chore(release): bump workspace version to 0.25.0

A minor bump: 98 commits since v0.24.0. The headline is DualSense pad audio
(PR #23) — a wired DualSense playing a game's voice-coil haptics and its own
speaker, streamed from the host, on Android and the desktop session client
against a Windows host with Steam's driver present. Behind it: the haptics
sweep's twelve milestones closing more than twenty controller faults across
every client and both hosts; the audio quality/latency work (256 kbps stereo,
the Steam Streaming Microphone endpoint root cause, and the de-jitter ratchet
that left audio permanently behind the picture); and MTU resilience plus
mid-session shard renegotiation, which turns the silent all-black stream on a
sub-1330-byte path into a diagnosed warning that heals itself. Plus the Decky
plugin reduced to a launcher, system-button routing with hold-Select, gamepad-UI
profiles on all three UIs, `discover`/`launch --request-access` in the CLI, and
the Sunshine false-conflict and crashed-host display-restore fixes.

The canary base is already 0.25 — scripts/ci/pf-version.sh derives it as one
minor ahead of the latest stable tag — so this is the version canary has been
publishing against all along.

Wire protocol stays at 2: every addition this cycle is optional or
capability-gated (an optional trailing max_shard_payload on Hello, the 0x08/0x09
renegotiation pair, the 0xD1 pad-audio plane, the 0xD2 redundant desktop-audio
plane, MAX_DATAGRAM_BYTES 2048 -> 9216). C ABI moves 14 -> 16 in two steps: 15
retroactively declares the floor that guarantees the rumble policy engine's C
surface (which shipped while the constant still read 7), and 16 adds the
pad-audio surface and mirrors its two capability bits. Four new capability bits
land in the client/host bytes (audio redundancy 0x04/0x20, pad audio 0x08/0x40);
the video-caps byte was NOT touched and stays full from 0.23.0, so the standing
"next video cap needs a second byte and an ABI bump" note still holds. host_caps
is now down to its last free bit (0x80). Virtual-display driver protocol 6 and
the Windows gamepad channel 3 are untouched — pf-driver-proto is byte-for-byte
identical to v0.24.0. The generated header is in sync (ABI 16, both cap mirrors).

Breaking for C embedders: 149 unprefixed macros are now PUNKTFUNK_-prefixed
(139 #defines renamed in the checked-in header). Mechanical to fix, and it
cannot break silently — the old spellings cease to exist, so it is always an
undeclared-identifier error rather than the wrong value a colliding #define
used to produce.

Lock touched for the 32 workspace members only, via `cargo update --workspace`:
diff against origin/main is versions-only, 32 insertions and 32 deletions. Unlike
the last cut there is no third-party crate sitting on the outgoing version to
trip the count — `wasapi` is at 0.23.0 and was never a candidate. `cargo metadata
--locked` resolves (35 members; fec-rs, pf-driver-proto and usbip-sim keep their
own versions by design). `cargo fmt --all --check` clean in both the main and
packaging/windows/drivers workspaces. Doc lazy-continuation scanner: 0 hits over
521 files — that regex is the exact defect that made the first v0.23.0 tag go red
on Windows clippy, and no Windows leg runs on a main push, so main being green
proves nothing about the tag fan-out.

api/openapi.json is deliberately left at 0.23.0: it tracks API edits and lags,
as in every prior cut. It is now two releases behind and worth a look.

Notes at docs/releases/v0.25.0.md, per docs/releases/README.md — authored with
the bump so CI's ensure_release seeds the release body at tag creation. Body
voice checked programmatically: 0 internal-vocabulary hits above `## Under the
hood`. Play's "What's new" at docs/releases/whatsnew/v0.25.0.txt (494/500 chars),
verified by running android.yml's gate logic verbatim against it, including the
byte-identical-to-another-release check.

Cargo.lock

@@ -947,7 +947,7 @@ dependencies = [
 
 [[package]]
 name = "cursor-probe"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "pf-capture",
@@ -1036,7 +1036,7 @@ dependencies = [
 
 [[package]]
 name = "display-disturb"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "windows 0.62.2 (registry+https://github.com/rust-lang/crates.io-index)",
 ]
@@ -2221,7 +2221,7 @@ dependencies = [
 
 [[package]]
 name = "latency-probe"
-version = "0.24.0"
+version = "0.25.0"
 
 [[package]]
 name = "lazy_static"
@@ -2326,7 +2326,7 @@ dependencies = [
 
 [[package]]
 name = "libvpl-sys"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "bindgen",
  "cmake",
@@ -2361,7 +2361,7 @@ checksum = "0ceec5bc11778974d1bcb055b18002eba7f4b3518b6a0081b3af5f21666da9ad"
 
 [[package]]
 name = "loss-harness"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "punktfunk-core",
 ]
@@ -2850,7 +2850,7 @@ checksum = "9b4f627cb1b25917193a259e49bdad08f671f8d9708acfd5fe0a8c1455d87220"
 
 [[package]]
 name = "pf-capture"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ashpd",
@@ -2871,7 +2871,7 @@ dependencies = [
 
 [[package]]
 name = "pf-client-core"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ash",
@@ -2893,11 +2893,12 @@ dependencies = [
  "ureq",
  "wasapi",
  "windows 0.62.2 (git+https://github.com/microsoft/windows-rs?rev=acb5a1a7441033d9312b16842af02eb0c2b403dc)",
+ "winreg",
 ]
 
 [[package]]
 name = "pf-clipboard"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ashpd",
@@ -2915,7 +2916,7 @@ dependencies = [
 
 [[package]]
 name = "pf-console-ui"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ash",
@@ -2936,7 +2937,7 @@ dependencies = [
 
 [[package]]
 name = "pf-encode"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ash",
@@ -2960,7 +2961,7 @@ dependencies = [
 
 [[package]]
 name = "pf-ffvk"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "ash",
  "bindgen",
@@ -2969,7 +2970,7 @@ dependencies = [
 
 [[package]]
 name = "pf-frame"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "libc",
@@ -2981,7 +2982,7 @@ dependencies = [
 
 [[package]]
 name = "pf-gpu"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "pf-host-config",
@@ -2995,11 +2996,11 @@ dependencies = [
 
 [[package]]
 name = "pf-host-config"
-version = "0.24.0"
+version = "0.25.0"
 
 [[package]]
 name = "pf-inject"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ashpd",
@@ -3028,14 +3029,14 @@ dependencies = [
 
 [[package]]
 name = "pf-paths"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "tracing",
 ]
 
 [[package]]
 name = "pf-presenter"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ash",
@@ -3050,7 +3051,7 @@ dependencies = [
 
 [[package]]
 name = "pf-update"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "serde",
  "serde_json",
@@ -3058,7 +3059,7 @@ dependencies = [
 
 [[package]]
 name = "pf-update-check"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "base64",
@@ -3070,7 +3071,7 @@ dependencies = [
 
 [[package]]
 name = "pf-vdisplay"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ashpd",
@@ -3103,7 +3104,7 @@ dependencies = [
 
 [[package]]
 name = "pf-win-display"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "pf-paths",
@@ -3115,7 +3116,7 @@ dependencies = [
 
 [[package]]
 name = "pf-zerocopy"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ash",
@@ -3323,7 +3324,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-cli"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "pf-client-core",
  "punktfunk-core",
@@ -3334,7 +3335,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-client-android"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "android_logger",
  "jni",
@@ -3346,11 +3347,13 @@ dependencies = [
  "opus",
  "punktfunk-core",
  "tracing",
+ "uac-host",
+ "usbfs-iso",
 ]
 
 [[package]]
 name = "punktfunk-client-linux"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "async-channel",
@@ -3367,7 +3370,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-client-session"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "pf-client-core",
@@ -3382,7 +3385,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-client-windows"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "async-channel",
  "ffmpeg-next",
@@ -3402,7 +3405,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-core"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "aes-gcm",
  "bytes",
@@ -3434,7 +3437,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-host"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "aes",
  "aes-gcm",
@@ -3519,7 +3522,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-probe"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "mdns-sd",
@@ -3533,7 +3536,7 @@ dependencies = [
 
 [[package]]
 name = "punktfunk-tray"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "anyhow",
  "ksni",
@@ -3556,7 +3559,7 @@ checksum = "d55d956fa96f5ec02be2e13af0e20391a5aa83d6a074e3ad368959d0fab299ea"
 
 [[package]]
 name = "pyrowave-sys"
-version = "0.24.0"
+version = "0.25.0"
 dependencies = [
  "bindgen",
  "cmake",
@@ -4985,6 +4988,14 @@ version = "1.20.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20"
 
+[[package]]
+name = "uac-host"
+version = "0.1.0"
+source = "git+https://github.com/unom-io/usbfs-iso?rev=f3de1fd62cec271d07f45664dc464f23e423e721#f3de1fd62cec271d07f45664dc464f23e423e721"
+dependencies = [
+ "usbfs-iso",
+]
+
 [[package]]
 name = "uds_windows"
 version = "1.2.1"
@@ -5064,6 +5075,14 @@ dependencies = [
  "serde",
 ]
 
+[[package]]
+name = "usbfs-iso"
+version = "0.1.0"
+source = "git+https://github.com/unom-io/usbfs-iso?rev=f3de1fd62cec271d07f45664dc464f23e423e721#f3de1fd62cec271d07f45664dc464f23e423e721"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "usbip-sim"
 version = "0.8.0"

Cargo.toml

@@ -53,7 +53,7 @@ exclude = [
 ndk = { path = "clients/android/native/vendor/ndk" }
 
 [workspace.package]
-version = "0.24.0"
+version = "0.25.0"
 edition = "2021"
 rust-version = "1.82"
 license = "MIT OR Apache-2.0"

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

@@ -410,17 +410,68 @@ private fun DsRow(usbDev: android.hardware.usb.UsbDevice) {
                         Text("Grant USB access")
                     }
                 }
-                else -> Text(
-                    if (model == DsDevice.Model.DUALSHOCK4) {
-                        "Ready — captured at stream start: rumble, lightbar and gyro are " +
-                            "driven directly."
-                    } else {
-                        "Ready — captured at stream start: rumble, adaptive triggers, lightbar " +
-                            "and gyro are driven directly."
-                    },
-                    style = MaterialTheme.typography.bodySmall,
-                    color = MaterialTheme.colorScheme.onSurfaceVariant,
-                )
+                else -> {
+                    Text(
+                        if (model == DsDevice.Model.DUALSHOCK4) {
+                            "Ready — captured at stream start: rumble, lightbar and gyro are " +
+                                "driven directly."
+                        } else {
+                            "Ready — captured at stream start: rumble, adaptive triggers, lightbar " +
+                                "and gyro are driven directly."
+                        },
+                        style = MaterialTheme.typography.bodySmall,
+                        color = MaterialTheme.colorScheme.onSurfaceVariant,
+                    )
+                    // Pad-audio self test. Deliberately reachable WITHOUT a stream: it exists to
+                    // answer "can this phone drive this pad's audio endpoint at all", and gating
+                    // that behind a live session would make it depend on the very thing one wants
+                    // to rule out when a session misbehaves. DualSense only — the DS4 has no
+                    // 4-channel haptics device.
+                    if (model != DsDevice.Model.DUALSHOCK4) {
+                        var testing by remember { mutableStateOf(false) }
+                        var result by remember { mutableStateOf<String?>(null) }
+                        result?.let {
+                            Text(
+                                it,
+                                style = MaterialTheme.typography.bodySmall,
+                                color = MaterialTheme.colorScheme.onSurfaceVariant,
+                            )
+                        }
+                        OutlinedButton(
+                            enabled = !testing,
+                            onClick = {
+                                testing = true
+                                result = null
+                                Thread({
+                                    // Its OWN connection: the renderer's descriptor must never be
+                                    // shared with another transfer engine, and that applies to
+                                    // this test as much as to the real path.
+                                    val conn = runCatching { usbManager.openDevice(usbDev) }.getOrNull()
+                                    val fd = conn?.fileDescriptor ?: -1
+                                    val r = if (fd >= 0) {
+                                        io.unom.punktfunk.kit.NativeBridge.nativePadAudioSelfTest(fd, 3, 60)
+                                    } else {
+                                        -1
+                                    }
+                                    conn?.close()
+                                    val msg = when {
+                                        r > 0 -> "Haptics test passed — $r frames to the pad."
+                                        r == -1 -> "Could not open the pad's audio interface. " +
+                                            "Some kernels refuse it; the pad still works normally."
+                                        r == -2 -> "The audio stream stopped part-way."
+                                        else -> "The stream opened but no audio reached the pad."
+                                    }
+                                    android.os.Handler(android.os.Looper.getMainLooper()).post {
+                                        result = msg
+                                        testing = false
+                                    }
+                                }, "pf-pad-selftest-ui").start()
+                            },
+                        ) {
+                            Text(if (testing) "Testing…" else "Test haptics")
+                        }
+                    }
+                }
             }
         }
     }

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

@@ -84,6 +84,9 @@ suspend fun connectToHost(
             // The host's approval-list / trust-store label for this device — the same
             // Build.MODEL convention the pairing dialogs use for nativePair.
             Build.MODEL ?: "Android",
+            // Tier-A pad audio: ask for the 0xD1 plane only when a setting would render it, so a
+            // user with it off does not make the host provision endpoints it will never feed.
+            settings.padHaptics || settings.padSpeaker,
         )
     }
 }

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

@@ -170,6 +170,26 @@ data class Settings(
      */
     val dsCapture: Boolean = true,
 
+    /**
+     * Render the host's DualSense **voice-coil haptics** on a captured USB pad (tier A).
+     *
+     * The pad's own 4-channel audio device carries them, driven directly over usbfs — Android's
+     * audio framework denylists that device by VID/PID, so there is no supported route to it. The
+     * two kinds are arbitrated rather than mixed, and on evidence: wire rumble is suppressed only
+     * while haptics frames are actually arriving, so a title that drives classic rumble and sends
+     * no haptics audio keeps rumbling. Off, or on an uncaptured/Bluetooth pad, the pad stays on
+     * ordinary rumble (tier C), which on this client already drives the same actuators.
+     */
+    val padHaptics: Boolean = true,
+
+    /**
+     * Render the pad's **built-in speaker** on a captured USB pad. Independent of [padHaptics] —
+     * the host sends the two as separate streams and either can play alone. Off by default: the
+     * speaker is a small, easily-startling loudspeaker in the user's hands, and unlike haptics it
+     * duplicates audio they are already hearing.
+     */
+    val padSpeaker: Boolean = false,
+
     /**
      * How a physical mouse drives the host — the cross-client mouse model (see [MouseMode]).
      * [MouseMode.DESKTOP] (default here) points absolutely; [MouseMode.CAPTURE] locks the pointer
@@ -271,6 +291,8 @@ class SettingsStore(context: Context) {
         rumbleOnPhone = prefs.getBoolean(K_RUMBLE_ON_PHONE, false),
         sc2Capture = prefs.getBoolean(K_SC2_CAPTURE, true),
         dsCapture = prefs.getBoolean(K_DS_CAPTURE, true),
+        padHaptics = prefs.getBoolean(K_PAD_HAPTICS, true),
+        padSpeaker = prefs.getBoolean(K_PAD_SPEAKER, false),
         mouseMode = prefs.getString(K_MOUSE_MODE, null)
             ?.let { name -> MouseMode.entries.firstOrNull { it.storedName == name } }
             // Migration: the pre-enum Boolean "pointer_capture" (true = lock the pointer). Its
@@ -308,6 +330,8 @@ class SettingsStore(context: Context) {
             .putBoolean(K_RUMBLE_ON_PHONE, s.rumbleOnPhone)
             .putBoolean(K_SC2_CAPTURE, s.sc2Capture)
             .putBoolean(K_DS_CAPTURE, s.dsCapture)
+            .putBoolean(K_PAD_HAPTICS, s.padHaptics)
+            .putBoolean(K_PAD_SPEAKER, s.padSpeaker)
             .putString(K_MOUSE_MODE, s.mouseMode.storedName)
             .putBoolean(K_INVERT_SCROLL, s.invertScroll)
             .apply()
@@ -355,6 +379,8 @@ class SettingsStore(context: Context) {
         const val K_RUMBLE_ON_PHONE = "rumble_on_phone"
         const val K_SC2_CAPTURE = "sc2_capture"
         const val K_DS_CAPTURE = "ds_capture"
+        const val K_PAD_HAPTICS = "pad_haptics"
+        const val K_PAD_SPEAKER = "pad_speaker"
         const val K_MOUSE_MODE = "mouse_mode"
 
         /** Legacy Boolean the [K_MOUSE_MODE] enum replaced — read once for migration, never written. */

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

@@ -896,6 +896,22 @@ private fun ControllerSettings(s: Settings, update: (Settings) -> Unit, onOpenCo
                 enabled = s.gamepadForwarding,
                 onCheckedChange = { on -> update(s.copy(dsCapture = on)) },
             )
+            // Both only ever apply to a captured pad, so they follow that row and gate on it.
+            ToggleRow(
+                title = "Controller haptics",
+                subtitle = "Play the host's fine-grained DualSense haptics on the pad itself — " +
+                    "the pad keeps ordinary rumble for games that don't send them",
+                checked = s.padHaptics,
+                enabled = s.gamepadForwarding && s.dsCapture,
+                onCheckedChange = { on -> update(s.copy(padHaptics = on)) },
+            )
+            ToggleRow(
+                title = "Controller speaker",
+                subtitle = "Play audio the game sends to the controller's own speaker",
+                checked = s.padSpeaker,
+                enabled = s.gamepadForwarding && s.dsCapture,
+                onCheckedChange = { on -> update(s.copy(padSpeaker = on)) },
+            )
         }
     }
 }

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

@@ -507,6 +507,28 @@ fun StreamScreen(session: ActiveSession, onDisconnect: () -> Unit) {
         var dsUsbReceiver: BroadcastReceiver? = null
         if (ds != null) {
             feedback.sink = ds
+            // Tier-A pad audio: render the host's 0xD1 streams on the pad's own 4-channel USB
+            // audio device. Bound here rather than inside DsCapture because the session handle
+            // lives at this layer; DsCapture decides WHEN (it knows the wire index and the link
+            // lifetime), this decides WHETHER.
+            if (initialSettings.padHaptics || initialSettings.padSpeaker) {
+                ds.padAudio = object : DsCapture.PadAudioHook {
+                    override fun start(pad: Int, fd: Int) {
+                        val ok = NativeBridge.nativeStartPadAudio(
+                            handle,
+                            pad,
+                            fd,
+                            initialSettings.padHaptics,
+                            initialSettings.padSpeaker,
+                        )
+                        Log.i("punktfunk", "pad audio on pad $pad: ${if (ok) "started" else "unavailable"}")
+                    }
+
+                    // Returns only once the render thread is joined — DsCapture calls this before
+                    // closing the connection whose descriptor that thread borrows.
+                    override fun stop(pad: Int) = NativeBridge.nativeStopPadAudio(handle, pad)
+                }
+            }
             val usbManager = context.getSystemService(Context.USB_SERVICE) as UsbManager
             val usbDev = ds.findUsbDevice()
             when {

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

@@ -23,8 +23,9 @@ import android.view.InputDevice
  * Input: parse ([DsDevice.parseState]) → typed mirror on an [GamepadRouter.ExternalPad] (buttons
  * diffed, axes on-change — the exit chord participates like any pad) + the rich plane (touch
  * normalized to the wire's 0..65535 screen space on-change; motion forwarded per report in raw
- * device units, the wire's contract). The wire slot is claimed lazily on the FIRST parsed report
- * and freed on unplug/[stop], so indices never leak.
+ * device units, the wire's contract). The wire slot is claimed when the capture engages, with the
+ * first parsed report as the fallback for a claim that found no free index, and freed on
+ * unplug/[stop], so indices never leak.
  *
  * Feedback: implements [GamepadFeedback.PadFeedbackSink] — rumble / trigger / lightbar / player
  * LED events addressed to this pad's wire index become USB output reports on the physical pad
@@ -78,6 +79,33 @@ class DsCapture(
     @Volatile
     var onActiveChanged: ((active: Boolean) -> Unit)? = null
 
+    /**
+     * Tier-A pad audio, bound by the app layer (which owns the session handle).
+     *
+     * [start] is called once the router has assigned this pad a wire index, which the host uses to
+     * address the `0xD1` stream. [stop] is called **before** the USB link closes — on [stop] and on
+     * unplug alike — and must not return until nothing is still writing to the descriptor.
+     */
+    interface PadAudioHook {
+        fun start(pad: Int, fd: Int)
+        fun stop(pad: Int)
+    }
+
+    @Volatile
+    var padAudio: PadAudioHook? = null
+
+    /** True once [PadAudioHook.start] has run for the current capture, so it fires exactly once. */
+    @Volatile private var padAudioStarted = false
+
+    /**
+     * The renderer's OWN connection to the pad.
+     *
+     * It must not share [usb]'s descriptor: two transfer engines on one usbfs descriptor reap each
+     * other's completions (see [HidUsbLink.openAuxConnection]), which strands both the HID reader
+     * and the audio ring. Closed only after the hook's stop has returned.
+     */
+    @Volatile private var padAudioConn: android.hardware.usb.UsbDeviceConnection? = null
+
     val isActive: Boolean get() = model != null
 
     /** First attached Sony USB pad, for the permission flow. Needs no permission to enumerate. */
@@ -105,12 +133,17 @@ class DsCapture(
         // (the same init hid-playstation/SDL send on open).
         if (m != DsDevice.Model.DUALSHOCK4) usb.writeRaw(0, DsDevice.ds5InitReport(m))
         Log.i(TAG, "Sony pad captured over USB: PID=0x%04x model=%s".format(dev.productId, m))
+        ensureSlot(m)
         onActiveChanged?.invoke(true)
         return true
     }
 
     /** Stop the link and free the wire slot (host tears the virtual pad down). Idempotent. */
     fun stop() {
+        // Before anything touches the link: the pad-audio renderer borrows this connection's
+        // descriptor, and `usb.stop()` closes it. The hook does not return until its thread is
+        // joined, so ordering this first is what makes the borrow sound.
+        stopPadAudio()
         val m = model
         if (m != null) {
             // The interfaces are about to release with the kernel driver still detached — a
@@ -136,16 +169,112 @@ class DsCapture(
     private fun onReport(report: ByteArray, len: Int) {
         val m = model ?: return
         if (!DsDevice.parseState(m, report, len, state)) return
-        val p = pad ?: router.openExternal(m.pref)?.also {
-            pad = it
-            Log.i(TAG, "captured $m → wire pad ${it.index}")
-        } ?: return // all 16 wire indices taken — drop until one frees
+        // Normally claimed already, at capture time; this is the retry for a capture that engaged
+        // while every wire index was taken.
+        val p = pad ?: ensureSlot(m) ?: return // all 16 taken — drop until one frees
         mirrorTyped(p)
         mirrorRich(p, m)
     }
 
