout_ms`. Amplitudes
 // are 0..0xFFFF (`low` = low-frequency motor, `high` = high-frequency), `(0, 0)` = stop.
@@ -960,6 +1063,18 @@ PunktfunkStatus punktfunk_connection_color_info(PunktfunkConnection *c,
                                                 uint8_t *bit_depth);
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Read the session's resolved chroma subsampling (from the host's Welcome) as the HEVC
+// `chroma_format_idc`: `1` = 4:2:0 (the default every pre-4:4:4 host produced), `3` = full-chroma
+// 4:4:4. `*out` is filled when non-NULL. The in-band SPS is authoritative; this lets the embedder
+// pre-size its decoder / pick a 4:4:4 pixel format up front. Available immediately after a
+// successful connect (it doesn't change without a reconfigure).
+//
+// # Safety
+// `c` is a valid connection handle; `out` is NULL or writable for one `u8`.
+PunktfunkStatus punktfunk_connection_chroma_format(PunktfunkConnection *c, uint8_t *out);
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // Send one input event to the host as a QUIC datagram (non-blocking enqueue).
 //

packaging/arch/PKGBUILD

@@ -72,6 +72,8 @@ package_punktfunk-host() {
               'xdg-desktop-portal-wlr: portal for the headless Sway session helper'
               'punktfunk-web: browser management console (device pairing + status)')
   install=punktfunk-host.install
+  # User-editable config: the headless game-mode drop-in (see below) — don't clobber local edits.
+  backup=('etc/gamescope-session-plus/sessions.d/steam')
   local R; R="$(_repo)"; local T="$srcdir/target/release"
 
   install -Dm0755 "$T/punktfunk-host"                 "$pkgdir/usr/bin/punktfunk-host"
@@ -86,6 +88,12 @@ package_punktfunk-host() {
   install -Dm0644 "$R/scripts/punktfunk-kde-session.service" "$pkgdir/usr/lib/systemd/user/punktfunk-kde-session.service"
   sed -i 's#%h/punktfunk/scripts/headless/run-headless-kde.sh#/usr/share/punktfunk/headless/run-headless-kde.sh#' \
     "$pkgdir/usr/lib/systemd/user/punktfunk-kde-session.service"
+  # KWin Desktop-mode authorization: non-launcher .desktop whose X-KDE-Wayland-Interfaces lets the
+  # host bind KWin's restricted zkde_screencast (virtual output) + fake_input globals on an
+  # interactive Plasma session. Must ship with the host (KWin caches the per-exe grant on first
+  # connect). See the file's header comment.
+  install -Dm0644 "$R/packaging/linux/io.unom.Punktfunk.Host.desktop" \
+    "$pkgdir/usr/share/applications/io.unom.Punktfunk.Host.desktop"
   # headless session helpers + env templates + OpenAPI doc
   install -Dm0755 "$R/scripts/headless/run-headless-kde.sh"  "$pkgdir/usr/share/punktfunk/headless/run-headless-kde.sh"
   install -Dm0755 "$R/scripts/headless/run-headless-sway.sh" "$pkgdir/usr/share/punktfunk/headless/run-headless-sway.sh"
@@ -94,6 +102,11 @@ package_punktfunk-host() {
   install -Dm0644 "$R/scripts/host.env.example"              "$pkgdir/usr/share/punktfunk/host.env.example"
   install -Dm0644 "$R/packaging/bazzite/host.env"            "$pkgdir/usr/share/punktfunk/host.env.bazzite"
   install -Dm0644 "$R/packaging/kde/host.env"                "$pkgdir/usr/share/punktfunk/host.env.kde"
+  # Headless GAME-mode fix: gamescope-session-plus drop-in that uses the headless backend when no
+  # display is connected (so SteamOS/Bazzite "Switch to Game Mode" works on a display-less streaming
+  # host). No-op on display-attached boxes; sourced as /etc/gamescope-session-plus/sessions.d/steam.
+  install -Dm0644 "$R/packaging/bazzite/gamescope-headless-session" \
+    "$pkgdir/etc/gamescope-session-plus/sessions.d/steam"
   install -Dm0644 "$R/api/openapi.json"                 "$pkgdir/usr/share/punktfunk/openapi.json"
   install -Dm0644 "$R/LICENSE-MIT"    "$pkgdir/usr/share/licenses/punktfunk-host/LICENSE-MIT"
   install -Dm0644 "$R/LICENSE-APACHE" "$pkgdir/usr/share/licenses/punktfunk-host/LICENSE-APACHE"