+    /**
+     * Claim this capture's wire slot and start pad audio on it. Idempotent; null when all 16
+     * indices are taken.
+     *
+     * Claimed when the capture engages rather than on the first report, because a pad that reports
+     * nothing is still a pad: with the lazy claim, a captured-but-silent pad left the host with no
+     * arrival, hence no virtual pad, no pad-audio capability and so no `0xD1` — a renderer sitting
+     * at zero frames, indistinguishable from a broken pipeline (it took a physical replug to
+     * clear). Callable from the main thread (capture start) and the link thread (the fallback).
+     */
+    @Synchronized
+    private fun ensureSlot(m: DsDevice.Model): GamepadRouter.ExternalPad? {
+        pad?.let { return it }
+        val p = router.openExternal(m.pref) ?: return null
+        pad = p
+        Log.i(TAG, "captured $m → wire pad ${p.index}")
+        // The wire index exists from here on, and the host addresses pad audio by it.
+        startPadAudio(p.index)
+        return p
+    }
+
+    /** Hand the renderer its own descriptor. Caller holds the monitor; fires once per capture. */
+    private fun startPadAudio(index: Int) {
+        val hook = padAudio ?: return
+        if (padAudioStarted) return
+        // A dedicated connection, NOT usb.fileDescriptor — see padAudioConn.
+        val conn = usb.openAuxConnection()
+        val fd = conn?.fileDescriptor ?: -1
+        if (fd < 0) {
+            conn?.close()
+            Log.w(TAG, "pad audio: could not open a second USB connection")
+            return
+        }
+        padAudioConn = conn
+        padAudioStarted = true
+        // Real-world self test, opt-in: `adb shell setprop debug.punktfunk.pad_audio_selftest 3`
+        // drives the voice coils for N seconds through the actual client path before the renderer
+        // takes over — the one check that proves the descriptor, the interface claim and the write
+        // path all work on THIS device, without needing a host to be streaming. Same convention as
+        // debug.punktfunk.force_parts.
+        val secs = runCatching {
+            Class.forName("android.os.SystemProperties")
+                .getMethod("get", String::class.java, String::class.java)
+                .invoke(null, "debug.punktfunk.pad_audio_selftest", "0") as String
+        }.getOrNull()?.toIntOrNull() ?: 0
+        if (secs > 0) {
+            // Diagnostic mode: the self test OWNS this descriptor for the capture, and the renderer
+            // must not also drive it — two engines on one usbfs descriptor reap each other's
+            // completions, which is precisely the fault this test exists to expose.
+            Thread({
+                val r = NativeBridge.nativePadAudioSelfTest(fd, secs, 60)
+                Log.i(TAG, "pad audio self-test → ${if (r > 0) "PASS ($r frames)" else "FAIL ($r)"}")
+            }, "pf-pad-selftest").start()
+        } else {
+            // B6: hand the coils back before the first haptics frame. Any rumble earlier in this
+            // session asserted HAPTICS_SELECT, which firmware-mutes them, and nothing else ever
+            // clears it — so without this the stream renders into a muted actuator and looks for
+            // all the world like the host is sending nothing.
+            restoreAudioHaptics()
+            hook.start(index, fd)
+        }
+    }
+
+    /**
+     * B6: clear the rumble/haptics-select bits so the pad's voice coils answer the audio-haptics
+     * path again. EP0-direct, like the other out-of-band writes here: this has to land even when
+     * the interrupt-OUT queue is busy or draining, and it is idempotent.
+     */
+    private fun restoreAudioHaptics() {
+        val m = model ?: return
+        if (m == DsDevice.Model.DUALSHOCK4) return // no voice coils, no audio-haptics path
+        if (!usb.writeControl(DsDevice.ds5AudioHapticsReport(m))) {
+            Log.w(TAG, "pad audio: could not hand the coils back to audio haptics")
+        }
+    }
+
+    /**
+     * Stop the renderer, then close the connection whose descriptor it borrows — in that order.
+     *
+     * Runs on [stop] and on unplug alike. Skipping it on unplug left the render thread writing to a
+     * descriptor whose device was gone, leaked the connection, and — because the started flag stayed
+     * set and the native tier-A registry stayed armed for that index — cost the pad both its pad
+     * audio and its wire rumble on the way back in.
+     */
+    @Synchronized
+    private fun stopPadAudio() {
+        if (!padAudioStarted) return
+        padAudioStarted = false
+        // The hook's stop joins the render thread, so nothing is using the descriptor once it
+        // returns — only then is it safe to close the connection that owns it.
+        pad?.let { padAudio?.stop(it.index) }
+        padAudioConn?.close()
+        padAudioConn = null
+    }
+
     private fun onLinkClosed() {
         Log.i(TAG, "Sony USB link closed (unplug)")
+        // Before releaseSlot(), which forgets the wire index the renderer is addressed by.
+        stopPadAudio()
         disarmBackstop()
         val wasActive = model != null
         model = null
@@ -238,6 +367,10 @@ class DsCapture(
             // write — as this used to — meant a discarded stop left the motors running with
             // nothing scheduled to try again; a USB pad holds its last level until told zero.
             if (sent) disarmBackstop() else armBackstop(STOP_RETRY_MS)
+            // B6: the stop report just re-asserted HAPTICS_SELECT on its way past, so if a
+            // haptics stream is live the coils it drives were muted by the very write that
+            // silenced the motors. Give them back.
+            if (sent && padAudioStarted) restoreAudioHaptics()
         }
     }
 

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

@@ -276,6 +276,21 @@ object DsDevice {
      * the classic compat-vibration path AND `VIBRATION2` (firmware ≥ 2.24's full-range replot;
      * older firmware ignores the unknown flag2 bit) — the host parser accepts either.
      */
+    /**
+     * B6: hand the voice coils back to the audio-haptics path.
+     *
+     * Every [ds5RumbleReport] asserts `HAPTICS_SELECT` (flag0 bit1), which is SDL's
+     * "disable audio haptics" bit — the firmware mutes the coils the 0xD1 haptics stream drives.
+     * Until now NOTHING ever cleared it again, so a single rumble anywhere in a session left tier-A
+     * haptics silent for the rest of that pad's life, with no error and nothing in a log.
+     *
+     * The undo is a report whose flag0 has BOTH bits clear (SDL's own comment: "Leaving emulated
+     * rumble bits off will restore audio haptics"). No other valid flag is set, so nothing else
+     * about the pad's state is touched. Mirrors `Ds5Feedback::audio_haptics_packet` on the desktop
+     * client, which is the same packet one transport over.
+     */
+    fun ds5AudioHapticsReport(model: Model): ByteArray = newDs5(model)
+
     fun ds5RumbleReport(model: Model, low: Int, high: Int): ByteArray = newDs5(model).also {
         it[1] = (DS5_FLAG0_COMPAT_VIBRATION or DS5_FLAG0_HAPTICS_SELECT).toByte()
         it[39] = DS5_FLAG2_VIBRATION2.toByte()

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

@@ -98,6 +98,40 @@ class HidUsbLink(
     /** First attached matching device, or null. Does not need USB permission to enumerate. */
     fun findDevice(): UsbDevice? = usb.deviceList.values.firstOrNull(config.deviceMatch)
 
+    /**
+     * Open a SECOND connection to the same device, for a consumer that needs its own descriptor.
+     *
+     * **Not a convenience — a correctness requirement.** `UsbDeviceConnection.requestWait()`
+     * returns *any* completed request on that connection, and the same is true of the usbfs reap
+     * ioctl underneath it: two independent transfer engines sharing one descriptor steal each
+     * other's completions. This link's reader owns its connection exclusively (see the note on
+     * [outQueue]), so anything else driving transfers on this device — the isochronous audio
+     * renderer — must open its own.
+     *
+     * usbfs allows the same device to be opened many times, and claims are per (descriptor,
+     * interface), so a claim made on this connection does not conflict with one made on that.
+     *
+     * The caller owns the returned connection and must close it.
+     */
+    fun openAuxConnection(): UsbDeviceConnection? {
+        val dev = device ?: return null
+        return usb.openDevice(dev)
+    }
+
+    /**
+     * The open connection's usbfs file descriptor, or -1 when the link is not running.
+     *
+     * Handed to native code that drives interfaces this link deliberately does NOT claim — the
+     * pad's isochronous audio endpoint (see `pad_audio` on the native side), which Android's own
+     * USB API cannot reach because `UsbRequest` rejects anything that is not bulk or interrupt.
+     * usbfs claims are per interface, so a native claim of the audio interface leaves this link's
+     * HID claim untouched.
+     *
+     * **The borrower must stop using it before [stop] runs**: closing the connection while a
+     * transfer is in flight pulls the descriptor out from under the kernel.
+     */
+    val fileDescriptor: Int get() = connection?.fileDescriptor ?: -1
+
     /**
      * Claim [dev]'s controller interface(s) and start the read loop. The caller has already
      * obtained USB permission. Returns false when nothing could be claimed.

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

@@ -69,6 +69,10 @@ object NativeBridge {
          *  list and trust store show for it, same convention as [nativePair]'s `name`. `null`/blank ⇒
          *  the host falls back to a fingerprint-derived "device abcd1234" label. */
         deviceName: String?,
+        /** Advertise `CLIENT_CAP_PAD_AUDIO` — the SESSION-level negotiation for the 0xD1 per-pad
+         *  DualSense plane. Without it the host never sets `HOST_CAP_PAD_AUDIO` and emits nothing,
+         *  so a captured pad's own render capabilities would have nothing to gate. */
+        padAudioOk: Boolean,
     ): Long
 
     /** 64-hex SHA-256 of the cert the host presented on [handle]; valid after a successful connect. */
@@ -332,6 +336,46 @@ object NativeBridge {
      */
     external fun nativeSetMicMuted(handle: Long, muted: Boolean)
 
+    /**
+     * Start tier-A DualSense pad audio: render the host's `0xD1` streams on the pad's own
+     * 4-channel USB audio device.
+     *
+     * [fd] is an open [android.hardware.usb.UsbDeviceConnection]'s file descriptor. Native code
+     * **borrows** it — it claims the pad's audio interface through usbfs (which leaves any HID
+     * claim on the same device alone) and never closes the descriptor. The caller must keep the
+     * connection open until [nativeStopPadAudio] returns.
+     *
+     * This also declares the pad's render capability to the host; without it no `0xD1` is sent.
+     *
+     * Returns false when there is nothing to render. A kernel that refuses the interface claim is
+     * NOT reported here — the renderer discovers that on its own thread and the session simply
+     * carries on without tier A, because some OEM kernels refuse and no app-side fix exists.
+     */
+    external fun nativeStartPadAudio(
+        handle: Long,
+        pad: Int,
+        fd: Int,
+        haptics: Boolean,
+        speaker: Boolean,
+    ): Boolean
+
+    /**
+     * Stop tier-A pad audio and join its render thread, and hand the pad back to wire rumble.
+     *
+     * Returns only once the thread is joined — so the `UsbDeviceConnection` may be closed as soon
+     * as this returns, and not before.
+     */
+    external fun nativeStopPadAudio(handle: Long, pad: Int)
+
+    /**
+     * Drive the pad with a test tone through the real render path — no host, no session.
+     *
+     * [fd] must come from a connection **nothing else is driving transfers on**: two engines on
+     * one usbfs descriptor reap each other's completions. Blocks for roughly [seconds]; run it off
+     * the main thread. Returns sample frames written, or negative on failure.
+     */
+    external fun nativePadAudioSelfTest(fd: Int, seconds: Int, hz: Int): Int
+
     /**
      * Is a mic capture actually RUNNING — i.e. did [nativeStartMic] open a stream, and has
      * [nativeStopMic] not been called since? Offer the in-stream mute control on THIS rather than

clients/android/native/Cargo.toml

@@ -64,6 +64,14 @@ libc = "0.2"
 # host + Linux client use. audiopus_sys vendors libopus (pure C) and builds it static via cmake —
 # the cargo-ndk build sets LIBOPUS_STATIC=1/LIBOPUS_NO_PKG=1 so it links the bundled lib, not the host's.
 opus = "0.3"
+# Tier-A pad audio (WP9). Android's audio framework denylists the DualSense's output by VID/PID,
+# so the pad's isochronous endpoint is driven directly on the fd `UsbDeviceConnection` hands over.
+# Our own crates, developed openly because the hole they fill — isochronous USB in Rust — is an
+# ecosystem-wide one: https://github.com/unom-io/usbfs-iso
+# Pinned by revision rather than floating: this is a transport under a real-time deadline and it
+# should move when we choose to. Becomes a plain version dependency once the crates are published.
+uac-host = { git = "https://github.com/unom-io/usbfs-iso", rev = "f3de1fd62cec271d07f45664dc464f23e423e721" }
+usbfs-iso = { git = "https://github.com/unom-io/usbfs-iso", rev = "f3de1fd62cec271d07f45664dc464f23e423e721" }
 
 [lints]
 workspace = true

clients/android/native/src/feedback.rs

@@ -77,6 +77,14 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeNextRumble(
         // handle.
         let h = unsafe { &*(handle as *const SessionHandle) };
         match h.client.next_rumble_command(PULL_TIMEOUT) {
+            // A pad whose coils are ACTIVELY being driven by the 0xD1 haptics stream must not see
+            // wire rumble: `DsDevice` sets `valid_flag0` bit 1 (`HAPTICS_SELECT`) on every rumble
+            // write, and that bit disables the audio-haptics path — so one replayed command would
+            // mute the coils the stream is driving. Gating on *arrival of haptics frames* rather
+            // than on "a stream is open" is what keeps a rumble-only title working: it renders no
+            // haptics audio, so the host emits nothing on 0xD1 and the pad keeps its rumble.
+            // Dropping it here rather than in Kotlin keeps the rule next to the reason.
+            Ok(cmd) if crate::pad_audio::haptics_owns_coils((cmd.pad & 0xF) as u8) => -1,
             Ok(cmd) => pack_rumble(cmd.pad, cmd.low, cmd.high, cmd.backstop_ms),
             Err(_) => -1, // NoFrame (timeout) or Closed — Kotlin loops on its running flag
         }
@@ -174,6 +182,11 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeNextHidout(
                 out[3..n].copy_from_slice(&data);
                 n
             }
+            HidOutput::AudioCtl { .. } => {
+                // DS5 pad-audio routing/volumes — no Android replay path yet (the 0xD1 sample
+                // plane isn't rendered here either); drop it like TrackpadHaptic.
+                return -1;
+            }
         };
         n as jint
     })

clients/android/native/src/lib.rs

@@ -37,6 +37,8 @@ mod discovery;
 mod feedback;
 #[cfg(target_os = "android")]
 mod mic;
+/// Tier-A DualSense pad audio: the 0xD1 plane rendered on the pad's own USB endpoint.
+mod pad_audio;
 mod session;
 mod stats;
 // Ungated like `discovery`: pure `jni` + `punktfunk_core::wol` (no Android framework), so it links

clients/android/native/src/session/connect.rs

@@ -145,6 +145,7 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeConnect<'lo
     timeout_ms: jint,
     launch: JString<'local>,
     device_name: JString<'local>,
+    pad_audio_ok: jboolean,
 ) -> jlong {
     let host: String = match env.get_string(&host) {
         Ok(s) => s.into(),
@@ -268,7 +269,16 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeConnect<'lo
         // CLIENT_CAP_PHASE_LOCK is honest: the async decode loop's presenter feeds
         // report_phase (advisory in v1 — the host arms on report receipt — but the Hello
         // should say what the client does).
-        punktfunk_core::quic::CLIENT_CAP_PHASE_LOCK,
+        // CLIENT_CAP_PAD_AUDIO is the SESSION-level negotiation, separate from the per-pad
+        // arrival bits: without it the host never sets HOST_CAP_PAD_AUDIO and never emits 0xD1,
+        // so declaring a pad's render caps later would have nothing to gate. Gated on the
+        // settings so a user with pad audio off does not make the host provision endpoints.
+        punktfunk_core::quic::CLIENT_CAP_PHASE_LOCK
+            | if pad_audio_ok != 0 {
+                punktfunk_core::quic::CLIENT_CAP_PAD_AUDIO
+            } else {
+                0
+            },
         // Slice-progressive delivery, by decoder truth (Kotlin probes FEATURE_PartialFrame on
         // every decoder this device would use; `debug.punktfunk.force_parts` overrides for the
         // on-glass experiment): AU prefixes then arrive as `Frame::part` pieces and the decode
@@ -291,6 +301,8 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeConnect<'lo
                 audio: Mutex::new(None),
                 #[cfg(target_os = "android")]
                 mic: Mutex::new(None),
+                #[cfg(target_os = "android")]
+                pad_audio: Mutex::new(None),
                 // A fresh session is never muted (mute is per-session UI state, not a setting).
                 mic_muted: Arc::new(std::sync::atomic::AtomicBool::new(false)),
             };

clients/android/native/src/session/mod.rs

@@ -61,6 +61,11 @@ pub(crate) struct SessionHandle {
     audio: Mutex<Option<crate::audio::AudioPlayback>>,
     #[cfg(target_os = "android")]
     mic: Mutex<Option<crate::mic::MicCapture>>,
+    /// Tier-A DualSense pad audio (the 0xD1 plane), started by `nativeStartPadAudio` once Kotlin
+    /// has claimed the pad's audio interface and handed its descriptor over. Session-lifetime and
+    /// `Option` because a session may have no wired DualSense at all, which is the common case.
+    #[cfg(target_os = "android")]
+    pub(crate) pad_audio: Mutex<Option<crate::pad_audio::PadAudio>>,
     /// In-stream mic mute, set via `nativeSetMicMuted` and read per 10 ms frame by the mic's
     /// encode loop ([`crate::mic`]). Session-lifetime rather than per-[`crate::mic::MicCapture`]
     /// for the same reason the stats gate is: the mic stops and restarts across a surface
@@ -99,6 +104,14 @@ impl SessionHandle {
     fn stop_mic(&self) {
         let _ = self.mic.lock().unwrap().take();
     }
+
+    /// Stop pad audio. Dropping the [`crate::pad_audio::PadAudio`] joins its render thread, which
+    /// is what guarantees nothing is still writing to the descriptor when Kotlin closes the
+    /// `UsbDeviceConnection`. Idempotent.
+    #[cfg(target_os = "android")]
+    pub(crate) fn stop_pad_audio(&self) {
+        let _ = self.pad_audio.lock().unwrap().take();
+    }
 }
 
 impl Drop for SessionHandle {
@@ -108,6 +121,8 @@ impl Drop for SessionHandle {
         self.stop_audio();
         #[cfg(target_os = "android")]
         self.stop_mic();
+        #[cfg(target_os = "android")]
+        self.stop_pad_audio();
     }
 }
 

clients/android/native/src/session/planes.rs

@@ -460,6 +460,111 @@ pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeStopMic(
     })
 }
 
+/// `NativeBridge.nativeStartPadAudio(handle, pad, fd, haptics, speaker): Boolean` — start tier-A
+/// DualSense pad audio on a descriptor Kotlin has already obtained.
+///
+/// `fd` comes from `UsbDeviceConnection.getFileDescriptor()` **after** claiming the pad's audio
+/// streaming interface. Kotlin owns that connection and **must keep it open until
+/// `nativeStopPadAudio` returns**: the renderer borrows the descriptor and never closes it, so
+/// closing early would pull it out from under an in-flight isochronous transfer.
+///
+/// Returns `false` when there is nothing to render (both kinds disabled) or the thread would not
+/// start. A kernel that refuses the interface claim is NOT reported here — the renderer discovers
+/// that on its own thread and degrades to tier C, because some OEM kernels refuse and there is no
+/// app-side fix worth blocking a session on.
+#[no_mangle]
+#[cfg(target_os = "android")]
+pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeStartPadAudio(
+    _env: JNIEnv,
+    _this: JObject,
+    handle: jlong,
+    pad: jni::sys::jint,
+    fd: jni::sys::jint,
+    haptics: jboolean,
+    speaker: jboolean,
+) -> jboolean {
+    jni_guard(0, || {
+        if handle == 0 || fd < 0 || !(0..16).contains(&pad) {
+            return 0;
+        }
+        // SAFETY: live handle per the nativeConnect/nativeClose contract.
+        let h = unsafe { &*(handle as *const SessionHandle) };
+        // Replace any previous renderer first: dropping it joins the old thread, so two of them
+        // can never hold the same descriptor at once.
+        h.stop_pad_audio();
+        // The capability declaration and the rumble suppression are NOT done here: the renderer
+        // makes both only once its USB stream actually opens (see `pad_audio::render`). Doing them
+        // at spawn time would, on a kernel that refuses the interface claim, take the pad off wire
+        // rumble and give it nothing in return — no haptics of any kind.
+        match crate::pad_audio::start(
+            std::sync::Arc::clone(&h.client),
+            pad as u8,
+            fd,
+            haptics != 0,
+            speaker != 0,
+        ) {
+            Some(p) => {
+                *h.pad_audio.lock().unwrap() = Some(p);
+                1
+            }
+            None => 0,
+        }
+    })
+}
+
+/// `NativeBridge.nativePadAudioSelfTest(fd, seconds, hz): Int` — drive the pad directly with a
+/// tone through the real client render path, with no host and no session involved.
+///
+/// The check a standalone harness cannot make: it owns its descriptor by construction, so it can
+/// never reveal that the client handed the renderer a descriptor something else was already
+/// driving. Returns sample frames written, or negative on failure (see `pad_audio::SelfTest`).
+#[no_mangle]
+#[cfg(target_os = "android")]
+pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativePadAudioSelfTest(
+    _env: JNIEnv,
+    _this: JObject,
+    fd: jni::sys::jint,
+    seconds: jni::sys::jint,
+    hz: jni::sys::jint,
+) -> jni::sys::jint {
+    jni_guard(-1, || {
+        if fd < 0 {
+            return -1;
+        }
+        // SAFETY: Kotlin holds the owning UsbDeviceConnection open across this call and drives no
+        // other transfers on it (it opens a dedicated connection for exactly this).
+        unsafe { crate::pad_audio::self_test(fd, seconds, hz) }
+    })
+}
+
+/// `NativeBridge.nativeStopPadAudio(handle, pad)` — stop tier-A pad audio and join its thread.
+///
+/// Returns only once the render thread is joined, which is the point: Kotlin may close the
+/// `UsbDeviceConnection` as soon as this returns and not before.
+#[no_mangle]
+#[cfg(target_os = "android")]
+pub extern "system" fn Java_io_unom_punktfunk_kit_NativeBridge_nativeStopPadAudio(
+    _env: JNIEnv,
+    _this: JObject,
+    handle: jlong,
+    pad: jni::sys::jint,
+) {
+    jni_guard((), || {
+        if handle != 0 {
+            // SAFETY: live handle per the nativeConnect/nativeClose contract.
+            let h = unsafe { &*(handle as *const SessionHandle) };
+            h.stop_pad_audio();
+            if (0..16).contains(&pad) {
+                // Withdraw the capability and hand the pad back to wire rumble, in that order:
+                // the host stops sending 0xD1 before tier C resumes, so the two never overlap.
+                h.client.set_pad_audio_caps(pad as u8, 0);
+                crate::pad_audio::set_tier_a(pad as u8, false);
+                crate::pad_audio::clear_haptics_liveness(pad as u8);
+            }
+        }
+    })
+}
+
 /// `NativeBridge.nativeSetMicMuted(handle, muted)` — mute/unmute the mic uplink mid-stream.
 ///
 /// Muting deliberately does NOT stop the capture: the AAudio input stream, the input-preset rung

clients/session/src/main.rs

@@ -286,6 +286,12 @@ mod session_main {
         // Spawned at first params-build so it exists for --connect AND console launches.
         #[cfg(unix)]
         crate::ctl_socket::spawn(gamepad.clone());
+        // Pad-audio prefs to OUR gamepad service (same reasoning as the pin above): tier-A
+        // slots declare their render caps at open time, which happens on attach — after this.
+        gamepad.set_pad_audio_prefs(
+            settings.pad_haptics,
+            pf_client_core::pad_audio::speaker_active(&settings.pad_speaker),
+        );
         let mode = Mode {
             width: if settings.width == 0 {
                 native.width
@@ -389,6 +395,11 @@ mod session_main {
             cursor_forward: settings.mouse_mode() == trust::MouseMode::Desktop,
             mic_enabled: settings.mic_enabled,
             echo_cancel: settings.echo_cancel,
+            // Pad audio (0xD1): the DualSense haptics/speaker render settings. The gamepad
+            // service learns the same prefs below so tier-A slots declare their render caps
+            // at open; the session pump gates CLIENT_CAP_PAD_AUDIO + the renderer on these.
+            pad_haptics: settings.pad_haptics,
+            pad_speaker: settings.pad_speaker.clone(),
             clipboard,
             // The Settings preference (auto → VAAPI where it exists; the presenter
             // demotes to software on boxes whose Vulkan can't import the dmabufs).

crates/pf-client-core/Cargo.toml

@@ -57,6 +57,10 @@ sdl3 = { version = "0.18", features = ["hidapi"] }
 
 [target.'cfg(windows)'.dependencies]
 wasapi = "0.23"
+# Pad-audio correlation (pad_audio.rs): the HID devnode's ContainerID and a render endpoint's
+# stamped PKEY_Device_ContainerId both live in the registry — read-only, which sidesteps COM
+# property stores entirely (the same version the host pins).
+winreg = "0.56"
 sdl3 = { version = "0.18", features = ["hidapi", "build-from-source"] }
 # D3D11VA decode (video_d3d11.rs): device/adapter selection, DXVA probes, and the shared
 # NT-handle hand-off ring. Same pinned rev as clients/windows so the workspace builds ONE

crates/pf-client-core/src/audio_wasapi.rs

@@ -98,13 +98,43 @@ pub fn devices() -> Result<(Vec<AudioDevice>, Vec<AudioDevice>)> {
 /// Settings device pickers via session main), or the OS default. A picked device that's
 /// gone (unplugged USB DAC, remote session) falls back to the default with a warning —
 /// audio keeps working, like the PipeWire twin's `target.object` behavior.
+/// Resolve an active endpoint by id WITHOUT `DeviceEnumerator::get_device`.
+///
+/// That helper builds its argument as `PCWSTR::from_raw(HSTRING::from(id).as_ptr())` — the
+/// `HSTRING` is a temporary, dropped at the end of that statement, so `GetDevice` reads freed
+/// memory and misses ids that are perfectly valid. Scanning the active collection touches only
+/// safe crate APIs, so it cannot regress the same way. (`punktfunk-host` fixes the same bug with
+/// raw COM instead; this crate cannot, because it pins a different `windows` revision than
+/// `wasapi` does, making the two `IMMDevice` types incompatible.)
+pub(crate) fn device_by_id(
+    enumerator: &DeviceEnumerator,
+    direction: &Direction,
+    id: &str,
+) -> Result<wasapi::Device> {
+    let devices = enumerator
+        .get_device_collection(direction)
+        .map_err(|e| anyhow!("enumerate {direction:?} endpoints: {e}"))?;
+    let count = devices
+        .get_nbr_devices()
+        .map_err(|e| anyhow!("endpoint count: {e}"))?;
+    for i in 0..count {
+        let dev = devices
+            .get_device_at_index(i)
+            .map_err(|e| anyhow!("endpoint {i}: {e}"))?;
+        if dev.get_id().is_ok_and(|got| got == id) {
+            return Ok(dev);
+        }
+    }
+    anyhow::bail!("no active {direction:?} endpoint with id {id}")
+}
+
 fn pick_device(
     enumerator: &DeviceEnumerator,
     direction: &Direction,
     var: &str,
 ) -> Result<wasapi::Device> {
     if let Some(id) = std::env::var(var).ok().filter(|v| !v.is_empty()) {
-        match enumerator.get_device(&id) {
+        match device_by_id(enumerator, direction, &id) {
             Ok(d) => {
                 tracing::info!(
                     var,

crates/pf-client-core/src/gamepad.rs

@@ -369,8 +369,14 @@ enum Ctl {
     Pin(Option<String>),
     KindOverride(GamepadPref),
     Forwarding(bool),
-    SystemButtons { forward_raw: bool, gesture: bool },
+    SystemButtons {
+        forward_raw: bool,
+        gesture: bool,
+    },
     TapButton(u32),
+    /// Which pad-audio streams the session's settings want rendered (bit0 = haptics, bit1 =
+    /// speaker) — the settings half of the per-pad tier-A capability declared at slot open.
+    PadAudioPrefs(u8),
     MenuMode(bool),
     MenuRumble(MenuPulse),
 }
@@ -573,6 +579,18 @@ impl GamepadService {
         let _ = self.ctl.send(Ctl::TapButton(wire::BTN_MISC1));
     }
 
+    /// Declare which pad-audio streams this session's settings want rendered (`haptics` =
+    /// [`Settings::pad_haptics`](crate::trust::Settings::pad_haptics), `speaker` =
+    /// `pad_speaker == "pad"` via [`crate::pad_audio::speaker_active`]). Drives the per-pad
+    /// tier-A capability bits declared to the core at slot open — a WIRED DualSense/Edge
+    /// declares exactly these; every other pad declares 0. Call before [`Self::attach`],
+    /// like [`Self::set_kind_override`]: slots declare at open time. Defaults to "nothing"
+    /// for an embedder that never calls it, keeping the wire bytes exactly as before.
+    pub fn set_pad_audio_prefs(&self, haptics: bool, speaker: bool) {
+        let bits = (haptics as u8) | ((speaker as u8) << 1);
+        let _ = self.ctl.send(Ctl::PadAudioPrefs(bits));
+    }
+
     pub fn attach(&self, connector: Arc<NativeClient>) {
         let _ = self.ctl.send(Ctl::Attach(connector));
     }
@@ -746,6 +764,8 @@ impl Ds5Feedback {
     /// The USB report offsets these are derived from — see the type doc. Kept beside the derived
     /// values so the subtraction is visible at the point of definition.
     const REPORT_ID_LEN: usize = 1;
+    /// The audio-control region (`ucHeadphoneVolume`…`ucAudioMuteBits`): report byte 5.
+    const AUDIO: usize = 5 - Self::REPORT_ID_LEN;
     const RIGHT_TRIGGER: usize = 11 - Self::REPORT_ID_LEN;
     const LEFT_TRIGGER: usize = 22 - Self::REPORT_ID_LEN;
     const PAD_LIGHTS: usize = 44 - Self::REPORT_ID_LEN;
@@ -782,6 +802,29 @@ impl Ds5Feedback {
         p[Self::PAD_LIGHTS] = bits & 0x1F;
         p
     }
+
+    /// The one-shot tier-A activation packet — the SDL disable-bit trap undone. `p[0]`
+    /// (`ucEnableBits1`) bit0 = "enable rumble emulation" and bit1 = "disable audio haptics"
+    /// (SDL_hidapi_ps5.c); SDL sets BOTH whenever its rumble path runs, which mutes the very
+    /// voice coils the 0xD1 haptics stream drives. Per SDL's own comment — "Leaving emulated
+    /// rumble bits off will restore audio haptics" — a packet with those bits CLEARED (and no
+    /// other valid flag, so nothing else is touched) puts the pad back on audio haptics.
+    fn audio_haptics_packet() -> [u8; 47] {
+        [0u8; 47]
+    }
+
+    /// Fold a host [`HidOutput::AudioCtl`] into an effects packet: `raw` is DS5 output report
+    /// `0x02` bytes 5..=10 verbatim → struct offsets 4..=9 ([`Self::AUDIO`] — headphone/
+    /// speaker/mic volumes + routing), and `p[0]` re-asserts the report's audio-valid flags
+    /// (`flags` bits1..4 = report `flag0` bits 4..7). `flags` bit0 (haptics-select, `flag0`
+    /// bit1 = SDL's "disable audio haptics") is deliberately NOT replayed: bits 0/1 stay
+    /// clear so the pad's audio haptics stay live (see [`audio_haptics_packet`]).
+    fn audio_ctl_packet(flags: u8, raw: &[u8; 6]) -> [u8; 47] {
+        let mut p = [0u8; 47];
+        p[0] = (flags & 0x1E) << 3;
+        p[Self::AUDIO..Self::AUDIO + 6].copy_from_slice(raw);
+        p
+    }
 }
 
 /// One forwarded controller during an attached session: the open SDL handle, its stable wire
@@ -818,6 +861,14 @@ struct Slot {
     /// Hold-Select→guide state ([`SelectGesture`]) — only fed while the worker's
     /// `guide_gesture` policy is on.
     gesture: SelectGesture,
+    /// Pad-audio render capabilities declared for this slot (bit0 = haptics, bit1 = speaker
+    /// — the [`NativeClient::set_pad_audio_caps`] bits). Nonzero only for a tier-A pad (a
+    /// WIRED DualSense/Edge, see [`crate::pad_audio::is_tier_a_ds5`]) under matching
+    /// settings; bit0 set additionally suppresses wire rumble for this slot (the SDL
+    /// disable-bit trap — see [`Worker::render_feedback`]).
+    audio_caps: u8,
+    /// The wire-rumble-suppressed notice fired for this slot (log once, not per command).
+    rumble_suppressed_logged: bool,
 }
 
 impl Slot {
@@ -834,6 +885,8 @@ impl Slot {
             held_clicks: [false; 2],
             last_accel: [0; 3],
             gesture: SelectGesture::default(),
+            audio_caps: 0,
+            rumble_suppressed_logged: false,
         }
     }
 
@@ -971,6 +1024,10 @@ struct Worker {
     /// Releases owed for synthetic taps ([`Ctl::TapButton`]): `(pad, bit, due)` — the
     /// down went out on receipt, the up goes out from the poll once `due` passes.
     synthetic_ups: Vec<(u8, u32, Instant)>,
+    /// Pad-audio streams the session's settings want rendered (bit0 = haptics, bit1 =
+    /// speaker — [`GamepadService::set_pad_audio_prefs`]). `0` (the default) until an embedder
+    /// declares some: tier-A detection then never runs and every arrival stays caps-less.
+    pad_audio_prefs: u8,
     attached: Option<Arc<NativeClient>>,
     /// Raises the UI escape signal; the escape chord fires it once per press.
     escape_tx: async_channel::Sender<()>,
@@ -1176,11 +1233,18 @@ impl Worker {
             Ok(pad) => {
                 let mut slot = Slot::new(id, index, pref, pad);
                 Self::set_slot_sensors(&mut slot, true);
+                slot.audio_caps = self.pad_audio_caps_for(id, &slot.pad);
                 // Declare this pad's kind BEFORE any of its input, so the host builds a matching
                 // virtual device (mixed types — pad 0 a DualSense, pad 1 an Xbox pad). The core
                 // re-sends it a few times against datagram loss; an older host ignores it and
                 // uses the session-default kind.
                 if let Some(c) = &self.attached {
+                    // Pad-audio render caps go in FIRST — the core ORs them into this (and
+                    // every re-sent) arrival's flags bits 8/9 toward a capable host. ALWAYS
+                    // set (0 for non-tier-A): wire indices are reused within a connection, so
+                    // a tier-A slot that closes must not leave its bits behind for the next
+                    // pad on the same index (the set_rumble_quirks rule).
+                    c.set_pad_audio_caps(index, slot.audio_caps);
                     send(
                         c,
                         InputKind::GamepadArrival,
@@ -1203,6 +1267,27 @@ impl Worker {
                     };
                     c.set_rumble_quirks(index as u16, quirks);
                 }
+                if slot.audio_caps != 0 {
+                    if slot.audio_caps & 0x01 != 0 {
+                        // Tier-A haptics activation: the SDL disable-bit trap. SDL's DS5
+                        // driver sets ucEnableBits1 0x01|0x02 ("enable rumble emulation" +
+                        // "disable audio haptics") whenever its rumble path runs — which
+                        // would MUTE the voice coils the 0xD1 stream drives. One effects
+                        // packet with those bits CLEARED puts the pad back on audio haptics
+                        // ("Leaving emulated rumble bits off will restore audio haptics" —
+                        // SDL_hidapi_ps5.c); wire rumble for this slot is suppressed in
+                        // render_feedback so SDL never re-arms them.
+                        let _ = slot.pad.send_effect(&Ds5Feedback::audio_haptics_packet());
+                    }
+                    // Hand the pad to the session's renderer worker. Windows correlation
+                    // needs the HID interface path; Linux matches the sink by signature.
+                    crate::pad_audio::register_tier_a(index, slot.pad.path());
+                    tracing::info!(
+                        index,
+                        caps = slot.audio_caps,
+                        "tier-A DualSense: pad-audio render caps declared"
+                    );
+                }
                 tracing::info!(
                     id,
                     index,
@@ -1216,6 +1301,35 @@ impl Worker {
         }
     }
 
+    /// This pad's pad-audio render capabilities (the bits [`NativeClient::set_pad_audio_caps`]
+    /// takes): the settings prefs for a tier-A pad — a physical DualSense/Edge (by VID:PID,
+    /// never the DECLARED kind: the stream renders on the controller in the user's hands) on
+    /// a WIRED connection — and `0` for everything else (tier B/C are out of scope). Wired
+    /// comes from `SDL_GetGamepadConnectionState`; when SDL answers Unknown, the pad's 4-ch
+    /// audio sibling existing is the fallback signal (Bluetooth exposes no audio device).
+    fn pad_audio_caps_for(&self, id: u32, pad: &sdl3::gamepad::Gamepad) -> u8 {
+        if self.pad_audio_prefs == 0 {
+            return 0; // nothing wanted — skip the (possibly probing) wired check entirely
+        }
+        let jid = sdl3::sys::joystick::SDL_JoystickID(id);
+        let vid = self.subsystem.vendor_for_id(jid).unwrap_or(0);
+        let pid = self.subsystem.product_for_id(jid).unwrap_or(0);
+        if !crate::pad_audio::is_tier_a_ds5(vid, pid, true) {
+            return 0; // not a DualSense/Edge — no wired check needed
+        }
+        use sdl3::joystick::ConnectionState;
+        let wired = match pad.connection_state() {
+            Ok(ConnectionState::Wired) => true,
+            Ok(ConnectionState::Wireless) => false,
+            _ => crate::pad_audio::wired_audio_sibling(pad.path().as_deref()),
+        };
+        if crate::pad_audio::is_tier_a_ds5(vid, pid, wired) {
+            self.pad_audio_prefs
+        } else {
+            0
+        }
+    }
+
     /// Flush a slot's held wire state (so nothing sticks down host-side) and drop it — closing
     /// the SDL handle. The flush only emits wire events, so it is safe even when the device is
     /// already gone (unplug).
@@ -1233,6 +1347,11 @@ impl Worker {
             send(&c, InputKind::GamepadRemove, 0, 0, self.slots[i].index);
         }
         let slot = self.slots.remove(i);
+        if slot.audio_caps != 0 {
+            // Take the pad back from the pad-audio renderer (its device-gone path then
+            // re-correlates — and finds nothing until a tier-A pad registers again).
+            crate::pad_audio::unregister_tier_a(slot.index);
+        }
         tracing::info!(
             id = slot.id,
             index = slot.index,
@@ -1654,6 +1773,7 @@ impl Worker {
                         set_valve_hidapi(false);
                     }
                 }
+                Ok(Ctl::PadAudioPrefs(bits)) => self.pad_audio_prefs = bits & 0x03,
                 Ok(Ctl::MenuMode(on)) => {
                     self.menu_mode = on;
                     if on {
@@ -1966,6 +2086,20 @@ impl Worker {
         // first; the physical silence backstop is in `close_slot_at`).
         while let Ok(cmd) = connector.next_rumble_command(Duration::ZERO) {
             if let Some(slot) = self.slots.iter_mut().find(|s| s.index as u16 == cmd.pad) {
+                // The SDL disable-bit trap: ANY SDL rumble write sets ucEnableBits1
+                // 0x01|0x02, muting the very voice coils the 0xD1 haptics stream drives —
+                // so a slot with tier-A haptics active never issues wire rumble (the stream
+                // carries the feedback; the game's rumble is in its haptics mix).
+                if slot.audio_caps & 0x01 != 0 {
+                    if !slot.rumble_suppressed_logged {
+                        slot.rumble_suppressed_logged = true;
+                        tracing::info!(
+                            pad = slot.index,
+                            "wire rumble suppressed — the pad-audio haptics stream carries feedback"
+                        );
+                    }
+                    continue;
+                }
                 Self::issue_rumble(slot, cmd.low, cmd.high, cmd.backstop_ms);
             }
         }
@@ -2003,13 +2137,27 @@ impl Worker {
                         .pad
                         .send_effect(&Ds5Feedback::trigger_packet(which, effect));
                 }
+                // The audio-control region of a DS5 output report a game wrote host-side
+                // (volumes + routing; the SAMPLES ride 0xD1) — folded back into the physical
+                // pad's effects packet, but only where a tier-A renderer is actually live
+                // (`audio_caps`): replaying speaker volumes at a pad whose audio device
+                // nothing streams to would just mute/blast a future session's start state.
+                // Non-tier-A pads keep dropping it (the pre-pad-audio behaviour).
+                HidOutput::AudioCtl { flags, raw, .. } if is_ds && slot.audio_caps != 0 => {
+                    let _ = slot
+                        .pad
+                        .send_effect(&Ds5Feedback::audio_ctl_packet(flags, &raw));
+                }
                 // Deliberately unhandled, listed rather than left to a bare `_` so a new
                 // variant cannot join them silently: adaptive triggers exist only on a
                 // DualSense, and the trackpad-haptic / raw-passthrough planes are DS-specific
-                // and carried by `send_effect` above when the pad is one.
+                // and carried by `send_effect` above when the pad is one. `AudioCtl` lands here
+                // only when the guarded arm above declined it — a non-DualSense pad, or one with
+                // no live tier-A renderer — which is the pre-pad-audio behaviour: drop it.
                 HidOutput::Trigger { .. }
                 | HidOutput::TrackpadHaptic { .. }
-                | HidOutput::HidRaw { .. } => {}
+                | HidOutput::HidRaw { .. }
+                | HidOutput::AudioCtl { .. } => {}
             }
         }
     }
@@ -2048,6 +2196,9 @@ fn hidout_pad(h: &HidOutput) -> u8 {
         | HidOutput::Trigger { pad, .. }
         | HidOutput::TrackpadHaptic { pad, .. }
         | HidOutput::HidRaw { pad, .. } => *pad,
+        // AudioCtl's pad is the plane's only u16. `HidOutput::decode` rejects anything at or
+        // above MAX_PADS (B27), so by the time one reaches here the narrowing is lossless.
+        HidOutput::AudioCtl { pad, .. } => *pad as u8,
     }
 }
 
@@ -2075,6 +2226,7 @@ impl Worker {
             system_forward: true,
             guide_gesture: false,
             synthetic_ups: Vec::new(),
+            pad_audio_prefs: 0,
             attached: None,
             escape_tx,
             disconnect_tx,
@@ -2520,6 +2672,44 @@ mod slot_tests {
             }),
             6
         );
+        // AudioCtl's wire pad is u16; the index space is 0..MAX_PADS end to end.
+        assert_eq!(
+            hidout_pad(&HidOutput::AudioCtl {
+                pad: 7,
+                flags: 0,
+                raw: [0; 6]
+            }),
+            7
+        );
+    }
+
+    /// The AudioCtl fold: the 6 raw bytes (DS5 report 0x02 bytes 5..=10) land at effect-struct
+    /// offsets 4..=9, the report's audio-valid flags (AudioCtl.flags bits1..4) come back as
+    /// p[0] bits 4..7, and the rumble-emulation / disable-audio-haptics bits (p[0] bits 0/1)
+    /// stay CLEAR — setting either would mute the voice coils the 0xD1 stream drives.
+    #[test]
+    fn audio_ctl_folds_report_bytes_into_effect_offsets() {
+        let raw = [0x50, 0x60, 0x70, 0x05, 0x11, 0x22];
+        // flags 0b1_0111: haptics-select (bit0) + audio-valid bits 1/2/4 of the condensed form.
+        let p = Ds5Feedback::audio_ctl_packet(0b1_0111, &raw);
+        assert_eq!(&p[4..10], &raw, "report bytes 5..=10 → struct 4..=9");
+        // bits1..4 (0b1011) → flag0 bits 4..7.
+        assert_eq!(p[0], 0b1011_0000);
+        assert_eq!(
+            p[0] & 0x03,
+            0,
+            "haptics-select must NOT replay into p[0] bits 0/1"
+        );
+        // Nothing else is touched: no trigger/LED enable bits, no stray bytes.
+        assert!(p[1..4].iter().all(|&b| b == 0));
+        assert!(p[10..].iter().all(|&b| b == 0));
+        // No audio-valid flags condenses to no enable bits (raw still carried verbatim).
+        let p = Ds5Feedback::audio_ctl_packet(0b0_0001, &raw);
+        assert_eq!(p[0], 0);
+        assert_eq!(&p[4..10], &raw);
+        // The tier-A activation packet is the all-clear: every enable bit off — per
+        // SDL_hidapi_ps5.c, leaving the emulated-rumble bits off restores audio haptics.
+        assert_eq!(Ds5Feedback::audio_haptics_packet(), [0u8; 47]);
     }
 }
 

crates/pf-client-core/src/lib.rs

@@ -47,6 +47,11 @@ pub mod os;
 // Client settings profiles: the override catalog + the one connect-time resolver
 // (design/client-settings-profiles.md §4). Sits beside `trust`, which owns the host records
 // the bindings live on.
+// Pad audio (the 0xD1 plane): DualSense voice-coil haptics + speaker rendered on the wired
+// physical pad's own 4-ch audio device — correlation, the per-session renderer worker, and
+// the tier-A pad registry the gamepad worker feeds it through.
+#[cfg(any(target_os = "linux", windows))]
+pub mod pad_audio;
 #[cfg(any(target_os = "linux", windows))]
 pub mod profiles;
 #[cfg(any(target_os = "linux", windows))]

crates/pf-client-core/src/session.rs

@@ -44,6 +44,14 @@ pub struct SessionParams {
     /// Run the uplink through the platform's echo cancellation ([`Settings::echo_cancel`]).
     /// Ignored when `mic_enabled` is false; `PUNKTFUNK_NO_AEC=1` overrides it off.
     pub echo_cancel: bool,
+    /// Render the host's per-pad DualSense voice-coil haptics stream (0xD1 kind 0) on a wired
+    /// physical DualSense ([`crate::trust::Settings::pad_haptics`]). With `pad_speaker` it
+    /// gates the `CLIENT_CAP_PAD_AUDIO` advertisement and the pad-audio renderer thread.
+    pub pad_haptics: bool,
+    /// Where the DualSense built-in-speaker stream (0xD1 kind 1) goes: `"pad"` | `"mix"` |
+    /// `"off"` ([`crate::trust::Settings::pad_speaker`]; `"mix"` is a TODO that renders as
+    /// off — see [`crate::pad_audio::speaker_active`]).
+    pub pad_speaker: String,
     /// Share the clipboard with this host (the per-host `KnownHost::clipboard_sync`). The
     /// bridge additionally needs the host to advertise `HOST_CAP_CLIPBOARD`.
     pub clipboard: bool,
@@ -356,6 +364,11 @@ fn pump(
             );
         }
     }
+    // Pad audio (0xD1): advertise only when the settings could render a stream — the per-pad
+    // tier-A detection at slot open (gamepad.rs) still decides which pads declare render caps
+    // on their arrivals, so this bit alone changes nothing without a wired DualSense.
+    let pad_speaker_on = crate::pad_audio::speaker_active(&params.pad_speaker);
+    let pad_audio_on = params.pad_haptics || pad_speaker_on;
     let connector = match NativeClient::connect(
         &params.host,
         params.port,
@@ -379,6 +392,11 @@ fn pump(
             0
         }) | (if params.phase_lock {
             punktfunk_core::quic::CLIENT_CAP_PHASE_LOCK
+        } else {
+            0
+            // PAD_AUDIO: the embedder can render per-pad DualSense haptics/speaker (see above).
+        }) | (if pad_audio_on {
+            punktfunk_core::quic::CLIENT_CAP_PAD_AUDIO
         } else {
             0
         }),
@@ -501,6 +519,20 @@ fn pump(
     // app-lifetime service's job (the UI attaches it on Connected). Audio runs on its own
     // thread (one puller per plane), blocking on the audio queue like the Apple client.
     let audio_thread = spawn_audio(connector.clone(), stop.clone());
+    // Pad audio (0xD1): its own drain thread (that plane's single consumer), spawned whenever
+    // the settings could render. The output device is opened LAZILY once frames actually
+    // arrive — which only happens after a tier-A pad declared render caps on its arrival — so
+    // a session without a wired DualSense costs one idle 10 ms poll loop.
+    let pad_audio_thread = pad_audio_on
+        .then(|| {
+            crate::pad_audio::spawn(
+                connector.clone(),
+                stop.clone(),
+                params.pad_haptics,
+                pad_speaker_on,
+            )
+        })
+        .flatten();
     // The shared clipboard (design/clipboard-and-file-transfer.md §5): its own thread, since
     // `next_clip` blocks and the OS clipboard calls can wait on other apps. Returns straight
     // away when the host has no clipboard capability, so spawning is unconditional.
@@ -1066,6 +1098,9 @@ fn pump(
     if let Some(t) = audio_thread {
         let _ = t.join(); // exits within its 100 ms pull timeout once `stop` is set
     }
+    if let Some(t) = pad_audio_thread {
+        let _ = t.join(); // exits within its 10 ms pull timeout once `stop` is set
+    }
     if let Some(t) = clipboard_thread {
         let _ = t.join(); // exits within its next_clip wait once `stop` is set
     }

crates/pf-client-core/src/trust.rs

@@ -1024,6 +1024,21 @@ pub struct Settings {
     /// `PUNKTFUNK_AUDIO_SOURCE`).
     #[serde(default)]
     pub mic_device: String,
+    /// Render the host's per-pad DualSense voice-coil haptics stream (the 0xD1 plane, kind 0)
+    /// on a WIRED physical DualSense's own audio device (tier A — Bluetooth pads expose no
+    /// audio device). Gates the `CLIENT_CAP_PAD_AUDIO` advertisement and the per-pad arrival
+    /// capability bit; wire rumble is suppressed for a pad whose haptics stream is live (the
+    /// stream carries the feedback — see `gamepad.rs`, the SDL disable-bit trap). Default ON:
+    /// the capable-and-agreed negotiation means it changes nothing without a capable host AND
+    /// a wired DS5. `default` so pre-existing stores load with it on.
+    #[serde(default = "default_true")]
+    pub pad_haptics: bool,
+    /// Where the DualSense built-in-speaker stream (0xD1 kind 1) is rendered: `"pad"` (default
+    /// — the physical pad's own speaker), `"mix"` (fold it into the main stream audio — a
+    /// declared TODO that renders as `"off"` today; see `pad_audio::speaker_active`), or
+    /// `"off"`. `default` so pre-existing stores load as `"pad"`.
+    #[serde(default = "default_pad_speaker")]
+    pub pad_speaker: String,
     /// Match-window resolution policy (design/midstream-resolution-resize.md D1): the
     /// stream mode follows the session window — the connect asks for the window's pixel
     /// size and a mid-session resize renegotiates the host's virtual display + encoder
@@ -1071,6 +1086,10 @@ fn default_true() -> bool {
     true
 }
 