packaging/bazzite/gamescope-headless-session

@@ -0,0 +1,22 @@
+# punktfunk: headless game-mode fallback for gamescope-session-plus.
+#
+# Installed as /etc/gamescope-session-plus/sessions.d/steam. The gamescope-session-plus launcher
+# SOURCES this (shell, with `set -a` so assignments auto-export) AFTER its /usr/share defaults, so it
+# can override the session's gamescope flags.
+#
+# Why: on a box with NO connected display (a dedicated streaming host), the stock Steam game mode runs
+# gamescope's DRM backend against a physical panel (`--prefer-output *,eDP-1`). With nothing to scan
+# out, gamescope crashes on launch; after 5 strikes Bazzite/SteamOS force-selects the desktop session
+# and "Switch to Game Mode" appears broken. Falling back to gamescope's HEADLESS backend makes game
+# mode render entirely offscreen and expose a PipeWire node, which the punktfunk host captures and
+# streams — full gamescope game mode (per-game res / FSR / HDR / VRR / frame-limit), no monitor needed.
+#
+# Safe by construction:
+#   * NO-OP when any display is connected -> the normal DRM game mode runs unchanged.
+#   * Only sets values that are still unset (`: "${VAR:=...}"`), so the punktfunk host's per-client
+#     mode (SCREEN_WIDTH/SCREEN_HEIGHT injected via systemd-run for a managed session) still wins.
+if ! grep -qx connected /sys/class/drm/*/status 2>/dev/null; then
+    : "${BACKEND:=headless}"
+    : "${SCREEN_WIDTH:=1920}"
+    : "${SCREEN_HEIGHT:=1080}"
+fi

packaging/bazzite/host.env

@@ -20,12 +20,25 @@ PUNKTFUNK_ZEROCOPY=1
 #   PUNKTFUNK_COMPOSITOR=kwin|mutter|wlroots|gamescope
 #   PUNKTFUNK_INPUT_BACKEND=libei|wlr|gamescope|uinput
 #
-# In Gaming Mode the host MANAGES a gamescope-session-plus at the CLIENT's resolution by default
-# (tears the TV's autologin down on connect; restores it on a debounced idle, reused on a quick
-# reconnect). To instead ATTACH to the running TV session at its own mode (couch-on-TV — gaming
-# stays live on the panel, no Steam restart), set:
-#   PUNKTFUNK_GAMESCOPE_ATTACH=1
-#   PUNKTFUNK_GAMESCOPE_APP=steam -gamepadui   # only for an ad-hoc bare-spawn fallback
+# GAME MODE = ATTACH (the box owns its session; the host follows). The box decides whether it's in
+# Steam Gaming Mode or a Desktop — you switch with the normal Steam UI / "Switch to Desktop". The
+# host just ATTACHES to whatever's live and captures it; it never tears the session down or relaunches
+# it. So switching Desktop<->Game is rock-solid, and when you disconnect the box STAYS in its current
+# mode — reconnecting drops you right back where you were. The streamed resolution in game mode is the
+# box's gamescope mode (see SCREEN_WIDTH/HEIGHT in /etc/gamescope-session-plus/sessions.d/steam).
+PUNKTFUNK_GAMESCOPE_ATTACH=1
 #
-# Follow a Gaming<->Desktop switch MID-STREAM (rebuild the backend in place, no reconnect):
-#   PUNKTFUNK_SESSION_WATCH=1
+# Opt OUT to the MANAGED model instead (host tears the box's gamescope down on connect and launches
+# its OWN at the CLIENT's exact resolution; restores on a debounced idle). Client-mode-following, but
+# it does not coexist with a box-owned game-mode session — pick one:
+#   PUNKTFUNK_GAMESCOPE_MANAGED=1   # (and remove PUNKTFUNK_GAMESCOPE_ATTACH above)
+#
+# Follow a Gaming<->Desktop switch MID-STREAM (rebuild the backend in place, no reconnect). This is
+# ON BY DEFAULT on Bazzite/SteamOS (the host detects the platform); set =0 to disable it:
+#   PUNKTFUNK_SESSION_WATCH=0
+#
+# HEADLESS GAME MODE: on a box with no display attached, Bazzite's "Switch to Game Mode" normally
+# crashes (gamescope's DRM backend has no panel to drive). The host package ships
+# /etc/gamescope-session-plus/sessions.d/steam, which auto-falls-back to gamescope's HEADLESS backend
+# when no display is connected — so game mode boots offscreen and streams, with no config here. It's a
+# no-op on display-attached boxes. (The host then auto-detects Gaming and streams it.)