+fn default_pad_speaker() -> String {
+    "pad".into()
+}
+
 impl Settings {
     /// The stats-overlay tier, resolving pre-tier stores: an old `show_stats = false`
     /// reads as Off, everything else as Normal (≈ what the pre-tier overlay showed).
@@ -1179,6 +1198,8 @@ impl Default for Settings {
             invert_scroll: false,
             speaker_device: String::new(),
             mic_device: String::new(),
+            pad_haptics: true,
+            pad_speaker: "pad".into(),
             match_window: false,
             last_window_w: 0,
             last_window_h: 0,

crates/pf-inject/src/inject/hidout_dedup.rs

@@ -24,14 +24,20 @@ const RENEW_EVERY: Duration = Duration::from_millis(1000);
 /// bundles rumble + lightbar + player-LEDs + adaptive-triggers into one report, so a pad that is
 /// merely *rumbling* re-sends its (unchanged) lightbar / LED / trigger state on every output report.
 /// The managers already dedup rumble; this does the same for the rich [`HidOutput`] feedback so the
-/// 0xCD plane carries only genuine changes. State (`Led` / `PlayerLeds` / `Trigger`) is deduped by
-/// value; a one-shot `TrackpadHaptic` pulse is always forwarded (each pulse must fire).
+/// 0xCD plane carries only genuine changes. State (`Led` / `PlayerLeds` / `Trigger` / `AudioCtl`)
+/// is deduped by value; a one-shot `TrackpadHaptic` pulse is always forwarded (each pulse must
+/// fire).
 #[derive(Clone, Default)]
 pub struct HidoutDedup {
     led: Option<(u8, u8, u8)>,
     player_leds: Option<u8>,
     /// Last-forwarded adaptive-trigger effect per side: `[0]` = L2, `[1]` = R2.
     trigger: [Option<Vec<u8>>; 2],
+    /// Last-forwarded audio-control state (`flags` + the raw volume/routing bytes).
+    audio_ctl: Option<(u8, [u8; 6])>,
+    /// Once-per-pad-lifetime field-diagnosis flag: set after the first forwarded `AudioCtl`
+    /// carrying the haptics-select bit was logged (cleared with the rest on (re)plug).
+    haptics_select_logged: bool,
     /// When anything was last put on the wire for this pad. `None` = nothing latched yet, so
     /// there is nothing to renew. See [`RENEW_EVERY`].
     last_sent: Option<Instant>,
@@ -123,6 +129,25 @@ impl HidoutDedup {
             }
             // One-shot haptic pulse (Steam voice-coil) — state-less, always fires.
             HidOutput::TrackpadHaptic { .. } => true,
+            HidOutput::AudioCtl { pad, flags, raw } => {
+                let v = Some((*flags, *raw));
+                if self.audio_ctl == v {
+                    false
+                } else {
+                    // Field-diagnosis signal, once per pad lifetime: a title driving the DS5's
+                    // audio haptics (not plain rumble emulation, whose all-zero audio region
+                    // never reaches here) — the trace that tells "the game does audio haptics"
+                    // apart from "the client just doesn't render them".
+                    if flags & 0x01 != 0 && !self.haptics_select_logged {
+                        self.haptics_select_logged = true;
+                        tracing::info!(
+                            "DS5 title asserted haptics-select (audio haptics) pad={pad}"
+                        );
+                    }
+                    self.audio_ctl = v;
+                    true
+                }
+            }
             // Raw as-is passthrough reports must NEVER dedup: the physical device's firmware
             // watchdogs RELY on identical periodic refreshes (Triton rumble re-sent every ~40 ms
             // against a ~50 ms safety timeout, lizard-off every ~3 s) — dropping a repeat would
@@ -302,4 +327,29 @@ mod tests {
         // The pulse stamped the clock but latched no state, so the renewal has nothing to repeat.
         assert!(d.renewals(0, t + Duration::from_millis(1000)).is_empty());
     }
+
+    /// `AudioCtl` dedups by value like the other state kinds: an identical repeat (every output
+    /// report re-sends the unchanged audio region) is dropped, a flags-only or raw-only change
+    /// forwards again, and `clear` re-arms — including the once-per-pad haptics-select log flag.
+    #[test]
+    fn audio_ctl_dedups_by_value() {
+        let mut d = HidoutDedup::default();
+        let t = Instant::now();
+        let audio = |flags, vol| HidOutput::AudioCtl {
+            pad: 0,
+            flags,
+            raw: [vol, 0, 0, 0, 0, 0],
+        };
+        // Identical twice → exactly one emission.
+        assert!(d.should_forward(&audio(0x17, 0x50), t));
+        assert!(!d.should_forward(&audio(0x17, 0x50), t));
+        // Either half changing (flags, or the raw region) forwards again.
+        assert!(d.should_forward(&audio(0x16, 0x50), t));
+        assert!(d.should_forward(&audio(0x16, 0x60), t));
+        // The other kinds' state is untouched by audio traffic.
+        assert!(d.should_forward(&HidOutput::PlayerLeds { pad: 0, bits: 1 }, t));
+        // `clear` (pad re-plug) re-arms the value dedup.
+        d.clear();
+        assert!(d.should_forward(&audio(0x16, 0x60), t));
+    }
 }

crates/pf-inject/src/inject/proto/dualsense_proto.rs

@@ -535,7 +535,7 @@ pub mod out_report {
 
 /// Parse a DualSense USB output report (`0x02`) into a [`DsFeedback`], indexed off
 /// [`out_report`]. Only the well-understood fields (motor rumble, lightbar RGB, player LEDs) are
-/// surfaced — adaptive-trigger blocks are forwarded raw for the client.
+/// surfaced — adaptive-trigger blocks and the audio-control region are forwarded raw for the client.
 ///
 /// Every field is gated on the report's valid-flags (`valid_flag0` at data[1], `valid_flag1`
 /// at data[2]) — writers only set the bits for fields they mean to change (the rest is zeroed),
@@ -592,6 +592,21 @@ pub fn parse_ds_output(pad: u8, data: &[u8], fb: &mut DsFeedback) {
             });
         }
     }
+    // The audio-control region (bytes 5..=10: headphone/speaker/mic volumes + routing), for the
+    // pad-audio path. The wire flags condense the report's audio bits: bit0 = haptics-select
+    // (flag0 BIT1 — set on every SDL rumble write too, which is why it alone never triggers an
+    // emission), bits1..4 = flag0 bits 4..7 (the audio-valid flags gating the region). Emitted
+    // whenever an audio-valid flag is present or the region carries data; downstream dedup
+    // ([`crate::hidout_dedup`]) reduces the per-report repeats to genuine changes.
+    let raw: [u8; 6] = data[5..11].try_into().unwrap();
+    if flag0 & 0xF0 != 0 || raw != [0u8; 6] {
+        let flags = ((flag0 >> 1) & 0x01) | ((flag0 >> 3) & 0x1E);
+        fb.hidout.push(HidOutput::AudioCtl {
+            pad: pad.into(),
+            flags,
+            raw,
+        });
+    }
 }
 
 #[cfg(test)]
@@ -917,6 +932,48 @@ mod tests {
         assert_eq!(*DUALSENSE_EDGE_RDESC.last().unwrap(), 0xC0);
     }
 
+    /// A 0x02 report driving the pad's audio (haptics-select + audio-valid flags + the volume/
+    /// routing bytes) surfaces an `AudioCtl` with the exact raw region and the condensed flags;
+    /// a plain rumble write (haptics-select but a silent audio region — every SDL rumble) does
+    /// NOT — that is what `parse_output_respects_valid_flags` pins with its `hidout.is_empty()`.
+    #[test]
+    fn parse_output_surfaces_audio_ctl() {
+        let mut data = vec![0u8; 48];
+        data[0] = 0x02;
+        data[1] = 0xB2; // flag0: haptics-select (BIT1) + audio-valid bits 4/5/7
+        data[5] = 0x50; // headphone volume
+        data[6] = 0x60; // speaker volume
+        data[7] = 0x70; // mic volume
+        data[8] = 0x05; // audio routing / enable bits
+        let mut fb = DsFeedback::default();
+        parse_ds_output(3, &data, &mut fb);
+        // flags: bit0 = flag0 bit1, bits1..4 = flag0 bits 4..7 (0b1011 → 0b10110).
+        assert_eq!(
+            fb.hidout,
+            vec![HidOutput::AudioCtl {
+                pad: 3,
+                flags: 0b1_0111,
+                raw: [0x50, 0x60, 0x70, 0x05, 0x00, 0x00],
+            }]
+        );
+        // A non-zero audio region with NO audio-valid flags still surfaces (dedup collapses the
+        // repeats downstream) — some writers leave stale volumes gated off; the host side wants
+        // the honest bytes either way.
+        let mut data = vec![0u8; 48];
+        data[0] = 0x02;
+        data[9] = 0x01;
+        let mut fb = DsFeedback::default();
+        parse_ds_output(0, &data, &mut fb);
+        assert_eq!(
+            fb.hidout,
+            vec![HidOutput::AudioCtl {
+                pad: 0,
+                flags: 0,
+                raw: [0, 0, 0, 0, 0x01, 0],
+            }]
+        );
+    }
+
     /// A short / wrong-id report yields nothing.
     #[test]
     fn parse_output_rejects_garbage() {

crates/pf-inject/src/inject/uhid_manager.rs

@@ -518,6 +518,7 @@ mod tests {
             index: 2,
             kind: 1,
             capabilities: 0,
+            audio_caps: 0,
         });
         assert!(m.slots.get(2).is_some());
     }

crates/punktfunk-core/src/abi.rs

@@ -670,6 +670,12 @@ pub const PUNKTFUNK_HIDOUT_TRIGGER: u8 = 3;
 /// side (0 = right pad, 1 = left pad); `effect[0..6]` packs `amplitude` / `period` / `count` as
 /// little-endian `u16`s with `effect_len = 6`. Clients without trackpad coils drop it.
 pub const PUNKTFUNK_HIDOUT_TRACKPAD_HAPTIC: u8 = 4;
+/// `PunktfunkHidOutput::kind` — the audio-control region of a DS5 output report (pad-audio
+/// routing/volumes; the audio SAMPLES arrive via [`punktfunk_connection_next_pad_audio`]).
+/// `which` = the condensed audio flags (bit0 = haptics-select, bits1..4 = the report's
+/// audio-valid flags); `effect[0..6]` = bytes 5..=10 of the report verbatim
+/// (headphone/speaker/mic volumes + routing) with `effect_len = 6`. Forwarded change-only.
+pub const PUNKTFUNK_HIDOUT_AUDIO_CTL: u8 = 5;
 /// Capacity of `PunktfunkHidOutput::effect` (the DualSense trigger parameter block).
 pub const PUNKTFUNK_HID_EFFECT_MAX: u8 = 11;
 
@@ -762,6 +768,17 @@ impl PunktfunkHidOutput {
                 out.effect_len = 6;
             }
             HidOutput::HidRaw { .. } => return None,
+            HidOutput::AudioCtl { pad, flags, raw } => {
+                // Same packing idiom as TrackpadHaptic: `which` carries the flags byte,
+                // `effect[0..6]` the raw audio region. The u16 wire pad narrows losslessly
+                // because `HidOutput::decode` refuses one at or above `input::MAX_PADS` (B27) —
+                // it is enforced there, not merely assumed here.
+                out.kind = PUNKTFUNK_HIDOUT_AUDIO_CTL;
+                out.pad = *pad as u8;
+                out.which = *flags;
+                out.effect[0..6].copy_from_slice(raw);
+                out.effect_len = 6;
+            }
         }
         Some(out)
     }
@@ -1175,6 +1192,25 @@ pub const PUNKTFUNK_HOST_CAP_CLIPBOARD: u8 = 0x02;
 /// the client keeps its pen-as-touch fallback. (Mirrors `quic::HOST_CAP_PEN`;
 /// design/pen-tablet-input.md.)
 pub const PUNKTFUNK_HOST_CAP_PEN: u8 = 0x10;
+/// Host-capability bit in [`punktfunk_connection_host_caps`]: the host can capture per-gamepad
+/// audio (DualSense voice-coil haptics + speaker) and emit it on the 0xD1 plane toward pads
+/// declared capable via [`punktfunk_connection_set_pad_audio_caps`]. Set only when the client
+/// asked via [`PUNKTFUNK_CLIENT_CAP_PAD_AUDIO`]. (Mirrors `quic::HOST_CAP_PAD_AUDIO`.)
+pub const PUNKTFUNK_HOST_CAP_PAD_AUDIO: u8 = 0x40;
+
+/// Pad-audio `kind` ([`punktfunk_connection_next_pad_audio`]): the BACK channel pair — DualSense
+/// voice-coil haptics, 5 ms Opus frames. (Mirrors `quic::PAD_AUDIO_KIND_HAPTICS`.)
+pub const PUNKTFUNK_PAD_AUDIO_KIND_HAPTICS: u8 = 0;
+/// Pad-audio `kind`: the FRONT channel pair — the controller's built-in speaker, 10 ms Opus
+/// frames. (Mirrors `quic::PAD_AUDIO_KIND_SPEAKER`.)
+pub const PUNKTFUNK_PAD_AUDIO_KIND_SPEAKER: u8 = 1;
+
+/// [`punktfunk_connection_set_pad_audio_caps`] `audio_caps` bit: the pad renders the HAPTICS
+/// stream (a real DualSense's voice coils).
+pub const PUNKTFUNK_PAD_AUDIO_CAP_HAPTICS: u8 = 0x01;
+/// [`punktfunk_connection_set_pad_audio_caps`] `audio_caps` bit: the pad renders the SPEAKER
+/// stream.
+pub const PUNKTFUNK_PAD_AUDIO_CAP_SPEAKER: u8 = 0x02;
 
 // Keep the ABI cap bits in lockstep with the wire constants (compile-time guard against drift).
 #[cfg(feature = "quic")]
@@ -1189,6 +1225,20 @@ const _: () = {
     assert!(PUNKTFUNK_HOST_CAP_GAMEPAD_STATE == crate::quic::HOST_CAP_GAMEPAD_STATE);
     assert!(PUNKTFUNK_HOST_CAP_CLIPBOARD == crate::quic::HOST_CAP_CLIPBOARD);
     assert!(PUNKTFUNK_HOST_CAP_PEN == crate::quic::HOST_CAP_PEN);
+    assert!(PUNKTFUNK_HOST_CAP_PAD_AUDIO == crate::quic::HOST_CAP_PAD_AUDIO);
+    assert!(PUNKTFUNK_CLIENT_CAP_PAD_AUDIO == crate::quic::CLIENT_CAP_PAD_AUDIO);
+    assert!(PUNKTFUNK_PAD_AUDIO_KIND_HAPTICS == crate::quic::PAD_AUDIO_KIND_HAPTICS);
+    assert!(PUNKTFUNK_PAD_AUDIO_KIND_SPEAKER == crate::quic::PAD_AUDIO_KIND_SPEAKER);
+    // The setter's caps bits are the arrival flags bits 8/9 shifted down (the wire packing
+    // `input::encode_gamepad_arrival` applies).
+    assert!(
+        (PUNKTFUNK_PAD_AUDIO_CAP_HAPTICS as u32) << 8
+            == crate::input::ARRIVAL_FLAG_PAD_AUDIO_HAPTICS
+    );
+    assert!(
+        (PUNKTFUNK_PAD_AUDIO_CAP_SPEAKER as u32) << 8
+            == crate::input::ARRIVAL_FLAG_PAD_AUDIO_SPEAKER
+    );
     assert!(PUNKTFUNK_PEN_IN_RANGE == crate::quic::PEN_IN_RANGE);
     assert!(PUNKTFUNK_PEN_TOUCHING == crate::quic::PEN_TOUCHING);
     assert!(PUNKTFUNK_PEN_BARREL1 == crate::quic::PEN_BARREL1);
@@ -1771,6 +1821,13 @@ pub const PUNKTFUNK_CLIENT_CAP_CURSOR: u8 = 0x01;
 /// forward-compatible.
 pub const PUNKTFUNK_CLIENT_CAP_PHASE_LOCK: u8 = 0x02;
 
+/// [`punktfunk_connect_ex9`] `client_caps` bit: the client understands the pad-audio plane
+/// (0xD1 — per-gamepad DualSense voice-coil haptics + speaker). The embedder MUST then drain
+/// [`punktfunk_connection_next_pad_audio`] and declare each capable pad via
+/// [`punktfunk_connection_set_pad_audio_caps`]; the host emits pad audio only when it answers
+/// with [`PUNKTFUNK_HOST_CAP_PAD_AUDIO`]. (Mirrors `quic::CLIENT_CAP_PAD_AUDIO`.)
+pub const PUNKTFUNK_CLIENT_CAP_PAD_AUDIO: u8 = 0x08;
+
 /// Shared body of [`punktfunk_connect_ex7`] / [`punktfunk_connect_ex8`]: `status_out`
 /// (nullable) is written on EVERY path — `Ok`, the mapped [`PunktfunkError`],
 /// `InvalidArg` for bad arguments, `Panic` if the connect panicked.
@@ -2315,6 +2372,117 @@ pub unsafe extern "C" fn punktfunk_connection_next_audio_pcm(
     })
 }
 
+/// Pull the next pad-audio frame (0xD1) — one Opus frame of DualSense voice-coil haptics
+/// (`kind` = [`PUNKTFUNK_PAD_AUDIO_KIND_HAPTICS`], 5 ms) or built-in-speaker audio
+/// ([`PUNKTFUNK_PAD_AUDIO_KIND_SPEAKER`], 10 ms) for gamepad `*out_pad` — waiting up to
+/// `timeout_ms`. The payload is COPIED into `buf` (no borrow-until-next-call slot); the return
+/// value is its length in bytes, `0` = nothing this poll (timeout — or a DTX/oversized frame,
+/// both of which an embedder treats the same way), `-1` = the session ended (or an invalid
+/// handle/buffer). All pads/kinds share one queue — fan out by `*out_pad`/`*out_kind` to
+/// per-actuator Opus decoders. A frame larger than `buf_len` is dropped like the timeout case
+/// (the plane is lossy by design; any real Opus frame fits a 1500-byte buffer). Only a session
+/// connected with [`PUNKTFUNK_CLIENT_CAP_PAD_AUDIO`] against a
+/// [`PUNKTFUNK_HOST_CAP_PAD_AUDIO`] host — with the pad declared via
+/// [`punktfunk_connection_set_pad_audio_caps`] — ever receives any. Drain from a dedicated
+/// thread (one puller, may run alongside the other planes' pullers).
+///
+/// # Safety
+/// `c` is a valid connection handle; the `out_*` pointers are writable (NULLs are skipped);
+/// `buf` is writable for `buf_len` bytes.
+#[cfg(feature = "quic")]
+#[no_mangle]
+pub unsafe extern "C" fn punktfunk_connection_next_pad_audio(
+    c: *mut PunktfunkConnection,
+    out_pad: *mut u8,
+    out_kind: *mut u8,
+    out_seq: *mut u32,
+    out_pts_ns: *mut u64,
+    buf: *mut u8,
+    buf_len: usize,
+    timeout_ms: u32,
+) -> i32 {
+    let r = std::panic::catch_unwind(AssertUnwindSafe(|| {
+        // SAFETY: per the ABI contract - an opaque handle from a `*_new`/`*_pair` that the caller
+        // has not yet freed, or null, which `as_mut`/`as_ref` reports as `None` and the `match`
+        // here handles.
+        let c = match unsafe { c.as_ref() } {
+            Some(c) => c,
+            None => return -1,
+        };
+        if buf.is_null() && buf_len != 0 {
+            return -1;
+        }
+        match c
+            .inner
+            .next_pad_audio(std::time::Duration::from_millis(timeout_ms as u64))
+        {
+            Some(f) => {
+                if f.opus.is_empty() || f.opus.len() > buf_len {
+                    // DTX silence (skipped like the audio-PCM path — decoding an empty payload
+                    // as loss would synthesize concealment) or doesn't fit — report "nothing
+                    // this poll" (the next_hidout HidRaw-skip precedent; truncated Opus would
+                    // be undecodable anyway).
+                    return 0;
+                }
+                // SAFETY: per the ABI contract - each out-param below is OPTIONAL, so it is null-
+                // checked before it is written; `buf` is a caller-owned writable region of
+                // `buf_len` bytes and the copy length was just bounds-checked against it.
+                unsafe {
+                    if !out_pad.is_null() {
+                        *out_pad = f.pad;
+                    }
+                    if !out_kind.is_null() {
+                        *out_kind = f.kind;
+                    }
+                    if !out_seq.is_null() {
+                        *out_seq = f.seq;
+                    }
+                    if !out_pts_ns.is_null() {
+                        *out_pts_ns = f.pts_ns;
+                    }
+                    std::ptr::copy_nonoverlapping(f.opus.as_ptr(), buf, f.opus.len());
+                }
+                f.opus.len() as i32
+            }
+            // `None` folds timeout and closed; the shutdown flag tells them apart so the
+            // embedder's plane loop can exit instead of polling a dead session forever.
+            None if c.inner.is_session_ended() => -1,
+            None => 0,
+        }
+    }));
+    r.unwrap_or(-1)
+}
+
+/// Declare wire pad `pad`'s pad-audio render capabilities (`audio_caps`: OR of
+/// [`PUNKTFUNK_PAD_AUDIO_CAP_HAPTICS`] / [`PUNKTFUNK_PAD_AUDIO_CAP_SPEAKER`]) — how a client
+/// tells the host WHICH pads can actually play the 0xD1 streams. Call at controller attach,
+/// BEFORE the pad's arrival event is sent (the [`punktfunk_connection_set_rumble_quirks`]
+/// timing): the core folds the bits into the arrival's flags (bits 8/9), and only toward a
+/// [`PUNKTFUNK_HOST_CAP_PAD_AUDIO`] host — never calling this leaves the wire bytes exactly as
+/// before. Latest-wins per pad; unknown bits are masked off.
+///
+/// # Safety
+/// `c` is a valid connection handle. Callable from any thread.
+#[cfg(feature = "quic")]
+#[no_mangle]
+pub unsafe extern "C" fn punktfunk_connection_set_pad_audio_caps(
+    c: *mut PunktfunkConnection,
+    pad: u8,
+    audio_caps: u8,
+) -> PunktfunkStatus {
+    guard(|| {
+        // SAFETY: per the ABI contract - an opaque handle from a `*_new`/`*_pair` that the caller
+        // has not yet freed, or null, which `as_mut`/`as_ref` reports as `None` and the `match`
+        // here handles.
+        let c = match unsafe { c.as_ref() } {
+            Some(c) => c,
+            None => return PunktfunkStatus::NullPointer,
+        };
+        c.inner.set_pad_audio_caps(pad, audio_caps);
+        PunktfunkStatus::Ok
+    })
+}
+
 /// 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`].
@@ -4417,3 +4585,36 @@ pub unsafe extern "C" fn punktfunk_reanchor_gate_is_holding(
         PunktfunkStatus::Ok
     })
 }
+
+#[cfg(all(test, feature = "quic"))]
+mod tests {
+    use super::*;
+
+    /// The `AudioCtl` → `PunktfunkHidOutput` mapping: kind 5, pad narrowed, `which` carries the
+    /// flags byte, `effect[0..6]` the raw audio region with `effect_len = 6` (the TrackpadHaptic
+    /// packing idiom — no struct growth, so the size guard above stays at 19).
+    #[test]
+    fn hidout_abi_maps_audio_ctl() {
+        let out = PunktfunkHidOutput::from_hid(&crate::quic::HidOutput::AudioCtl {
+            pad: 3,
+            flags: 0x17,
+            raw: [0x50, 0x60, 0x70, 0x05, 0, 0],
+        })
+        .unwrap();
+        assert_eq!(out.kind, PUNKTFUNK_HIDOUT_AUDIO_CTL);
+        assert_eq!(out.pad, 3);
+        assert_eq!(out.which, 0x17);
+        assert_eq!(out.effect_len, 6);
+        assert_eq!(out.effect[..6], [0x50, 0x60, 0x70, 0x05, 0, 0]);
+        assert_eq!(out.effect[6..], [0; 5]);
+        // A raw passthrough report still has no C representation (skipped at the pull site).
+        assert!(
+            PunktfunkHidOutput::from_hid(&crate::quic::HidOutput::HidRaw {
+                pad: 0,
+                kind: 0,
+                data: vec![0x80],
+            })
+            .is_none()
+        );
+    }
+}

crates/punktfunk-core/src/client/mod.rs

@@ -16,11 +16,13 @@ use crate::config::{CompositorPref, GamepadPref, Mode};
 use crate::error::{PunktfunkError, Result};
 use crate::input::InputEvent;
 use crate::quic::{
-    endpoint, ClipControl, ClipKind, ClipOffer, ColorInfo, HdrMeta, HidOutput, ProbeRequest,
-    RfiRequest, RichInput,
+    endpoint, ClipControl, ClipKind, ClipOffer, ColorInfo, HdrMeta, HidOutput, PadAudioFrame,
+    ProbeRequest, RfiRequest, RichInput,
 };
 use crate::session::Frame;
-use std::sync::atomic::{AtomicBool, AtomicI64, AtomicU16, AtomicU32, AtomicU64, Ordering};
+use std::sync::atomic::{
+    AtomicBool, AtomicI64, AtomicU16, AtomicU32, AtomicU64, AtomicU8, Ordering,
+};
 use std::sync::mpsc::{Receiver, RecvTimeoutError};
 use std::sync::{Arc, Mutex};
 use std::time::{Duration, Instant};
@@ -43,7 +45,7 @@ use self::control::{CtrlRequest, Negotiated};
 use self::frame_channel::{DecodeLatAcc, FrameChannel, FramePop};
 use self::planes::{
     RumbleUpdate, AUDIO_QUEUE, CLIP_EVENT_QUEUE, CURSOR_SHAPE_QUEUE, CURSOR_STATE_QUEUE,
-    HDR_META_QUEUE, HIDOUT_QUEUE, HOST_TIMING_QUEUE, RUMBLE_QUEUE,
+    HDR_META_QUEUE, HIDOUT_QUEUE, HOST_TIMING_QUEUE, PAD_AUDIO_QUEUE, RUMBLE_QUEUE,
 };
 use self::probe::ProbeState;
 use self::pump::run_pump;
@@ -122,6 +124,14 @@ pub struct NativeClient {
     rumble_sched: Arc<rumble::RumbleShared>,
     /// Inbound DualSense feedback (lightbar / player LEDs / adaptive triggers) — 0xCD datagrams.
     hidout: Mutex<Receiver<HidOutput>>,
+    /// Inbound pad audio (DualSense voice-coil haptics + speaker Opus frames) — 0xD1 datagrams.
+    /// Only a session that advertised [`quic::CLIENT_CAP_PAD_AUDIO`] against a
+    /// [`quic::HOST_CAP_PAD_AUDIO`] host ever receives any.
+    pad_audio: Mutex<Receiver<PadAudioFrame>>,
+    /// Per-pad pad-audio render capabilities (bit0 haptics, bit1 speaker), written by
+    /// [`NativeClient::set_pad_audio_caps`] and OR'd into outgoing gamepad-arrival flags
+    /// (bits 8/9) by the worker's input task — toward a `HOST_CAP_PAD_AUDIO` host only.
+    pad_audio_caps: Arc<[AtomicU8; crate::input::MAX_PADS]>,
     /// Inbound static HDR metadata (ST.2086 mastering + content light level) — 0xCE datagrams.
     hdr_meta: Mutex<Receiver<HdrMeta>>,
     /// Inbound per-AU host capture→send timings — 0xCF datagrams (the client always advertises
@@ -418,6 +428,10 @@ impl NativeClient {
         let rumble_sched = Arc::new(rumble::RumbleShared::new());
         let rumble_feed = rumble::RumbleFeed(rumble_sched.clone());
         let (hidout_tx, hidout_rx) = std::sync::mpsc::sync_channel::<HidOutput>(HIDOUT_QUEUE);
+        let (pad_audio_tx, pad_audio_rx) =
+            std::sync::mpsc::sync_channel::<PadAudioFrame>(PAD_AUDIO_QUEUE);
+        let pad_audio_caps: Arc<[AtomicU8; crate::input::MAX_PADS]> =
+            Arc::new(std::array::from_fn(|_| AtomicU8::new(0)));
         let (hdr_meta_tx, hdr_meta_rx) = std::sync::mpsc::sync_channel::<HdrMeta>(HDR_META_QUEUE);
         let (host_timing_tx, host_timing_rx) =
             std::sync::mpsc::sync_channel::<crate::quic::HostTiming>(HOST_TIMING_QUEUE);
@@ -459,6 +473,7 @@ impl NativeClient {
         let clock_offset_w = clock_offset.clone();
         let decode_lat_w = decode_lat.clone();
         let live_bitrate_w = live_bitrate.clone();
+        let pad_audio_caps_w = pad_audio_caps.clone();
         let ctrl_tx_pump = ctrl_tx.clone(); // the data-plane pump sends adaptive-FEC LossReports
         let worker = std::thread::Builder::new()
             .name("punktfunk-client".into())
@@ -508,6 +523,8 @@ impl NativeClient {
                     rumble_tx,
                     rumble_feed,
                     hidout_tx,
+                    pad_audio_tx,
+                    pad_audio_caps: pad_audio_caps_w,
                     hdr_meta_tx,
                     host_timing_tx,
                     cursor_shape_tx,
@@ -556,6 +573,8 @@ impl NativeClient {
             rumble: Mutex::new(rumble_rx),
             rumble_sched,
             hidout: Mutex::new(hidout_rx),
+            pad_audio: Mutex::new(pad_audio_rx),
+            pad_audio_caps,
             hdr_meta: Mutex::new(hdr_meta_rx),
             host_timing: Mutex::new(host_timing_rx),
             cursor_shape: Mutex::new(cursor_shape_rx),
@@ -1061,6 +1080,33 @@ impl NativeClient {
         }
     }
 
+    /// Pull the next pad-audio frame (0xD1): one Opus frame of DualSense voice-coil haptics
+    /// ([`quic::PAD_AUDIO_KIND_HAPTICS`], 5 ms) or built-in-speaker audio
+    /// ([`quic::PAD_AUDIO_KIND_SPEAKER`], 10 ms) for gamepad `pad`. All pads/kinds share the
+    /// queue — the embedder fans out by `pad`/`kind` to per-actuator Opus decoders. `None` on
+    /// timeout AND once the session ended ([`is_session_ended`](Self::is_session_ended)
+    /// distinguishes, and the plane is best-effort either way). Only a session that advertised
+    /// [`quic::CLIENT_CAP_PAD_AUDIO`] against a [`quic::HOST_CAP_PAD_AUDIO`] host — with the
+    /// pad's render caps declared via [`set_pad_audio_caps`](Self::set_pad_audio_caps) — ever
+    /// receives any. Drain on a dedicated thread like [`next_audio`](Self::next_audio); one
+    /// puller per the plane contract.
+    pub fn next_pad_audio(&self, timeout: Duration) -> Option<PadAudioFrame> {
+        self.pad_audio.lock().unwrap().recv_timeout(timeout).ok()
+    }
+
+    /// Declare wire pad `pad`'s pad-audio render capabilities: `audio_caps` bit0 = the pad can
+    /// play the HAPTICS stream (a real DualSense's voice coils), bit1 = the SPEAKER stream.
+    /// Call at controller attach, BEFORE the pad's arrival is sent (like
+    /// [`set_rumble_quirks`](Self::set_rumble_quirks)) — the worker ORs the bits into the
+    /// arrival's flags (bits 8/9), and only toward a [`quic::HOST_CAP_PAD_AUDIO`] host, so an
+    /// embedder that never calls this (or a host that can't capture pad audio) leaves the wire
+    /// bytes exactly as before. Latest-wins per pad; unknown bits are masked off.
+    pub fn set_pad_audio_caps(&self, pad: u8, audio_caps: u8) {
+        if let Some(slot) = self.pad_audio_caps.get(pad as usize) {
+            slot.store(audio_caps & 0x03, Ordering::Relaxed);
+        }
+    }
+
     /// Pull the next static HDR metadata update (ST.2086 mastering display + content light level)
     /// the host sent for an HDR session; same timeout/closed semantics as
     /// [`NativeClient::next_hidout`]. The host sends one near session start and re-sends it on

crates/punktfunk-core/src/client/planes.rs

@@ -20,6 +20,12 @@ pub(crate) type RumbleUpdate = (u16, u16, u16, Option<u16>);
 /// Same overflow discipline as rumble; the host re-sends on the next feedback change.
 pub(crate) const HIDOUT_QUEUE: usize = 32;
 
+/// Pad-audio frames (`0xD1` — DualSense voice-coil haptics + speaker) buffered for the embedder,
+/// ALL pads and kinds on one queue (the embedder fans out by `pad`/`kind`): 64 × 5 ms = 320 ms of
+/// slack on a haptics-only stream, the [`AUDIO_QUEUE`] discipline. A lagging embedder drops the
+/// newest frame (the renderer conceals the gap).
+pub(crate) const PAD_AUDIO_QUEUE: usize = 64;
+
 /// Static HDR metadata (ST.2086 mastering + content light level) buffered for the embedder. Tiny
 /// and low-rate (one on start, re-sent on mastering changes / keyframes); a small ring is ample.
 pub(crate) const HDR_META_QUEUE: usize = 8;

crates/punktfunk-core/src/client/pump.rs

@@ -50,6 +50,8 @@ pub(super) async fn run_pump(args: WorkerArgs) {
         rumble_tx,
         rumble_feed,
         hidout_tx,
+        pad_audio_tx,
+        pad_audio_caps,
         hdr_meta_tx,
         host_timing_tx,
         cursor_shape_tx,
@@ -92,9 +94,17 @@ pub(super) async fn run_pump(args: WorkerArgs) {
 
     // Input task: embedder events → uplink datagrams, with per-transition gamepad events
     // folded into idempotent seq-stamped snapshots toward a HOST_CAP_GAMEPAD_STATE host
-    // (see [`input_task`]).
+    // (see [`input_task`]). Pad-audio render caps ride arrival flags bits 8/9 ONLY toward a
+    // HOST_CAP_PAD_AUDIO host — an older host reads the whole flags word as the pad index.
     let gamepad_snapshots = host_caps & crate::quic::HOST_CAP_GAMEPAD_STATE != 0;
-    tokio::spawn(input_task::run(conn.clone(), input_rx, gamepad_snapshots));
+    let pad_audio_arrivals = host_caps & crate::quic::HOST_CAP_PAD_AUDIO != 0;
+    tokio::spawn(input_task::run(
+        conn.clone(),
+        input_rx,
+        gamepad_snapshots,
+        pad_audio_arrivals,
+        pad_audio_caps,
+    ));
 
     // Mic task: embedder Opus mic frames → 0xCB uplink datagrams (best-effort, dropped on loss).
     // Self-healing latency bound: every frame still queued once this task catches up is standing
@@ -166,6 +176,7 @@ pub(super) async fn run_pump(args: WorkerArgs) {
         rumble_tx,
         rumble_feed,
         hidout_tx,
+        pad_audio_tx,
         hdr_meta_tx,
         host_timing_tx,
         encode_lat.clone(),

crates/punktfunk-core/src/client/pump/datagram_task.rs

@@ -12,6 +12,7 @@ pub(super) async fn run(
     rumble_tx: std::sync::mpsc::SyncSender<RumbleUpdate>,
     rumble_feed: super::super::rumble::RumbleFeed,
     hidout_tx: std::sync::mpsc::SyncSender<crate::quic::HidOutput>,
+    pad_audio_tx: std::sync::mpsc::SyncSender<crate::quic::PadAudioFrame>,
     hdr_meta_tx: std::sync::mpsc::SyncSender<crate::quic::HdrMeta>,
     host_timing_tx: std::sync::mpsc::SyncSender<crate::quic::HostTiming>,
     // The ABR encode signal's accumulator (see [`EncodeLatAcc`]) — fed HERE, not off
@@ -100,6 +101,11 @@ pub(super) async fn run(
                     let _ = hidout_tx.try_send(h);
                 }
             }
+            Some(&crate::quic::PAD_AUDIO_MAGIC) => {
+                if let Some(f) = crate::quic::decode_pad_audio_datagram(&d) {
+                    let _ = pad_audio_tx.try_send(f);
+                }
+            }
             Some(&crate::quic::HDR_META_MAGIC) => {
                 if let Some(m) = crate::quic::decode_hdr_meta_datagram(&d) {
                     let _ = hdr_meta_tx.try_send(m);

crates/punktfunk-core/src/client/pump/input_task.rs

@@ -15,8 +15,16 @@ pub(super) async fn run(
     conn: quinn::Connection,
     mut input_rx: tokio::sync::mpsc::UnboundedReceiver<InputEvent>,
     gamepad_snapshots: bool,
+    // Whether the host advertised HOST_CAP_PAD_AUDIO: only then do arrivals carry the per-pad
+    // audio-render bits (flags 8/9) — an older host reads the whole flags word as the pad index,
+    // so unexpected high bits would make it drop the kind declaration entirely.
+    pad_audio: bool,
+    // Per-pad audio-render capabilities (bit0 haptics, bit1 speaker), fed by the embedder via
+    // [`NativeClient::set_pad_audio_caps`] and by arrival events already carrying the bits.
+    pad_audio_caps: std::sync::Arc<[std::sync::atomic::AtomicU8; crate::input::MAX_PADS]>,
 ) {
     use crate::input::{GamepadSnapshot, InputKind, MAX_PADS};
+    use std::sync::atomic::Ordering;
     // Touched pads only: an entry appears on the first gamepad event for that index, so the
     // refresh never conjures a virtual pad the embedder didn't drive.
     let mut pads: [Option<GamepadSnapshot>; MAX_PADS] = [None; MAX_PADS];
@@ -37,6 +45,28 @@ pub(super) async fn run(
     const ARRIVAL_RESENDS: u8 = 2;
     let mut arrival: [Option<u8>; MAX_PADS] = [None; MAX_PADS];
     let mut arrival_owed: [u8; MAX_PADS] = [0; MAX_PADS];
+    // An arrival's outgoing flags word: the pad index, plus the pad's audio-render bits (8/9)
+    // toward a HOST_CAP_PAD_AUDIO host. With no declared caps (or an older host) this is
+    // byte-identical to the plain index — the pre-pad-audio wire.
+    // B7: the caps a pad's LAST arrival actually carried. `set_pad_audio_caps` only stores into
+    // the registry — it cannot reach this task — so a declaration that lands after the arrival
+    // burst has drained (the renderer commits the trade only once its sink opens, which is well
+    // past the two 100 ms ticks) used to never reach the host at all: the client believed it had
+    // pad audio and the host emitted nothing on 0xD1, silently, forever. Comparing this against
+    // the live registry on every tick re-arms the burst by itself, with no new plumbing and no
+    // extra traffic when nothing changed.
+    let mut arrival_caps_sent: [u8; MAX_PADS] = [0; MAX_PADS];
+    let caps_now = |idx: usize| -> u8 {
+        if pad_audio {
+            pad_audio_caps[idx].load(Ordering::Relaxed)
+        } else {
+            0
+        }
+    };
+    let arrival_flags = |idx: usize| -> u32 {
+        let caps = caps_now(idx);
+        crate::input::encode_gamepad_arrival(idx as u8, caps)
+    };
     let mut refresh = tokio::time::interval(Duration::from_millis(100));
     refresh.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
     loop {
@@ -81,30 +111,56 @@ pub(super) async fn run(
                     let _ = conn.send_datagram(rem.encode().to_vec().into());
                     continue;
                 }
-                if gamepad_snapshots && ev.kind == InputKind::GamepadArrival && idx < MAX_PADS {
-                    // Remember the declared kind (`code`) and forward it, arming a re-send burst
-                    // so the host learns it before the pad's first frame even under loss.
-                    arrival[idx] = Some(ev.code as u8);
-                    arrival_owed[idx] = ARRIVAL_RESENDS;
-                    let _ = conn.send_datagram(ev.encode().to_vec().into());
-                    continue;
+                if gamepad_snapshots && ev.kind == InputKind::GamepadArrival {
+                    // The index is the LOW BYTE only — bits 8/9 may carry the pad's audio-render
+                    // caps (an embedder building raw events; the `set_pad_audio_caps` registry is
+                    // the usual source). Fold event-carried bits into the registry so the re-send
+                    // burst keeps them, then send with the negotiation-gated flags word.
+                    let (pad, ev_caps) = crate::input::decode_gamepad_arrival(ev.flags);
+                    let idx = pad as usize;
+                    if idx < MAX_PADS {
+                        if ev_caps != 0 {
+                            pad_audio_caps[idx].fetch_or(ev_caps, Ordering::Relaxed);
+                        }
+                        // Remember the declared kind (`code`) and forward it, arming a re-send
+                        // burst so the host learns it before the pad's first frame even under loss.
+                        arrival[idx] = Some(ev.code as u8);
+                        arrival_owed[idx] = ARRIVAL_RESENDS;
+                        arrival_caps_sent[idx] = caps_now(idx);
+                        let arr = crate::input::InputEvent {
+                            flags: arrival_flags(idx),
+                            ..ev
+                        };
+                        let _ = conn.send_datagram(arr.encode().to_vec().into());
+                        continue;
+                    }
                 }
                 let _ = conn.send_datagram(ev.encode().to_vec().into());
             }
             _ = refresh.tick() => {
                 for idx in 0..MAX_PADS {
+                    // B7: caps declared after the burst drained — re-announce this pad's arrival.
+                    // Only for a pad that HAS an arrival (so it is a live, declared controller),
+                    // and only when the value actually moved, so a steady session sends nothing.
+                    if arrival[idx].is_some()
+                        && arrival_owed[idx] == 0
+                        && caps_now(idx) != arrival_caps_sent[idx]
+                    {
+                        arrival_owed[idx] = ARRIVAL_RESENDS;
+                    }
                     // Re-send an owed kind declaration (independent of whether the pad has state
                     // yet — it may be idle-but-connected). Idempotent on the host.
                     if arrival_owed[idx] > 0 {
                         if let Some(kind) = arrival[idx] {
                             arrival_owed[idx] -= 1;
+                            arrival_caps_sent[idx] = caps_now(idx);
                             let arr = crate::input::InputEvent {
                                 kind: InputKind::GamepadArrival,
                                 _pad: [0; 3],
                                 code: kind as u32,
                                 x: 0,
                                 y: 0,
-                                flags: idx as u32,
+                                flags: arrival_flags(idx),
                             };
                             let _ = conn.send_datagram(arr.encode().to_vec().into());
                         } else {

crates/punktfunk-core/src/client/worker.rs

@@ -5,8 +5,8 @@ use crate::clipboard::{ClipCommand, ClipEventCore};
 use crate::config::{CompositorPref, GamepadPref, Mode};
 use crate::error::Result;
 use crate::input::InputEvent;
-use crate::quic::{HdrMeta, HidOutput};
-use std::sync::atomic::{AtomicBool, AtomicI64, AtomicU32, AtomicU64};
+use crate::quic::{HdrMeta, HidOutput, PadAudioFrame};
+use std::sync::atomic::{AtomicBool, AtomicI64, AtomicU32, AtomicU64, AtomicU8};
 use std::sync::mpsc::SyncSender;
 use std::sync::{Arc, Mutex};
 
@@ -43,6 +43,14 @@ pub(crate) struct WorkerArgs {
     /// closed, so the command API always observes connection teardown.
     pub(crate) rumble_feed: super::rumble::RumbleFeed,
     pub(crate) hidout_tx: SyncSender<HidOutput>,
+    /// Inbound pad-audio frames (`0xD1` — DualSense voice-coil haptics + speaker), drained by
+    /// [`NativeClient::next_pad_audio`].
+    pub(crate) pad_audio_tx: SyncSender<PadAudioFrame>,
+    /// Per-pad pad-audio render capabilities (bit0 haptics, bit1 speaker), written by
+    /// [`NativeClient::set_pad_audio_caps`] and OR'd into outgoing
+    /// [`GamepadArrival`](crate::input::InputKind::GamepadArrival) flags (bits 8/9) by the input
+    /// task — toward a `HOST_CAP_PAD_AUDIO` host only.
+    pub(crate) pad_audio_caps: Arc<[AtomicU8; crate::input::MAX_PADS]>,
     pub(crate) hdr_meta_tx: SyncSender<HdrMeta>,
     pub(crate) host_timing_tx: SyncSender<crate::quic::HostTiming>,
     pub(crate) cursor_shape_tx: SyncSender<crate::quic::CursorShape>,

crates/punktfunk-core/src/input.rs

@@ -64,7 +64,11 @@ pub enum InputKind {
     GamepadRemove = 13,
     /// Declares which controller KIND a pad presents so a session can MIX types (pad 0 a
     /// DualSense, pad 1 an Xbox pad). `code` = the [`GamepadPref`](crate::config::GamepadPref)
-    /// wire byte, `flags` = pad index. Sent when the client opens a pad slot — before that pad's
+    /// wire byte, `flags` = pad index in the low byte plus the pad's render capabilities in bits
+    /// 8/9 ([`ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`]/[`ARRIVAL_FLAG_PAD_AUDIO_SPEAKER`] — sent only
+    /// toward a [`HOST_CAP_PAD_AUDIO`](crate::quic::HOST_CAP_PAD_AUDIO) host, so an older host
+    /// keeps reading the whole word as the index; hosts decode via [`decode_gamepad_arrival`]).
+    /// Sent when the client opens a pad slot — before that pad's
     /// first input — and re-sent a few times against datagram loss (like [`GamepadRemove`]). The
     /// host resolves the kind to a buildable backend and routes that pad's virtual device to it; a
     /// pad the client never declares (an older client, or a fully-lost declaration) falls back to
@@ -97,6 +101,34 @@ pub fn decode_gamepad_remove(flags: u32) -> (u8, u8) {
     (flags as u8, (flags >> 24) as u8)
 }
 