packaging/bazzite/kde-desktop-setup.sh

@@ -1,35 +1,36 @@
 #!/usr/bin/env bash
-# One-shot setup so the punktfunk host can stream the Bazzite KDE *Desktop* session (KWin virtual
-# output at the client's resolution). Run ONCE as the streaming user (no root needed). Gaming Mode
-# (gamescope) needs none of this — it auto-attaches. Idempotent: safe to re-run.
+# One-shot setup so the punktfunk host can INJECT INPUT while streaming the Bazzite KDE *Desktop*
+# session. Run ONCE as the streaming user (no root needed). Gaming Mode (gamescope) needs none of
+# this — it auto-attaches. Idempotent: safe to re-run.
 #
 #   bash /usr/share/punktfunk/bazzite/kde-desktop-setup.sh
 #
-# Two things a normal KDE login lacks that the headless host needs:
-#  1. KWIN_WAYLAND_NO_PERMISSION_CHECKS=1 — so KWin exposes the privileged `zkde_screencast`
-#     virtual-output protocol to the host (an external client) at all.
-#  2. The `kde-authorized` RemoteDesktop grant — so libei input setup auto-approves instead of
-#     popping an "Allow remote control?" dialog the headless host can't answer.
-# After running, log out + back into the KDE Desktop session once (or reboot) so KWin restarts
-# with the flag. Gaming Mode is unaffected.
+# The VIRTUAL OUTPUT (video) needs no setup: the host package ships io.unom.Punktfunk.Host.desktop,
+# whose X-KDE-Wayland-Interfaces grants the host KWin's restricted zkde_screencast protocol on a
+# normal interactive Plasma session — least-privilege (only the host, only that interface), the same
+# mechanism krfb/krdp use. No session-wide KWIN_WAYLAND_NO_PERMISSION_CHECKS hack is needed. KWin
+# caches the grant per-executable on first connect, so after a FRESH host install log out + back into
+# the Desktop session once so KWin re-reads the file.
+#
+# The one thing a normal KDE login still lacks is the `kde-authorized` RemoteDesktop grant — so the
+# host's libei input setup auto-approves instead of popping an "Allow remote control?" dialog the
+# headless host can't answer. That's what this script seeds.
 set -euo pipefail
 
 GRANT_SRC="${PUNKTFUNK_GRANT_SRC:-/usr/share/punktfunk/headless/kde-authorized}"
-ENVD="$HOME/.config/environment.d/10-punktfunk-kwin.conf"
 GRANT_DST="$HOME/.local/share/flatpak/db/kde-authorized"
+# Older versions of this script wrote a session-wide KWIN_WAYLAND_NO_PERMISSION_CHECKS=1 env file to
+# unlock screencast. The shipped .desktop replaces it; remove the stale, over-broad override.
+STALE_ENVD="$HOME/.config/environment.d/10-punktfunk-kwin.conf"
 
-echo "punktfunk: KDE Desktop-mode setup"
+echo "punktfunk: KDE Desktop-mode input setup"
 