+/// [`InputKind::GamepadArrival`] `flags` bit: this pad renders pad-audio HAPTICS — it is (or
+/// forwards to) a real DualSense whose voice-coil actuators can play the
+/// [`PAD_AUDIO_KIND_HAPTICS`](crate::quic::PAD_AUDIO_KIND_HAPTICS) stream. Rides above the pad
+/// index byte; sent only toward a [`HOST_CAP_PAD_AUDIO`](crate::quic::HOST_CAP_PAD_AUDIO) host
+/// (an older host reads the whole `flags` word as the index, so unexpected high bits would make
+/// it drop the declaration).
+pub const ARRIVAL_FLAG_PAD_AUDIO_HAPTICS: u32 = 1 << 8;
+/// [`InputKind::GamepadArrival`] `flags` bit: this pad renders pad-audio SPEAKER — the
+/// [`PAD_AUDIO_KIND_SPEAKER`](crate::quic::PAD_AUDIO_KIND_SPEAKER) stream. Same wire discipline
+/// as [`ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`].
+pub const ARRIVAL_FLAG_PAD_AUDIO_SPEAKER: u32 = 1 << 9;
+
+/// Pack a [`InputKind::GamepadArrival`] `flags` word: the pad index in the low byte plus
+/// `audio_caps` (bit0 = haptics, bit1 = speaker) as bits 8/9. `audio_caps = 0` reproduces the
+/// pre-pad-audio wire bytes exactly.
+pub fn encode_gamepad_arrival(pad: u8, audio_caps: u8) -> u32 {
+    (pad as u32) | (((audio_caps & 0x03) as u32) << 8)
+}
+
+/// Unpack a [`InputKind::GamepadArrival`] `flags` word into `(pad, audio_caps)`. The pad index
+/// is `flags & 0xFF` — hosts MUST mask rather than take the whole word, or a capability bit
+/// reads as a phantom index; `audio_caps` is bits 8/9 (bit0 = haptics, bit1 = speaker — the
+/// [`ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`]/[`ARRIVAL_FLAG_PAD_AUDIO_SPEAKER`] bits shifted down).
+/// An old-format word (index only) yields `audio_caps = 0`.
+pub fn decode_gamepad_arrival(flags: u32) -> (u8, u8) {
+    (flags as u8, ((flags >> 8) & 0x03) as u8)
+}
+
 /// The gamepad wire contract for [`InputKind::GamepadButton`]/[`InputKind::GamepadAxis`].
 ///
 /// Everything follows the GameStream/XInput conventions end to end: buttons reuse
@@ -348,6 +380,11 @@ pub enum GamepadEvent {
         kind: u8,
         /// LI_CCAP_* bits (0x02 = rumble).
         capabilities: u16,
+        /// Pad-audio render capabilities from a NATIVE-plane arrival's `flags` bits 8/9
+        /// (bit0 = haptics, bit1 = speaker — see [`decode_gamepad_arrival`]). NOT a GameStream
+        /// LI_CCAP bit (that vocabulary lives in `capabilities`); the GameStream plane cannot
+        /// express pad audio and always sets `0`, as does an old client.
+        audio_caps: u8,
     },
 }
 
@@ -443,6 +480,31 @@ mod tests {
         assert_eq!((pad, seq), (9, 123));
     }
 
+    #[test]
+    fn gamepad_arrival_flags_roundtrip() {
+        // The capability bits ride bits 8/9; the index stays the low byte.
+        for (pad, caps) in [(0u8, 0u8), (3, 0b01), (15, 0b10), (7, 0b11)] {
+            let flags = encode_gamepad_arrival(pad, caps);
+            assert_eq!(decode_gamepad_arrival(flags), (pad, caps));
+            assert_eq!(flags & 0xFF, pad as u32);
+        }
+        assert_eq!(
+            encode_gamepad_arrival(2, 0b11),
+            2 | ARRIVAL_FLAG_PAD_AUDIO_HAPTICS | ARRIVAL_FLAG_PAD_AUDIO_SPEAKER
+        );
+        // Old-format compat both ways: a caps-less word (an old client, or a new one toward an
+        // old host) is byte-identical to the plain index, and decodes with caps 0.
+        assert_eq!(encode_gamepad_arrival(5, 0), 5);
+        assert_eq!(decode_gamepad_arrival(5), (5, 0));
+        // Undefined high bits (a future extension) never leak into the index OR the caps.
+        assert_eq!(
+            decode_gamepad_arrival(0xFFFF_0000 | (0b01 << 8) | 9),
+            (9, 1)
+        );
+        // encode masks unknown caps bits, so a sloppy embedder can't corrupt the index space.
+        assert_eq!(encode_gamepad_arrival(1, 0xFF), 1 | (0b11 << 8));
+    }
+
     #[test]
     fn gamepad_snapshot_roundtrip() {
         let s = GamepadSnapshot {

crates/punktfunk-core/src/lib.rs

@@ -132,7 +132,13 @@ pub use stats::Stats;
 /// says what it says — so v15 is the floor that *guarantees* them: at or above it the surface is
 /// present, below it an embedder must probe for the symbol. Purely a version statement; no code
 /// changed with this bump, and no wire change, so [`WIRE_VERSION`] is unchanged.
-pub const ABI_VERSION: u32 = 15;
+/// v16: added the pad-audio client surface — `punktfunk_connection_next_pad_audio` (the 0xD1
+/// per-gamepad DualSense haptics/speaker plane) + `punktfunk_connection_set_pad_audio_caps` and
+/// the `PUNKTFUNK_CLIENT_CAP_PAD_AUDIO` / `PUNKTFUNK_HOST_CAP_PAD_AUDIO` mirrors. Additive and
+/// capability-gated end to end: the wire grows a new datagram tag (0xD1) an old client never
+/// receives (double-gated caps), a new 0xCD kind (0x06, dropped as unknown by old clients) and
+/// arrival flag bits 8/9 sent only toward a capable host, so [`WIRE_VERSION`] is unchanged.
+pub const ABI_VERSION: u32 = 16;
 
 /// The punktfunk/1 **wire** version — what `Hello`/`Welcome` carry and hosts equality-check.
 /// Deliberately its own constant: [`ABI_VERSION`] tracks the embeddable **C surface**

crates/punktfunk-core/src/quic/caps.rs

@@ -121,6 +121,15 @@ pub const CLIENT_CAP_PHASE_LOCK: u8 = 0x02;
 /// clean, the client keeps receiving the plain `0xC9` plane — so a client may always set this bit.
 /// `0x04` — `0x01`/`0x02` are cursor / phase-lock.
 pub const CLIENT_CAP_AUDIO_RED: u8 = 0x04;
+/// [`Hello::client_caps`] bit: the client understands the pad-audio plane
+/// ([`PAD_AUDIO_MAGIC`](super::datagram::PAD_AUDIO_MAGIC), `0xD1`) — per-gamepad DualSense
+/// voice-coil haptics + speaker Opus frames, plus the [`HidOutput::AudioCtl`]
+/// (super::datagram::HidOutput) routing/volume events. Active only when the host answers with
+/// [`HOST_CAP_PAD_AUDIO`] AND the pad's arrival declared a renderer for the kind
+/// ([`crate::input::ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`]/`_SPEAKER`) — the capable-and-agreed
+/// precedent, per pad; toward an older or incapable host nothing changes. `0x08` — `0x01` is [`CLIENT_CAP_CURSOR`],
+/// `0x02` is [`CLIENT_CAP_PHASE_LOCK`], `0x04` is [`CLIENT_CAP_AUDIO_RED`].
+pub const CLIENT_CAP_PAD_AUDIO: u8 = 0x08;
 
 /// [`Welcome::host_caps`] bit: the host CAN forward the cursor out-of-band (it captures cursor
 /// metadata separately from the frame — the Linux portal `SPA_META_Cursor` path; NOT gamescope,
@@ -154,6 +163,16 @@ pub const HOST_CAP_PEN: u8 = 0x10;
 /// unconditionally and treat this bit as "expect redundancy", not "only redundancy".
 /// `0x20` — `0x10` is [`HOST_CAP_PEN`], `0x08` is [`HOST_CAP_CURSOR`].
 pub const HOST_CAP_AUDIO_RED: u8 = 0x20;
+/// [`Welcome::host_caps`] bit: the host can capture pad audio — its virtual DualSense exposes
+/// the pad's audio endpoints (voice-coil haptics + speaker), so a game's per-pad audio can be
+/// captured and shipped on the [`PAD_AUDIO_MAGIC`](super::datagram::PAD_AUDIO_MAGIC) plane.
+/// Set only when the client asked via [`CLIENT_CAP_PAD_AUDIO`]; when both bits agree, a
+/// capable client marks its pads' render capabilities on their arrivals
+/// ([`crate::input::ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`]/`_SPEAKER`) and the host emits `0xD1`
+/// toward exactly those pads. `0x40` — `0x20` is [`HOST_CAP_AUDIO_RED`], `0x10` is
+/// [`HOST_CAP_PEN`], `0x08` is [`HOST_CAP_CURSOR`], `0x04` is [`HOST_CAP_TEXT_INPUT`],
+/// `0x01`/`0x02` are gamepad-state / clipboard.
+pub const HOST_CAP_PAD_AUDIO: u8 = 0x40;
 
 /// [`Hello::video_codecs`] bit: the client can decode H.264 / AVC. The GPU-less **software**
 /// encode path (openh264) emits H.264, so a client that wants to stream from a software host MUST
@@ -337,6 +356,28 @@ mod tests {
         );
     }
 
+    #[test]
+    fn pad_audio_cap_bits_are_distinct() {
+        // The new pad-audio bits pack into the existing caps bytes without colliding with any
+        // taken bit (a collision would silently negotiate an unrelated feature).
+        assert_eq!(
+            CLIENT_CAP_PAD_AUDIO & (CLIENT_CAP_CURSOR | CLIENT_CAP_PHASE_LOCK),
+            0
+        );
+        assert_eq!(
+            HOST_CAP_PAD_AUDIO
+                & (HOST_CAP_GAMEPAD_STATE
+                    | HOST_CAP_CLIPBOARD
+                    | HOST_CAP_TEXT_INPUT
+                    | HOST_CAP_CURSOR
+                    | HOST_CAP_PEN),
+            0
+        );
+        // Single-bit values (a multi-bit cap would OR neighbours in).
+        assert_eq!(CLIENT_CAP_PAD_AUDIO.count_ones(), 1);
+        assert_eq!(HOST_CAP_PAD_AUDIO.count_ones(), 1);
+    }
+
     #[test]
     fn resolve_codec_canonicalizes_a_multi_bit_preference() {
         // A non-conformant peer may stuff its capability MASK into `preferred` — the result

crates/punktfunk-core/src/quic/datagram.rs

@@ -1,12 +1,15 @@
-//! The QUIC-datagram side planes, demultiplexed by their first byte (0xC9–0xCF):
-//! audio, rumble, mic uplink, rich input, HID output, HDR metadata, host timing.
+//! The QUIC-datagram side planes, demultiplexed by their first byte (0xC9–0xD1):
+//! audio, rumble, mic uplink, rich input, HID output, HDR metadata, host timing,
+//! cursor state, pad audio.
 
 /// Datagram wire tags. Video rides UDP; everything low-rate rides QUIC datagrams,
 /// demultiplexed by the first byte: input = [`crate::input::INPUT_MAGIC`] (0xC8, client→host),
 /// audio = [`AUDIO_MAGIC`] (0xC9, host→client), rumble = [`RUMBLE_MAGIC`] (0xCA, host→client),
 /// mic = [`MIC_MAGIC`] (0xCB, client→host), rich-input = [`RICH_INPUT_MAGIC`] (0xCC, client→host),
 /// HID-output = [`HIDOUT_MAGIC`] (0xCD, host→client), HDR metadata = [`HDR_META_MAGIC`]
-/// (0xCE, host→client).
+/// (0xCE, host→client), host timing = [`HOST_TIMING_MAGIC`] (0xCF, host→client), cursor state =
+/// [`CURSOR_STATE_MAGIC`] (0xD0, host→client), pad audio = [`PAD_AUDIO_MAGIC`] (0xD1,
+/// host→client).
 pub const AUDIO_MAGIC: u8 = 0xC9;
 pub const RUMBLE_MAGIC: u8 = 0xCA;
 /// Microphone uplink: the client's mic, Opus-encoded, client → host (the inverse of
@@ -416,6 +419,7 @@ const HIDOUT_PLAYER_LEDS: u8 = 0x02;
 const HIDOUT_TRIGGER: u8 = 0x03;
 const HIDOUT_TRACKPAD_HAPTIC: u8 = 0x04;
 const HIDOUT_HID_RAW: u8 = 0x05;
+const HIDOUT_AUDIO_CTL: u8 = 0x06;
 
 /// [`HidOutput::HidRaw`] `kind`: an OUTPUT report — what the host's hidraw client wrote with
 /// `write()`/`SDL_hid_write` (Triton rumble `0x80`, haptic pulse `0x81`, …). The client replays
@@ -464,6 +468,16 @@ pub enum HidOutput {
     /// hardware safety timeout, and settings (lizard/IMU) are refreshed every ~3 s against the
     /// firmware watchdog — a lost datagram heals on the next refresh.
     HidRaw { pad: u8, kind: u8, data: Vec<u8> },
+    /// The audio-control region of a DS5 output report `0x02` a game wrote to the host's virtual
+    /// pad — the routing/volume side of pad audio (the audio SAMPLES ride the [`PAD_AUDIO_MAGIC`]
+    /// plane). `raw` is bytes 5..=10 of the report verbatim (headphone/speaker/mic volumes +
+    /// audio routing); `flags` condenses the report's audio valid-flags: bit0 = haptics-select
+    /// (`valid_flag0` bit1 — the title asked for audio haptics on the voice coils), bits1..4 =
+    /// `valid_flag0` bits 4..7 (the audio-valid flags gating `raw`). Wire form
+    /// `[0xCD][0x06][u16 pad LE][u8 flags][6 raw bytes]`. Forwarded change-only (deduped by
+    /// value host-side, like `Led`/`Trigger`) — a merely-rumbling pad re-sends unchanged audio
+    /// state on every output report.
+    AudioCtl { pad: u16, flags: u8, raw: [u8; 6] },
 }
 
 impl HidOutput {
@@ -496,6 +510,12 @@ impl HidOutput {
                 out.extend_from_slice(&[HIDOUT_HID_RAW, *pad, *kind]);
                 out.extend_from_slice(&data[..data.len().min(HID_REPORT_MAX)]);
             }
+            HidOutput::AudioCtl { pad, flags, raw } => {
+                out.push(HIDOUT_AUDIO_CTL);
+                out.extend_from_slice(&pad.to_le_bytes());
+                out.push(*flags);
+                out.extend_from_slice(raw);
+            }
         }
         out
     }
@@ -540,6 +560,22 @@ impl HidOutput {
                 // Bounded: at most HID_REPORT_MAX bytes are kept from the (attacker-sized) tail.
                 data: b[4..b.len().min(4 + HID_REPORT_MAX)].to_vec(),
             }),
+            // B27: the pad is the only u16 index on this plane, and every consumer narrows it
+            // with `as u8` on the stated assumption that pads are 0..MAX_PADS. Nothing enforced
+            // that, so wire pad 256 silently ALIASED onto slot 0 — a malformed or hostile
+            // datagram steering a real controller's speaker volumes. Rejected here, at the one
+            // place the u16 exists, so the narrowings downstream are lossless by construction
+            // (the same fix R10 applied to the rumble plane).
+            HIDOUT_AUDIO_CTL
+                if b.len() >= 11
+                    && u16::from_le_bytes([b[2], b[3]]) < crate::input::MAX_PADS as u16 =>
+            {
+                Some(HidOutput::AudioCtl {
+                    pad: u16::from_le_bytes([b[2], b[3]]),
+                    flags: b[4],
+                    raw: b[5..11].try_into().unwrap(),
+                })
+            }
             _ => None,
         }
     }
@@ -798,6 +834,72 @@ pub fn decode_cursor_state_datagram(b: &[u8]) -> Option<CursorState> {
     })
 }
 
+/// Pad-audio datagram tag, host → client: per-gamepad audio a game routed
+/// to the host's virtual DualSense — voice-coil haptics and the built-in speaker — for the client
+/// to render on the matching real controller. Next tag after [`CURSOR_STATE_MAGIC`]. The
+/// per-pad AUDIO plane (Opus frames, the [`AUDIO_MAGIC`]/[`MIC_MAGIC`] shape plus pad + kind);
+/// the routing/volume CONTROL side rides [`HidOutput::AudioCtl`]. Emitted only when the session
+/// negotiated it ([`CLIENT_CAP_PAD_AUDIO`](super::caps::CLIENT_CAP_PAD_AUDIO) ∧
+/// [`HOST_CAP_PAD_AUDIO`](super::caps::HOST_CAP_PAD_AUDIO)) and the pad's arrival declared a
+/// renderer for the kind ([`crate::input::ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`]/`_SPEAKER`).
+/// Best-effort like every audio datagram: a lost frame is a concealed gap, never state.
+pub const PAD_AUDIO_MAGIC: u8 = 0xD1;
+
+/// [`PadAudioFrame::kind`]: the BACK channel pair — the DualSense voice-coil actuators (audio
+/// haptics). 5 ms Opus frames, matching the [`AUDIO_MAGIC`] cadence: haptics are felt latency.
+pub const PAD_AUDIO_KIND_HAPTICS: u8 = 0;
+/// [`PadAudioFrame::kind`]: the FRONT channel pair — the controller's built-in speaker. 10 ms
+/// Opus frames (speaker content tolerates the extra buffering for the better coding efficiency).
+pub const PAD_AUDIO_KIND_SPEAKER: u8 = 1;
+
+/// Wire length of a pad-audio datagram header: tag + pad + kind + u32 seq + u64 pts = 15 bytes.
+const PAD_AUDIO_HEADER_LEN: usize = 1 + 1 + 1 + 4 + 8;
+
+/// One decoded pad-audio frame (owned — the client's plane queue stores it). `seq`/`pts_ns` are
+/// per-(pad, kind) counters from the host's capture clock, for gap concealment and lip-sync
+/// against the main audio plane.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct PadAudioFrame {
+    /// Gamepad index (the wire pad space, same as rumble/HID-output).
+    pub pad: u8,
+    /// [`PAD_AUDIO_KIND_HAPTICS`] or [`PAD_AUDIO_KIND_SPEAKER`].
+    pub kind: u8,
+    pub seq: u32,
+    pub pts_ns: u64,
+    /// The raw Opus payload — feed it to an Opus decoder as one frame. Empty = DTX silence.
+    pub opus: Vec<u8>,
+}
+
+/// Pad-audio datagram, host → client:
+/// `[0xD1][u8 pad][u8 kind][u32 seq LE][u64 pts_ns LE][opus payload]` — the
+/// [`encode_audio_datagram`]/[`encode_mic_datagram`] layout with a pad + kind prefix, one Opus
+/// frame per datagram (5/10 ms — well under any MTU); QUIC already encrypts.
+pub fn encode_pad_audio_datagram(pad: u8, kind: u8, seq: u32, pts_ns: u64, opus: &[u8]) -> Vec<u8> {
+    let mut b = Vec::with_capacity(PAD_AUDIO_HEADER_LEN + opus.len());
+    b.push(PAD_AUDIO_MAGIC);
+    b.push(pad);
+    b.push(kind);
+    b.extend_from_slice(&seq.to_le_bytes());
+    b.extend_from_slice(&pts_ns.to_le_bytes());
+    b.extend_from_slice(opus);
+    b
+}
+
+/// Parse a pad-audio datagram → [`PadAudioFrame`]. `None` on bad tag/length (the fixed header
+/// length bounds every read before it happens).
+pub fn decode_pad_audio_datagram(buf: &[u8]) -> Option<PadAudioFrame> {
+    if buf.len() < PAD_AUDIO_HEADER_LEN || buf[0] != PAD_AUDIO_MAGIC {
+        return None;
+    }
+    Some(PadAudioFrame {
+        pad: buf[1],
+        kind: buf[2],
+        seq: u32::from_le_bytes(buf[3..7].try_into().unwrap()),
+        pts_ns: u64::from_le_bytes(buf[7..15].try_into().unwrap()),
+        opus: buf[15..].to_vec(),
+    })
+}
+
 #[cfg(test)]
 mod tests {
     use crate::quic::*;
@@ -1281,6 +1383,12 @@ mod tests {
                     f
                 },
             },
+            // The DS5 audio-control region (haptics-select + speaker volume asserted).
+            HidOutput::AudioCtl {
+                pad: 1,
+                flags: 0b0_0101,
+                raw: [0x50, 0x60, 0x70, 0x05, 0x00, 0x00],
+            },
         ];
         for ev in &cases {
             let d = ev.encode();
@@ -1299,6 +1407,92 @@ mod tests {
         )
         .is_none());
     }
+
+    #[test]
+    fn audio_ctl_wire_layout_and_truncation() {
+        // The exact 11-byte layout: [0xCD][0x06][u16 pad LE][u8 flags][6 raw bytes].
+        // The pad is deliberately a REPRESENTABLE one: this used to assert that 0x0201 (513)
+        // round-tripped, which pinned B27's aliasing in place as if it were the contract.
+        let a = HidOutput::AudioCtl {
+            pad: 0x000B,
+            flags: 0x17,
+            raw: [1, 2, 3, 4, 5, 6],
+        };
+        let d = a.encode();
+        assert_eq!(d, [0xCD, 0x06, 0x0B, 0x00, 0x17, 1, 2, 3, 4, 5, 6]);
+        assert_eq!(HidOutput::decode(&d), Some(a));
+        // Truncated buffers are rejected outright (fixed length — never a partial read).
+        for n in 2..d.len() {
+            assert_eq!(HidOutput::decode(&d[..n]), None);
+        }
+    }
+
+    #[test]
+    fn pad_audio_datagram_roundtrip_and_truncation() {
+        let opus = [0x5Au8; 61];
+        let d = encode_pad_audio_datagram(3, PAD_AUDIO_KIND_HAPTICS, 42, 9_999, &opus);
+        assert_eq!(d[0], PAD_AUDIO_MAGIC);
+        assert_eq!(d.len(), 15 + opus.len());
+        let f = decode_pad_audio_datagram(&d).unwrap();
+        assert_eq!((f.pad, f.kind, f.seq, f.pts_ns), (3, 0, 42, 9_999));
+        assert_eq!(f.opus, opus);
+        // Truncated headers are rejected outright (never partially read).
+        for n in 0..15 {
+            assert_eq!(decode_pad_audio_datagram(&d[..n]), None);
+        }
+        // Tag separation: a pad-audio datagram is not a session-audio/mic datagram and vice-versa.
+        assert!(decode_audio_datagram(&d).is_none());
+        assert!(decode_mic_datagram(&d).is_none());
+        assert!(decode_pad_audio_datagram(&encode_audio_datagram(1, 2, &opus)).is_none());
+        // Empty payload (DTX) is legal — header-only datagram.
+        let hdr = encode_pad_audio_datagram(0, PAD_AUDIO_KIND_SPEAKER, 0, 0, &[]);
+        assert_eq!(hdr.len(), 15);
+        assert!(decode_pad_audio_datagram(&hdr).unwrap().opus.is_empty());
+    }
+
+    /// B27: the pad is the only u16 index on the 0xCD plane and every consumer narrows it with
+    /// `as u8`. An out-of-range one used to alias onto a real slot instead of being refused —
+    /// wire pad 256 steering pad 0's speaker volumes.
+    #[test]
+    fn audio_ctl_rejects_a_pad_outside_the_index_space() {
+        let ok = HidOutput::AudioCtl {
+            pad: (crate::input::MAX_PADS - 1) as u16,
+            flags: 0x12,
+            raw: [1, 2, 3, 4, 5, 6],
+        };
+        assert_eq!(
+            HidOutput::decode(&ok.encode()),
+            Some(ok),
+            "the last valid pad must still decode"
+        );
+
+        // Anything at or above MAX_PADS is refused outright, not truncated.
+        for pad in [crate::input::MAX_PADS as u16, 256, u16::MAX] {
+            let d = HidOutput::AudioCtl {
+                pad,
+                flags: 0x12,
+                raw: [1, 2, 3, 4, 5, 6],
+            }
+            .encode();
+            assert_eq!(HidOutput::decode(&d), None, "pad {pad} must not decode");
+        }
+
+        // The specific alias the bug produced: 256 as u8 == 0.
+        let d = HidOutput::AudioCtl {
+            pad: 256,
+            flags: 0,
+            raw: [0; 6],
+        }
+        .encode();
+        assert!(
+            !matches!(
+                HidOutput::decode(&d),
+                Some(HidOutput::AudioCtl { pad: 0, .. })
+            ),
+            "wire pad 256 must never surface as pad 0"
+        );
+    }
+
     #[test]
     fn cursor_state_roundtrip() {
         for (flags, x, y) in [