-# 1. KWin permission-check bypass (persistent, picked up by the next KDE session via systemd).
-mkdir -p "$(dirname "$ENVD")"
-cat > "$ENVD" <<'EOF'
-# punktfunk: let the streaming host bind KWin's privileged zkde_screencast (virtual output).
-# A dedicated streaming box; this relaxes KWin's Wayland permission checks for the desktop path.
-KWIN_WAYLAND_NO_PERMISSION_CHECKS=1
-EOF
-echo "  wrote $ENVD"
+if [[ -f "$STALE_ENVD" ]] && grep -q KWIN_WAYLAND_NO_PERMISSION_CHECKS "$STALE_ENVD" 2>/dev/null; then
+    rm -f "$STALE_ENVD"
+    echo "  removed stale $STALE_ENVD (screencast is now granted via the shipped .desktop)"
+fi
 
-# 2. RemoteDesktop portal grant for headless libei input (never clobber an existing one).
+# RemoteDesktop portal grant for headless libei input (never clobber an existing one).
 if [[ -s "$GRANT_DST" ]]; then
     echo "  grant DB already present ($GRANT_DST) — leaving it"
 elif [[ -s "$GRANT_SRC" ]]; then
@@ -44,5 +45,5 @@ else
     echo "  WARN: grant source not found at $GRANT_SRC — input will need a manual portal approval" >&2
 fi
 
-echo "punktfunk: done. Log out + back into the KDE Desktop session (or reboot) so KWin restarts"
-echo "          with the flag, then connect a client while in Desktop Mode."
+echo "punktfunk: done. On a fresh host install, log out + back into the KDE Desktop session once"
+echo "          (so KWin authorizes the host's virtual output), then connect a client in Desktop Mode."

packaging/debian/build-deb.sh

@@ -50,6 +50,13 @@ sed -i 's#%h/punktfunk/target/release/punktfunk-host#/usr/bin/punktfunk-host#' \
 install -Dm0644 scripts/punktfunk-kde-session.service "$STAGE/usr/lib/systemd/user/punktfunk-kde-session.service"
 sed -i 's#%h/punktfunk/scripts/headless/run-headless-kde.sh#/usr/share/punktfunk-host/headless/run-headless-kde.sh#' \
     "$STAGE/usr/lib/systemd/user/punktfunk-kde-session.service"
+
+# KWin Desktop-mode authorization: non-launcher .desktop whose X-KDE-Wayland-Interfaces lets the
+# host bind KWin's restricted zkde_screencast (virtual output) + fake_input globals on an
+# interactive Plasma session. Must ship with the host — KWin caches the per-exe grant on first
+# connect, so it has to be present before the host ever connects. See the file's header comment.
+install -Dm0644 packaging/linux/io.unom.Punktfunk.Host.desktop \
+    "$STAGE/usr/share/applications/io.unom.Punktfunk.Host.desktop"
 install -Dm0755 scripts/headless/run-headless-kde.sh   "$SHAREDIR/headless/run-headless-kde.sh"
 install -Dm0755 scripts/headless/run-headless-sway.sh  "$SHAREDIR/headless/run-headless-sway.sh"
 install -Dm0644 scripts/headless/kde-authorized        "$SHAREDIR/headless/kde-authorized"

packaging/linux/io.unom.Punktfunk.Host.desktop

@@ -0,0 +1,19 @@
+[Desktop Entry]
+Type=Application
+Name=Punktfunk Host
+Comment=punktfunk streaming host — KWin virtual-output / input authorization
+Exec=/usr/bin/punktfunk-host
+Terminal=false
+NoDisplay=true
+# This file is NOT a launcher — it exists so KWin authorizes the host to bind its restricted
+# Wayland globals when streaming the *Desktop* (KWin) session. KWin maps a connecting client to a
+# .desktop by resolving /proc/<pid>/exe against `Exec` (hence the absolute /usr/bin path), then
+# grants only the interfaces listed here (the same mechanism krfb-virtualmonitor / krdpserver use):
+#   * zkde_screencast_unstable_v1 — create the per-session virtual output at the client's mode.
+#   * org_kde_kwin_fake_input     — inject input directly (no RemoteDesktop portal dialog).
+# Comma-separated, per KWin's parser. Without this file KWin never advertises these to the host and
+# desktop-mode streaming fails with "KWin does not expose zkde_screencast_unstable_v1". Gaming Mode
+# (gamescope) does not use this path. NOTE: KWin caches the per-executable grant on first connect,
+# so this must be installed *before* the host first connects (a package install satisfies that; an
+# already-running KWin session needs a re-login to pick it up).
+X-KDE-Wayland-Interfaces=zkde_screencast_unstable_v1,org_kde_kwin_fake_input