crates/punktfunk-core/src/quic/mod.rs

@@ -25,7 +25,7 @@
 //! Split by concern (networking-audit deferred plan §3 — a pure move): `handshake` the
 //! positional Hello/Welcome/Start codecs, `caps` the capability/codec-negotiation
 //! vocabulary, `control` the typed control + clipboard messages, `pairing` the pairing
-//! message codecs with [`pake`] the SPAKE2 itself, `datagram` the 0xC9–0xCF plane codecs,
+//! message codecs with [`pake`] the SPAKE2 itself, `datagram` the 0xC9–0xD1 plane codecs,
 //! `pen` the stylus batch (0xCC kind 0x05) + host stroke tracker,
 //! [`io`] framed stream IO, `clock` skew estimation + mid-stream re-sync, [`endpoint`] the
 //! quinn constructors, [`clipstream`] the per-transfer clipboard fetch streams. Every item

crates/punktfunk-host/Cargo.toml

@@ -259,6 +259,17 @@ windows = { version = "0.62", features = [
     # CoCreateInstance(PolicyConfigClient) — set the default audio playback/recording endpoints via the
     # undocumented IPolicyConfig (audio/windows/audio_control.rs) so mic + desktop audio auto-wire.
     "Win32_System_Com",
+    # Pad-audio endpoint provisioning (audio/windows/pad_endpoint.rs): IMMDevice + IPropertyStore
+    # to stamp the DualSense identity onto the minted endpoints (PROPVARIANT lives in
+    # StructuredStorage and is gated on the Variant feature), DEVPKEY_Device_DriverInfPath to
+    # resolve the installed Steam Streaming Speakers INF, and raw Reg* calls behind the MMDevices
+    # ACL repair + the devnode's pad-index marker value.
+    "Win32_Media_Audio",
+    "Win32_UI_Shell_PropertiesSystem",
+    "Win32_System_Com_StructuredStorage",
+    "Win32_System_Variant",
+    "Win32_Devices_Properties",
+    "Win32_System_Registry",
     # SetUnhandledExceptionFilter + EXCEPTION_POINTERS — the last-resort native-crash logger
     # (src/windows/crash.rs); Kernel gates the CONTEXT type EXCEPTION_POINTERS embeds.
     "Win32_System_Diagnostics_Debug",

crates/punktfunk-host/src/audio.rs

@@ -183,6 +183,12 @@ pub fn open_virtual_mic(_channels: u32) -> Result<Box<dyn VirtualMic>> {
 mod audio_control;
 #[cfg(target_os = "linux")]
 mod linux;
+// DualSense pad-audio endpoint provisioning + loopback capture (design: pad haptics/audio).
+// pub(crate): the session layer queries endpoints by pad index and the CLI exposes the
+// `pad-endpoint` devtest.
+#[cfg(target_os = "windows")]
+#[path = "audio/windows/pad_endpoint.rs"]
+pub(crate) mod pad_endpoint;
 #[cfg(target_os = "windows")]
 #[path = "audio/windows/wasapi_cap.rs"]
 mod wasapi_cap;

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

@@ -143,6 +143,17 @@ pub(crate) fn wire_now(set_playback: bool) -> Wiring {
     wire_now_full(set_playback).wiring
 }
 
+/// Endpoint ids among `renders` that are the host's own pad-audio endpoints — the exclusion
+/// data [`plan`] runs on. Detection lives in [`super::pad_endpoint`] (stamped PFDS container /
+/// devnode marker, registry-only reads); this is just the per-pass collection.
+fn pad_render_ids(renders: &[Endpoint]) -> Vec<String> {
+    renders
+        .iter()
+        .filter(|(_, id)| super::pad_endpoint::is_pad_render_endpoint(id))
+        .map(|(_, id)| id.clone())
+        .collect()
+}
+
 /// Enumerate endpoints, compute the assignment, apply the default-device changes (unless
 /// `PUNKTFUNK_KEEP_DEFAULT`), and return the plan for the caller to act on (mic target / loopback
 /// echo guard). `set_playback` — true only from the desktop-audio capture open — additionally
@@ -159,6 +170,10 @@ pub(crate) fn wire_now_full(set_playback: bool) -> WiredPlan {
     let want = std::env::var("PUNKTFUNK_MIC_DEVICE")
         .ok()
         .map(|s| s.to_lowercase());
+    // The host's own pad-audio ("DualSense speaker") endpoints, by id — the pure plan filters
+    // them out of every role. Identity is platform data (stamped container / devnode marker),
+    // so it is collected HERE and passed in, like the candidate lists themselves.
+    let pad_ids = pad_render_ids(&renders);
     // Mix formats are read only when we are actually going to park the playback default (i.e. a
     // desktop-audio capture is opening). The mic pump wires on every open while the host is idle
     // and does not care which loopback endpoint wins, so it must not pay an IAudioClient
@@ -179,6 +194,7 @@ pub(crate) fn wire_now_full(set_playback: bool) -> WiredPlan {
         // only count a *narrowing* verdict can be made against without guessing: an endpoint that
         // cannot carry stereo cannot carry 5.1 either.
         2,
+        &pad_ids,
     );
     let done = |wiring: Wiring| WiredPlan {
         wiring,
@@ -245,7 +261,7 @@ pub(crate) fn wire_now_full(set_playback: bool) -> WiredPlan {
     if let Some((mic_name, mic_id)) = &wiring.mic_render {
         if default_render_id().as_deref() == Some(mic_id.as_str()) {
             // Audible preference = the host_audio plan's loopback pick (real hardware first).
-            match plan(&renders, &captures, want.as_deref(), true).loopback_render {
+            match plan(&renders, &captures, want.as_deref(), true, &pad_ids).loopback_render {
                 Some((name, id)) => match set_default_endpoint(&id) {
                     Ok(()) => tracing::info!(mic = %mic_name, device = %name,
                         "default playback was the virtual-mic target — moved it so desktop \
@@ -302,8 +318,10 @@ fn park_marker_path() -> std::path::PathBuf {
     pf_paths::config_dir().join("audio-default.prev")
 }
 
-/// The current default RENDER endpoint id, if any.
-fn default_render_id() -> Option<String> {
+/// The current default RENDER endpoint id, if any. pub(crate): the pad-endpoint provisioning
+/// uses it for its default-device guard (a freshly minted pad endpoint must never stay the
+/// default playback device).
+pub(crate) fn default_render_id() -> Option<String> {
     wasapi::DeviceEnumerator::new()
         .ok()?
         .get_default_device(&Direction::Render)
@@ -430,11 +448,13 @@ pub(crate) fn restore_default_playback() {
 }
 
 /// Open a device by endpoint id, with a name for error context.
+///
+/// Resolves through [`super::pad_endpoint::open_wasapi_device`], NOT the `wasapi` crate's
+/// `DeviceEnumerator::get_device` — that one hands `GetDevice` a freed string (see the helper's
+/// docs), so it fails at random on ids that are perfectly valid.
 pub(crate) fn open_endpoint(ep: &Endpoint) -> Result<wasapi::Device> {
-    wasapi::DeviceEnumerator::new()
-        .map_err(|e| anyhow!("DeviceEnumerator: {e}"))?
-        .get_device(&ep.1)
-        .map_err(|e| anyhow!("open endpoint {:?}: {e}", ep.0))
+    super::pad_endpoint::open_wasapi_device(&ep.1)
+        .map_err(|e| anyhow!("open endpoint {:?}: {e:#}", ep.0))
 }
 
 // --- IPolicyConfig (undocumented): set a default audio endpoint by id, for all three roles. ---
@@ -481,8 +501,9 @@ const _: () = {
 
 /// Set `device_id` as the default audio endpoint for eConsole/eMultimedia/eCommunications via the
 /// undocumented `IPolicyConfig::SetDefaultEndpoint` (the call `mmsys.cpl` makes). Errs if any role
-/// fails.
-fn set_default_endpoint(device_id: &str) -> Result<()> {
+/// fails. pub(crate): the pad-endpoint default-device guard restores the operator's default
+/// through the same machinery.
+pub(crate) fn set_default_endpoint(device_id: &str) -> Result<()> {
     use windows::core::{IUnknown, Interface, GUID, PCWSTR};
     use windows::Win32::System::Com::{CoCreateInstance, CLSCTX_ALL};
 

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

@@ -511,7 +511,7 @@ fn capture_once(
     if assert_plan {
         if let Some(d) = seen_default.as_deref() {
             if d != dev_id {
-                match judge_default(&en, wiring, d) {
+                match judge_default(wiring, d) {
                     DefaultKind::Capturable(name) => {
                         tracing::info!(default = %name, planned = %dev_name,
                             "could not park the default playback on the planned endpoint — \
@@ -639,7 +639,7 @@ fn capture_once(
                             );
                             return Ok(Next::Reopen(TargetMode::Follow));
                         }
-                        match judge_default(&en, wiring, &nid) {
+                        match judge_default(wiring, &nid) {
                             DefaultKind::Capturable(name) => {
                                 audio_client.stop_stream().ok();
                                 tracing::info!(device = %name,
@@ -726,8 +726,11 @@ enum DefaultKind {
     Unknown,
 }
 
-fn judge_default(en: &DeviceEnumerator, wiring: &wiring_plan::Wiring, id: &str) -> DefaultKind {
-    let Ok(dev) = en.get_device(id) else {
+/// Resolves through [`super::pad_endpoint::open_wasapi_device`], NOT the `wasapi` crate's
+/// `DeviceEnumerator::get_device` — that one hands `GetDevice` a freed string (see the helper's
+/// docs), and a spurious miss here silently downgrades a capturable default to `Unknown`.
+fn judge_default(wiring: &wiring_plan::Wiring, id: &str) -> DefaultKind {
+    let Ok(dev) = super::pad_endpoint::open_wasapi_device(id) else {
         return DefaultKind::Unknown;
     };
     let name = dev.get_friendlyname().unwrap_or_default();
@@ -736,7 +739,15 @@ fn judge_default(en: &DeviceEnumerator, wiring: &wiring_plan::Wiring, id: &str)
         .mic_render
         .as_ref()
         .is_some_and(|(_, mic_id)| mic_id == id);
-    if is_mic || wiring_plan::excluded_from_loopback(&ln) {
+    // B10: a pad's audio endpoint is not ordinary hardware, and the name rules cannot see that —
+    // it is deliberately stamped with the controller's own name ("DualSense Wireless Controller")
+    // so games treat it as the pad's speaker, which means `excluded_from_loopback` passes it
+    // straight through as `Capturable`. The pure plan filtered these out, but the plan is not the
+    // only reader: this classifier drives the watchdog, Follow mode and the parked default, so a
+    // pad endpoint that happened to be the system default could be adopted as the desktop capture
+    // source — sending the whole desktop mix to a controller's voice coils. Identity, not name.
+    let is_pad = super::pad_endpoint::is_pad_render_endpoint(id);
+    if is_mic || is_pad || wiring_plan::excluded_from_loopback(&ln) {
         DefaultKind::Dud(name)
     } else {
         DefaultKind::Capturable(name)

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

@@ -253,25 +253,16 @@ pub(crate) fn install_steam_audio_pair() -> bool {
     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.
-///
-/// Safe: `inf_name` is a `&str` and every FFI argument is built locally from it, so there is no
-/// precondition a caller could break — the `unsafe` is the `LoadLibraryExW`/`transmute`/call chain
-/// inside, which is this function's own business.
-fn try_install_steam_audio(inf_name: &str) -> bool {
-    use windows::core::{s, w, PCWSTR};
-    use windows::Win32::Foundation::HWND;
+/// Full path of a Steam Remote Play driver INF under Steam's per-arch driver directory
+/// (`%CommonProgramFiles(x86)%\Steam\drivers\Windows10\{arch}\<inf_name>`), as a NUL-terminated
+/// UTF-16 buffer. Shared by [`try_install_steam_audio`] and the pad-endpoint provisioning
+/// ([`super::pad_endpoint`]), which feeds the same INF to `UpdateDriverForPlugAndPlayDevicesW`
+/// when no installed Steam Streaming Speakers devnode exposes its `oemNN.inf`. `None` when the
+/// environment expansion fails (existence is the caller's check).
+pub(crate) fn steam_driver_inf_path(inf_name: &str) -> Option<Vec<u16>> {
+    use windows::core::PCWSTR;
     use windows::Win32::System::Environment::ExpandEnvironmentStringsW;
-    use windows::Win32::System::LibraryLoader::{
-        GetProcAddress, LoadLibraryExW, LOAD_LIBRARY_SEARCH_SYSTEM32,
-    };
 
-    if std::env::var_os("PUNKTFUNK_NO_MIC_INSTALL").is_some() {
-        return false;
-    }
     // Steam ships per-arch driver INFs under `Steam\drivers\Windows10\{arch}\`.
     #[cfg(target_arch = "x86_64")]
     let subdir = "x64";
@@ -290,8 +281,33 @@ fn try_install_steam_audio(inf_name: &str) -> bool {
     let n =
         unsafe { ExpandEnvironmentStringsW(PCWSTR(template.as_ptr()), Some(path.as_mut_slice())) };
     if n == 0 || n as usize > path.len() {
+        return None;
+    }
+    path.truncate(n as usize); // keeps the NUL
+    Some(path)
+}
+
+/// 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.
+///
+/// Safe: `inf_name` is a `&str` and every FFI argument is built locally from it, so there is no
+/// precondition a caller could break — the `unsafe` is the `LoadLibraryExW`/`transmute`/call chain
+/// inside, which is this function's own business.
+fn try_install_steam_audio(inf_name: &str) -> bool {
+    use windows::core::{s, w, PCWSTR};
+    use windows::Win32::Foundation::HWND;
+    use windows::Win32::System::LibraryLoader::{
+        GetProcAddress, LoadLibraryExW, LOAD_LIBRARY_SEARCH_SYSTEM32,
+    };
+
+    if std::env::var_os("PUNKTFUNK_NO_MIC_INSTALL").is_some() {
         return false;
     }
+    let Some(path) = steam_driver_inf_path(inf_name) else {
+        return false;
+    };
 
     // SAFETY: a static NUL-terminated literal, loaded from System32 only (the flag), so this cannot
     // pick up a planted `newdev.dll` from the working directory. The handle is checked before use.

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

@@ -186,6 +186,17 @@ fn virtualish(lname: &str) -> bool {
         || lname.contains("voicemeeter")
 }
 
+/// Is this render endpoint id one of the virtual pad's audio endpoints?
+///
+/// Pulled out of [`plan`] because the plan is NOT the only place that must not treat these as
+/// ordinary hardware — see [`excluded_from_loopback`]'s callers. A pad endpoint is deliberately
+/// stamped with the controller's own name ("DualSense Wireless Controller") so games read it as
+/// the pad's speaker, which means no name-based rule can recognise one; the only reliable test is
+/// identity against the ids the pad-endpoint provisioner created.
+pub(crate) fn is_pad_render(id: &str, pad_renders: &[String]) -> bool {
+    pad_renders.iter().any(|p| p == id)
+}
+
 /// Compute the assignment. `mic_want` is the operator override (`PUNKTFUNK_MIC_DEVICE`,
 /// lowercased): when set it beats the built-in candidate order for the mic target. `host_audio`
 /// flips the loopback preference to real hardware (audio audible on the host too); the default
@@ -195,8 +206,17 @@ pub(crate) fn plan(
     captures: &[Endpoint],
     mic_want: Option<&str>,
     host_audio: bool,
+    pad_renders: &[String],
 ) -> Wiring {
-    plan_with_formats(renders, captures, mic_want, host_audio, &no_formats, 2)
+    plan_with_formats(
+        renders,
+        captures,
+        mic_want,
+        host_audio,
+        &no_formats,
+        2,
+        pad_renders,
+    )
 }
 
 /// [`plan`] with knowledge of each render endpoint's engine mix format, and the channel count the
@@ -221,7 +241,20 @@ pub(crate) fn plan_with_formats(
     host_audio: bool,
     format_of: FormatProbe,
     want_channels: u8,
+    pad_renders: &[String],
 ) -> Wiring {
+    // 0. Pad-audio endpoints are invisible to the plan: never the mic target (client voice
+    //    would play out of a pad "speaker"), never a loopback source (a game's controller
+    //    audio cues would stream as desktop audio), and — since this shadows `renders` for
+    //    every tier below — never the flagged last resort either. Their names carry no virtual
+    //    marker (they are stamped "DualSense Wireless Controller" on purpose, so games read
+    //    them as the pad's speaker), so the name rules alone would take one for real hardware.
+    let renders: Vec<Endpoint> = renders
+        .iter()
+        .filter(|(_, id)| !is_pad_render(id, pad_renders))
+        .cloned()
+        .collect();
+    let renders = renders.as_slice();
     let find_render = |needle: &str| {
         renders
             .iter()
@@ -422,7 +455,7 @@ mod tests {
             ep("Microphone (Webcam)"),
             ep("CABLE Output (VB-Audio Virtual Cable)"),
         ];
-        let w = plan(&renders, &captures, None, false);
+        let w = plan(&renders, &captures, None, false, &[]);
         assert_eq!(
             w.mic_render.unwrap().0,
             "CABLE Input (VB-Audio Virtual Cable)"
@@ -451,7 +484,7 @@ mod tests {
             ep("CABLE Output (VB-Audio Virtual Cable)"),
             ep("Microphone (Steam Streaming Microphone)"),
         ];
-        let w = plan(&renders, &captures, None, false);
+        let w = plan(&renders, &captures, None, false, &[]);
         assert_eq!(
             w.mic_render.unwrap().0,
             "CABLE Input (VB-Audio Virtual Cable)"
@@ -471,7 +504,7 @@ mod tests {
             ep("CABLE Input (VB-Audio Virtual Cable)"),
             ep("Speakers (Steam Streaming Microphone)"),
         ];
-        let w = plan(&renders, &[], None, true);
+        let w = plan(&renders, &[], None, true, &[]);
         assert_eq!(
             w.loopback_render.unwrap().0,
             "Speakers (Apple Audio Device)"
@@ -488,7 +521,7 @@ mod tests {
             ep("CABLE In 16ch (VB-Audio Virtual Cable)"),
         ];
         for host_audio in [false, true] {
-            let w = plan(&renders, &[], None, host_audio);
+            let w = plan(&renders, &[], None, host_audio, &[]);
             assert!(w.loopback_render.is_none(), "host_audio={host_audio}");
         }
     }
@@ -500,7 +533,7 @@ mod tests {
     fn headless_cable_only_mic_wins() {
         let renders = [ep("CABLE Input (VB-Audio Virtual Cable)")];
         let captures = [ep("CABLE Output (VB-Audio Virtual Cable)")];
-        let w = plan(&renders, &captures, None, false);
+        let w = plan(&renders, &captures, None, false, &[]);
         assert!(w.mic_render.is_some(), "mic must claim the only cable");
         assert!(w.loopback_render.is_none(), "no echo-safe loopback exists");
     }
@@ -518,7 +551,7 @@ mod tests {
             ep("CABLE Output (VB-Audio Virtual Cable)"),
             ep("Microphone (Steam Streaming Microphone)"),
         ];
-        let w = plan(&renders, &captures, None, false);
+        let w = plan(&renders, &captures, None, false, &[]);
         assert_eq!(
             w.mic_render.unwrap().0,
             "CABLE Input (VB-Audio Virtual Cable)"
@@ -546,7 +579,7 @@ mod tests {
             ep("Speakers (Realtek HD Audio)"),
         ];
         let captures = [ep("Microphone (Steam Streaming Microphone)")];
-        let w = plan(&renders, &captures, None, false);
+        let w = plan(&renders, &captures, None, false, &[]);
         assert_eq!(
             w.mic_render.unwrap().0,
             "Speakers (Steam Streaming Microphone)"
@@ -560,7 +593,7 @@ mod tests {
     fn steam_mic_only_no_echo() {
         let renders = [ep("Speakers (Steam Streaming Microphone)")];
         let captures = [ep("Microphone (Steam Streaming Microphone)")];
-        let w = plan(&renders, &captures, None, false);
+        let w = plan(&renders, &captures, None, false, &[]);
         assert!(w.mic_render.is_some());
         assert!(w.loopback_render.is_none());
     }
@@ -576,7 +609,7 @@ mod tests {
             ep("Speakers (Steam Streaming Speakers)"),
         ];
         for host_audio in [false, true] {
-            let w = plan(&renders, &[], None, host_audio);
+            let w = plan(&renders, &[], None, host_audio, &[]);
             assert_eq!(
                 w.loopback_render.as_ref().unwrap().0,
                 "Speakers (Steam Streaming Speakers)",
@@ -597,7 +630,7 @@ mod tests {
             ep("Altavoces (Steam Streaming Microphone)"),
         ];
         let captures = [ep("Microphone (Steam Streaming Microphone)")];
-        let w = plan(&renders, &captures, None, false);
+        let w = plan(&renders, &captures, None, false, &[]);
         assert_eq!(
             w.mic_render.unwrap().0,
             "Altavoces (Steam Streaming Microphone)"
@@ -620,7 +653,7 @@ mod tests {
         ];
         let captures = [ep("Microphone (Steam Streaming Microphone)")];
         for host_audio in [false, true] {
-            let w = plan(&renders, &captures, None, host_audio);
+            let w = plan(&renders, &captures, None, host_audio, &[]);
             assert_eq!(
                 w.loopback_render.as_ref().unwrap().0,
                 "Speakers (Realtek HD Audio)",
@@ -642,7 +675,7 @@ mod tests {
         ];
         let captures = [ep("CABLE Output (VB-Audio Virtual Cable)")];
         for host_audio in [false, true] {
-            let w = plan(&renders, &captures, None, host_audio);
+            let w = plan(&renders, &captures, None, host_audio, &[]);
             assert!(w.loopback_render.is_none(), "host_audio={host_audio}");
             assert!(!w.loopback_last_resort, "host_audio={host_audio}");
             assert!(w.loopback_unsatisfiable(), "host_audio={host_audio}");
@@ -691,7 +724,7 @@ mod tests {
             ("steam streaming microphone", fmt(24_000, 1)),
             ("odyssey", fmt(48_000, 2)),
         ]);
-        let w = plan_with_formats(&renders, &captures, None, false, &p, 2);
+        let w = plan_with_formats(&renders, &captures, None, false, &p, 2, &[]);
         assert_eq!(
             w.loopback_render.as_ref().unwrap().0,
             "1 - Odyssey G60SD (AMD High Definition Audio Device)",
@@ -721,7 +754,7 @@ mod tests {
             ("steam streaming microphone", fmt(48_000, 2)),
             ("realtek", fmt(48_000, 2)),
         ]);
-        let w = plan_with_formats(&renders, &[], None, false, &p, 2);
+        let w = plan_with_formats(&renders, &[], None, false, &p, 2, &[]);
         assert_eq!(
             w.loopback_render.unwrap().0,
             "Speakers (Steam Streaming Microphone)"
@@ -737,7 +770,7 @@ mod tests {
             ep("Speakers (Steam Streaming Microphone)"),
         ];
         let p = probe(vec![("steam streaming microphone", fmt(16_000, 1))]);
-        let w = plan_with_formats(&renders, &[], None, false, &p, 2);
+        let w = plan_with_formats(&renders, &[], None, false, &p, 2, &[]);
         assert_eq!(
             w.loopback_render.as_ref().unwrap().0,
             "Speakers (Steam Streaming Microphone)"
@@ -753,7 +786,7 @@ mod tests {
     fn narrowing_is_reported_for_real_hardware_too() {
         let renders = [ep("Headset (Hands-Free AG Audio)")];
         let p = probe(vec![("headset", fmt(16_000, 1))]);
-        let w = plan_with_formats(&renders, &[], None, false, &p, 2);
+        let w = plan_with_formats(&renders, &[], None, false, &p, 2, &[]);
         assert_eq!(
             w.loopback_render.as_ref().unwrap().0,
             "Headset (Hands-Free AG Audio)"
@@ -773,8 +806,8 @@ mod tests {
         ];
         let captures = [ep("CABLE Output (VB-Audio Virtual Cable)")];
         for host_audio in [false, true] {
-            let a = plan(&renders, &captures, None, host_audio);
-            let b = plan_with_formats(&renders, &captures, None, host_audio, &no_formats, 2);
+            let a = plan(&renders, &captures, None, host_audio, &[]);
+            let b = plan_with_formats(&renders, &captures, None, host_audio, &no_formats, 2, &[]);
             assert_eq!(a, b, "host_audio={host_audio}");
             assert!(a.loopback_narrowing.is_none());
         }
@@ -792,7 +825,7 @@ mod tests {
             ("steam streaming microphone", fmt(24_000, 1)),
             ("realtek", fmt(48_000, 2)),
         ]);
-        let w = plan_with_formats(&renders, &[], None, true, &p, 2);
+        let w = plan_with_formats(&renders, &[], None, true, &p, 2, &[]);
         assert_eq!(w.loopback_render.unwrap().0, "Speakers (Realtek HD Audio)");
     }
 