packaging/rpm/punktfunk.spec

@@ -196,6 +196,14 @@ sed -i 's#%h/punktfunk/target/release/punktfunk-host#%{_bindir}/punktfunk-host#'
 install -Dm0644 scripts/punktfunk-kde-session.service %{buildroot}%{_userunitdir}/punktfunk-kde-session.service
 sed -i 's#%h/punktfunk/scripts/headless/run-headless-kde.sh#%{_datadir}/%{name}/headless/run-headless-kde.sh#' %{buildroot}%{_userunitdir}/punktfunk-kde-session.service
 
+# KWin authorization for Desktop-mode (KWin) streaming: a non-launcher .desktop whose
+# X-KDE-Wayland-Interfaces grants the host the restricted zkde_screencast (virtual output) +
+# fake_input globals on an interactive Plasma session. Must ship with the host so it is present
+# before the host first connects (KWin caches the per-exe grant). Replaces the old manual
+# KWIN_WAYLAND_NO_PERMISSION_CHECKS hack for the screencast permission.
+install -Dm0644 packaging/linux/io.unom.Punktfunk.Host.desktop \
+                %{buildroot}%{_datadir}/applications/io.unom.Punktfunk.Host.desktop
+
 # --- client subpackage ---
 install -Dm0755 target/release/punktfunk-client %{buildroot}%{_bindir}/punktfunk-client
 install -Dm0644 packaging/linux/io.unom.Punktfunk.desktop \
@@ -221,9 +229,17 @@ install -Dm0644 scripts/headless/punktfunk-sink.conf   %{buildroot}%{_datadir}/%
 install -Dm0644 scripts/host.env.example               %{buildroot}%{_datadir}/%{name}/host.env.example
 install -Dm0644 packaging/bazzite/host.env             %{buildroot}%{_datadir}/%{name}/host.env.bazzite
 install -Dm0644 packaging/kde/host.env                 %{buildroot}%{_datadir}/%{name}/host.env.kde
-# Bazzite KDE Desktop-mode one-shot setup (KWIN_WAYLAND_NO_PERMISSION_CHECKS + RemoteDesktop grant).
+# Bazzite KDE Desktop-mode one-shot setup (seeds the RemoteDesktop grant for libei input; the
+# screencast/virtual-output grant ships as io.unom.Punktfunk.Host.desktop, installed above).
 install -d %{buildroot}%{_datadir}/%{name}/bazzite
 install -Dm0755 packaging/bazzite/kde-desktop-setup.sh %{buildroot}%{_datadir}/%{name}/bazzite/kde-desktop-setup.sh
+# Headless GAME-mode fix: a gamescope-session-plus sessions.d drop-in that falls back to gamescope's
+# headless backend when no display is connected (so "Switch to Game Mode" works on a display-less
+# streaming host instead of crashing + 5-striking back to desktop). No-op on display-attached boxes.
+# Sourced by gamescope-session-plus as /etc/gamescope-session-plus/sessions.d/steam (after its
+# /usr/share defaults). Harmless on non-gamescope systems (the file is simply never read).
+install -Dm0644 packaging/bazzite/gamescope-headless-session \
+                %{buildroot}/etc/gamescope-session-plus/sessions.d/steam
 install -Dm0644 api/openapi.json                  %{buildroot}%{_datadir}/%{name}/openapi.json
 
 %if %{with web}
@@ -252,6 +268,10 @@ install -Dm0644 web/web.env.example                %{buildroot}%{_datadir}/punkt
 %{_prefix}/lib/sysctl.d/99-punktfunk-net.conf
 %{_userunitdir}/punktfunk-host.service
 %{_userunitdir}/punktfunk-kde-session.service
+%{_datadir}/applications/io.unom.Punktfunk.Host.desktop
+%dir /etc/gamescope-session-plus
+%dir /etc/gamescope-session-plus/sessions.d
+%config(noreplace) /etc/gamescope-session-plus/sessions.d/steam
 %dir %{_datadir}/%{name}
 %{_datadir}/%{name}/*