@@ -820,7 +853,7 @@ mod tests {
             ep("Voicemeeter Input (VB-Audio Voicemeeter VAIO)"),
         ];
         let captures = [ep("Voicemeeter Out B1 (VB-Audio Voicemeeter VAIO)")];
-        let w = plan(&renders, &captures, Some("voicemeeter input"), false);
+        let w = plan(&renders, &captures, Some("voicemeeter input"), false, &[]);
         assert_eq!(
             w.mic_render.unwrap().0,
             "Voicemeeter Input (VB-Audio Voicemeeter VAIO)"
@@ -836,7 +869,7 @@ mod tests {
     #[test]
     fn no_virtual_device() {
         let renders = [ep("Speakers (Realtek HD Audio)")];
-        let w = plan(&renders, &[], None, false);
+        let w = plan(&renders, &[], None, false, &[]);
         assert!(w.mic_render.is_none());
         assert_eq!(w.loopback_render.unwrap().0, "Speakers (Realtek HD Audio)");
     }
@@ -854,7 +887,7 @@ mod tests {
         ];
         let captures = [ep("Voicemeeter Out B1 (VB-Audio Voicemeeter VAIO)")];
         for host_audio in [false, true] {
-            let w = plan(&renders, &captures, None, host_audio);
+            let w = plan(&renders, &captures, None, host_audio, &[]);
             assert_eq!(
                 w.mic_render.as_ref().unwrap().0,
                 "Voicemeeter Input (VB-Audio Voicemeeter VAIO)",
@@ -877,7 +910,7 @@ mod tests {
             ep("Voicemeeter Aux Input (VB-Audio Voicemeeter AUX VAIO)"),
         ];
         for host_audio in [false, true] {
-            let w = plan(&renders, &[], None, host_audio);
+            let w = plan(&renders, &[], None, host_audio, &[]);
             assert!(w.mic_render.is_some(), "host_audio={host_audio}");
             assert!(w.loopback_render.is_none(), "host_audio={host_audio}");
         }
@@ -892,7 +925,7 @@ mod tests {
             ep("CABLE Input (VB-Audio Virtual Cable)"),
             ep("Speakers (Some Virtual Audio Device)"),
         ];
-        let w = plan(&renders, &[], None, false);
+        let w = plan(&renders, &[], None, false, &[]);
         assert!(w.loopback_render.is_none());
     }
 
@@ -918,7 +951,7 @@ mod tests {
         // Field shape minus the Speakers (mic holds the Streaming Microphone, nothing else).
         let renders = [ep("Altavoces (Steam Streaming Microphone)")];
         let captures = [ep("Microphone (Steam Streaming Microphone)")];
-        let w = plan(&renders, &captures, None, false);
+        let w = plan(&renders, &captures, None, false, &[]);
         assert!(w.loopback_unsatisfiable());
         let msg = describe_no_loopback(&renders, &w);
         assert!(msg.contains("reserved for the virtual mic"), "{msg}");
@@ -929,10 +962,70 @@ mod tests {
         // anyway), while the Steam pair is the remedy that adds a capturable sink.
         let renders = [ep("CABLE Input (VB-Audio Virtual Cable)")];
         let captures = [ep("CABLE Output (VB-Audio Virtual Cable)")];
-        let w = plan(&renders, &captures, None, false);
+        let w = plan(&renders, &captures, None, false, &[]);
         assert!(w.loopback_unsatisfiable());
         let msg = describe_no_loopback(&renders, &w);
         assert!(msg.contains("install Steam"), "{msg}");
         assert!(!msg.contains("install VB-Audio Virtual Cable"), "{msg}");
     }
+
+    /// A stamped pad endpoint is invisible to the plan. Its name carries NO virtual marker — on
+    /// purpose, games must read it as the pad's speaker — so the name rules alone would classify
+    /// it as real hardware and hand it the loopback; only the id exclusion prevents that.
+    /// Measured fact: the wiring plan on the target box already enumerated a stamped endpoint.
+    #[test]
+    fn pad_endpoints_invisible() {
+        let renders = [
+            ep("DualSense Wireless Controller"),
+            ep("Speakers (Realtek HD Audio)"),
+        ];
+        let pads = [renders[0].1.clone()];
+        let w = plan(&renders, &[], None, false, &pads);
+        assert_eq!(w.loopback_render.unwrap().0, "Speakers (Realtek HD Audio)");
+        // Even an operator mic override matching the pad's name must not claim it; with the
+        // pad as the only render endpoint there is honestly no mic target and no loopback.
+        let w = plan(
+            &renders[..1],
+            &[],
+            Some("wireless controller"),
+            false,
+            &pads,
+        );
+        assert!(w.mic_render.is_none());
+        assert!(w.loopback_render.is_none());
+    }
+
+    /// The exclusion has to survive the LAST RESORT tier, which this merge introduced alongside
+    /// pad audio. `last_resort` matches on the Steam-Speakers name, but it reads the same
+    /// shadowed `renders`, so a pad can never be reached through it either — otherwise the whole
+    /// desktop mix would be routed into the controller's voice coils.
+    #[test]
+    fn a_pad_is_never_the_last_resort() {
+        // Only the pad and the Steam pair exist; the mic reserves the Streaming Microphone, so
+        // the plan falls all the way through to the last resort.
+        let renders = [
+            ep("DualSense Wireless Controller"),
+            ep("Speakers (Steam Streaming Microphone)"),
+            ep("Speakers (Steam Streaming Speakers)"),
+        ];
+        let captures = [ep("Microphone (Steam Streaming Microphone)")];
+        let pads = [renders[0].1.clone()];
+        let w = plan(&renders, &captures, None, false, &pads);
+        assert_eq!(
+            w.loopback_render.as_ref().unwrap().0,
+            "Speakers (Steam Streaming Speakers)",
+            "the last resort must skip the pad"
+        );
+        assert!(w.loopback_last_resort);
+
+        // …and with the pad as the ONLY candidate left, the plan stays honestly unsatisfiable
+        // rather than falling back onto the coils.
+        let w = plan(&renders[..1], &captures, None, false, &pads);
+        assert!(
+            w.loopback_render.is_none(),
+            "a pad was taken as the last resort"
+        );
+        assert!(!w.loopback_last_resort);
+        assert!(w.loopback_unsatisfiable());
+    }
 }

crates/punktfunk-host/src/devtest.rs

@@ -384,6 +384,7 @@ pub fn dualsense_windows_test(args: &[String]) -> Result<()> {
                 index: idx,
                 kind: 2,
                 capabilities: 0,
+                audio_caps: 0,
             });
             println!(
                 "virtual {} up — cycling Cross + sweeping the left stick for {secs}s. Watch \
@@ -430,6 +431,7 @@ pub fn dualsense_windows_test(args: &[String]) -> Result<()> {
             index: idx,
             kind: 1,
             capabilities: 0,
+            audio_caps: 0,
         });
         println!(
             "virtual Xbox 360 (XUSB) up — sweeping LS + toggling A for {secs}s. Check with \
@@ -486,6 +488,119 @@ pub fn dualsense_windows_test(args: &[String]) -> Result<()> {
     Ok(())
 }
 
+/// Windows: pad-audio endpoint provisioning — `pad-endpoint ensure|remove|status [--index N]`.
+/// `ensure` runs the idempotent startup path (reuse-or-create the devnode, bind the Steam
+/// Streaming Speakers driver, stamp the DualSense identity + 4ch/48k formats, report whether
+/// the stamps are SERVED); `status` prints the devnode/endpoint and per-stamp stored vs served
+/// state without changing anything; `remove` deletes the devnode via pnputil — the escape
+/// hatch only, endpoints are persistent by design. Stamping needs SYSTEM (the MMDevices ACL);
+/// run `ensure` under the service account or PsExec when the property-store route is denied.
+#[cfg(target_os = "windows")]
+pub fn pad_endpoint(args: &[String]) -> Result<()> {
+    use crate::audio::pad_endpoint as pe;
+    let idx: u8 = args
+        .iter()
+        .skip_while(|a| *a != "--index")
+        .nth(1)
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(0);
+    // `--endpoint <id>` drives ANY render endpoint, not just a provisioned pad one. It is the
+    // discriminator between "this process cannot activate anything" and "our endpoint is broken":
+    // aim the same binary at a known-good endpoint and see whether it succeeds there.
+    let endpoint_override: Option<String> = args
+        .iter()
+        .skip_while(|a| *a != "--endpoint")
+        .nth(1)
+        .cloned();
+    match args.get(1).map(String::as_str) {
+        Some("ensure") => {
+            let p = pe::ensure(idx)?;
+            println!(
+                "pad-endpoint ensure: pad {} devnode {} endpoint {} needs_aeb_kick={}",
+                p.pad_index, p.device_instance, p.endpoint_id, p.needs_aeb_kick
+            );
+            Ok(())
+        }
+        Some("remove") => match pe::find(idx)? {
+            Some(p) => {
+                pe::remove(&p);
+                println!(
+                    "pad-endpoint remove: requested removal of {}",
+                    p.device_instance
+                );
+                Ok(())
+            }
+            None => {
+                println!("pad-endpoint remove: no pad-audio devnode for index {idx}");
+                Ok(())
+            }
+        },
+        // `punktfunk-host pad-endpoint <n> tone [seconds] [hz]` — drive the endpoint directly so
+        // the whole pad-audio chain can be exercised without a game. Without this, every attempt
+        // costs a game launch and a failure does not say which link broke.
+        Some("tone") => {
+            let secs: u32 = args.get(2).and_then(|s| s.parse().ok()).unwrap_or(5);
+            let hz: f32 = args.get(3).and_then(|s| s.parse().ok()).unwrap_or(60.0);
+            let endpoint_id = match endpoint_override {
+                Some(id) => id,
+                None => {
+                    // `find` (a system lookup), NOT `endpoint_for` (the service's in-process
+                    // cache): this runs as a separate CLI process and has no cache of its own.
+                    let Some(ep) = pe::find(idx)? else {
+                        println!(
+                            "pad-endpoint tone: no pad-audio devnode for pad {idx} — run \
+                             `ensure` first"
+                        );
+                        return Ok(());
+                    };
+                    if ep.endpoint_id.is_empty() {
+                        println!("pad-endpoint tone: pad {idx} has no endpoint id yet");
+                        return Ok(());
+                    }
+                    ep.endpoint_id
+                }
+            };
+            // `--pair front` drives the pad's SPEAKER instead of the voice coils — the only way to
+            // exercise the speaker kind without a game that renders one.
+            let pair = args
+                .iter()
+                .skip_while(|a| *a != "--pair")
+                .nth(1)
+                .map_or(pe::TonePair::Back, |s| pe::TonePair::parse(s));
+            println!(
+                "pad-endpoint tone: {hz} Hz into the {} of {endpoint_id} for {secs}s",
+                pair.label()
+            );
+            pe::render_test_tone(&endpoint_id, secs, hz, pair)?;
+            println!(
+                "pad-endpoint tone: done. A connected client with pad audio enabled should have \
+                 buzzed; the host log shows whether the gate opened."
+            );
+            Ok(())
+        }
+        // `punktfunk-host pad-endpoint capture [seconds]` — the receiving half of `tone`. Run
+        // both at once to exercise render -> engine -> loopback -> pair routing with no game and
+        // no client attached.
+        Some("capture") => {
+            let secs: u32 = args.get(2).and_then(|s| s.parse().ok()).unwrap_or(5);
+            let endpoint_id = match endpoint_override {
+                Some(id) => id,
+                None => match pe::find(idx)? {
+                    Some(ep) if !ep.endpoint_id.is_empty() => ep.endpoint_id,
+                    _ => {
+                        println!("pad-endpoint capture: pad {idx} has no endpoint — run `ensure`");
+                        return Ok(());
+                    }
+                },
+            };
+            println!("pad-endpoint capture: listening on {endpoint_id} for {secs}s");
+            pe::capture_probe(&endpoint_id, secs)
+        }
+        Some("status") => pe::print_status(idx),
+        _ => anyhow::bail!("usage: punktfunk-host pad-endpoint <ensure|remove|status> [--index N]"),
+    }
+}
+
 /// Mirror a physical monitor and pull frames from it — the on-glass gate for per-monitor capture
 /// (`design/per-monitor-portal-capture.md` P2/P3), without needing a client to connect.
 ///

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

@@ -65,6 +65,8 @@ pub fn decode(plaintext: &[u8]) -> Option<GamepadEvent> {
             index: *b.first()?,
             kind: *b.get(1)?,
             capabilities: le16(2)? as u16,
+            // GameStream's LI_CCAP vocabulary can't express pad audio — native-plane only.
+            audio_caps: 0,
         }),
         _ => None,
     }
@@ -138,6 +140,7 @@ mod tests {
             index,
             kind,
             capabilities,
+            ..
         }) = decode(&wrap(MAGIC_CONTROLLER_ARRIVAL, &body))
         else {
             panic!("expected Arrival");

crates/punktfunk-host/src/main.rs

@@ -618,6 +618,10 @@ fn real_main() -> Result<()> {
         // hold it, driving the real *WindowsManager end to end. `--index N`, `--seconds N`.
         #[cfg(target_os = "windows")]
         Some("dualsense-windows-test") => devtest::dualsense_windows_test(&args),
+        // Windows: pad-audio endpoint provisioning (`ensure`/`status`) + the pnputil removal
+        // escape hatch (`remove`). `--index N` selects the pad slot (default 0).
+        #[cfg(target_os = "windows")]
+        Some("pad-endpoint") => devtest::pad_endpoint(&args),
         // Capture→encode→file pipeline spike (dev tool).
         Some("spike") => spike::run(parse_spike(&args[1..])?),
         // Native punktfunk/1 host (QUIC control plane + UDP data plane).

crates/punktfunk-host/src/native.rs

@@ -62,6 +62,12 @@ use pairing::pair_ceremony;
 mod audio;
 use audio::audio_thread;
 
+/// Per-pad DualSense audio (the 0xD1 plane): loopback capture of the pre-provisioned pad
+/// endpoints → per-kind silence gate → stereo Opus → `PAD_AUDIO_MAGIC` datagrams. The input
+/// thread spawns/reaps one streamer per arriving pad (`input`); the Welcome advertises the cap
+/// via `pad_audio::host_cap` (`handshake`).
+mod pad_audio;
+
 /// The native input plane (plan §W1); the session setup spawns `input_thread` and feeds it a
 /// channel of `ClientInput`. The `Pads` router + rumble live there too.
 mod input;
@@ -345,6 +351,14 @@ pub(crate) async fn serve(
     // binds its capture device) and self-heals when the backend dies (PipeWire restart, Windows
     // endpoint churn).
     let mic_service = crate::audio::MicPump::start();
+    // Windows, env-gated (PUNKTFUNK_PAD_AUDIO / _SLOTS): pre-provision the per-pad "DualSense
+    // speaker" render endpoints once per host lifetime — idempotent devnode + stamp work on a
+    // dedicated COM thread, results published for sessions to query by pad index
+    // (crate::audio::pad_endpoint::endpoint_for). If any stamp is stored-but-not-served, the
+    // worker performs ONE AudioEndpointBuilder+Audiosrv restart now, before any session exists.
+    // Failures log once and leave the feature off: pads still work, just without pad audio.
+    #[cfg(target_os = "windows")]
+    crate::audio::pad_endpoint::provision_at_startup();
     // Host-lifetime worker that fires debounced TV-session restores (the managed gamescope path
     // restores the box's autologin gaming session on idle, not per-disconnect — see
     // `vdisplay::restore_managed_session`). Held for serve()'s lifetime; dropping it stops it.
@@ -1203,9 +1217,14 @@ async fn serve_session(
     let input_handle = {
         let conn = conn.clone();
         let gamepad = welcome.gamepad;
+        // Pad audio (0xD1) negotiated: the Welcome advertised the cap (Windows + provisioned
+        // endpoints + the client asked — handshake reads `pad_audio::host_cap`). Read back off
+        // the Welcome rather than recomputed, so the input thread's spawns cannot disagree
+        // with what the client was told.
+        let pad_audio_on = welcome.host_caps & punktfunk_core::quic::HOST_CAP_PAD_AUDIO != 0;
         std::thread::Builder::new()
             .name("punktfunk1-input".into())
-            .spawn(move || input_thread(input_rx, conn, inj_tx, gamepad))
+            .spawn(move || input_thread(input_rx, conn, inj_tx, gamepad, pad_audio_on))
             .context("spawn input thread")?
     };
     // One reader for ALL client→host datagrams, demuxed by magic byte (two read_datagram loops

crates/punktfunk-host/src/native/handshake.rs

@@ -640,6 +640,16 @@ pub(super) async fn negotiate(
                 punktfunk_core::quic::HOST_CAP_AUDIO_RED
             } else {
                 0
+            }
+            // Per-pad DualSense audio (0xD1 + HidOutput::AudioCtl): granted only when the
+            // client asked AND this host can capture it — Windows with the feature enabled
+            // and at least one pad endpoint provisioned at startup. A capable client then
+            // marks its pads' renderers on their arrivals; the input thread streams toward
+            // exactly those pads (`super::pad_audio`).
+            | if super::pad_audio::host_cap(hello.client_caps) {
+                punktfunk_core::quic::HOST_CAP_PAD_AUDIO
+            } else {
+                0
             },
         // The negotiated session AEAD (resolved above) + its 32-byte key toward a ChaCha
         // client; toward everyone else cipher 0 keeps the Welcome byte-identical to the

crates/punktfunk-host/src/native/input.rs

@@ -515,6 +515,100 @@ impl Pads {
     }
 }
 
+/// Per-pad 0xD1 streamers (`super::pad_audio`), keyed by pad index like every per-pad table
+/// here (bounded by [`MAX_WIRE_PADS`]; only slots 0..4 can ever have a provisioned endpoint —
+/// `spawn` refuses the rest). Spawned when a negotiated session's DualSense-family arrival
+/// declares renderer bits, reaped on remove / re-declare / session teardown.
+struct PadAudioSlots {
+    /// `(kinds, handle)` per running pad — `kinds` is the arrival's audio-caps mask, kept so
+    /// an identical re-arrival (they are re-sent against datagram loss) is a no-op.
+    slots: [Option<(u8, pad_audio::PadAudioHandle)>; MAX_WIRE_PADS],
+    /// Kind-change restarts spent per pad this session (R3). The trigger is a client-sent
+    /// arrival, so without a ceiling the client decides how many WASAPI captures the host opens.
+    restarts: [u8; MAX_WIRE_PADS],
+}
+
+/// R3: how many times one pad may change its declared audio kinds before the host stops
+/// obliging. A real controller declares once at open and never again; the re-sent arrivals are
+/// identical and take the no-op path above, so this is only reached by a client that keeps
+/// changing its mind.
+const MAX_PAD_AUDIO_RESTARTS: u8 = 8;
+
+impl PadAudioSlots {
+    fn new() -> PadAudioSlots {
+        PadAudioSlots {
+            slots: std::array::from_fn(|_| None),
+            restarts: [0; MAX_WIRE_PADS],
+        }
+    }
+
+    /// Idempotent spawn: same kinds → keep the running streamer; changed kinds → restart with
+    /// the new mask; not running → spawn (a slot without an endpoint stays empty — bounded
+    /// retries, since arrivals are only re-sent a few times per slot open).
+    fn ensure(&mut self, conn: &quinn::Connection, pad: u8, kinds: u8) {
+        let idx = pad as usize;
+        if idx >= MAX_WIRE_PADS {
+            return;
+        }
+        if let Some((have, _)) = &self.slots[idx] {
+            if *have == kinds {
+                return; // identical re-arrival — keep the running streamer
+            }
+            // R3: the restart trigger is a CLIENT-sent arrival, so the count is client-driven.
+            // Nothing bounded it: a client alternating its declared kinds could make the host
+            // tear down and re-spawn a WASAPI loopback capture indefinitely, each cycle paying a
+            // thread spawn and an endpoint activation. Cheap to bound, and a pad that has already
+            // changed its mind this many times in one session is not doing anything legitimate.
+            if self.restarts[idx] >= MAX_PAD_AUDIO_RESTARTS {
+                tracing::warn!(
+                    pad = idx,
+                    "pad-audio kinds changed again after {MAX_PAD_AUDIO_RESTARTS} restarts — \
+                     ignoring; the streamer keeps its current kinds for this session"
+                );
+                return;
+            }
+            self.restarts[idx] += 1;
+            tracing::info!(
+                pad = idx,
+                restarts = self.restarts[idx],
+                "pad-audio kinds changed — restarting the streamer"
+            );
+            self.stop(idx);
+        }
+        let stop = Arc::new(AtomicBool::new(false));
+        if let Some(h) = pad_audio::spawn(conn.clone(), pad, kinds, stop) {
+            self.slots[idx] = Some((kinds, h));
+        }
+    }
+
+    /// Stop + reap one pad's streamer. The join rides a detached reaper thread: a quiet pad's
+    /// capturer can sit out its ~5 s recv timeout, and this thread must keep its ≤4 ms
+    /// feedback cadence (games block on GET_REPORT handshakes) — the reaper still joins, just
+    /// not here. A failed reaper spawn falls back to the handle's own drop (signal + join).
+    fn stop(&mut self, idx: usize) {
+        if let Some((_, h)) = self.slots.get_mut(idx).and_then(|s| s.take()) {
+            h.signal();
+            let _ = std::thread::Builder::new()
+                .name("punktfunk1-padreap".into())
+                .spawn(move || h.stop());
+        }
+    }
+
+    /// Session teardown: flag every streamer FIRST so they wind down concurrently, then join —
+    /// the worst case is ONE quiet-endpoint recv timeout (~5 s), well inside the session's
+    /// 10 s side-thread join grace, not one per pad.
+    fn stop_all(&mut self) {
+        for s in self.slots.iter().flatten() {
+            s.1.signal();
+        }
+        for s in &mut self.slots {
+            if let Some((_, h)) = s.take() {
+                h.stop();
+            }
+        }
+    }
+}
+
 /// One client→host input item, both planes on ONE channel so the input thread wakes the
 /// moment either arrives (a second rich channel drained after the 4 ms recv timeout cost
 /// every pure-gyro motion sample up to 4 ms of quantization).
@@ -683,8 +777,13 @@ pub(super) fn input_thread(
     conn: quinn::Connection,
     inj_tx: std::sync::mpsc::Sender<InputEvent>,
     gamepad: GamepadPref,
+    pad_audio_on: bool,
 ) {
     let mut pads = Pads::new(gamepad);
+    // Per-pad 0xD1 audio streamers, live only when the Welcome granted the cap (`pad_audio_on`
+    // — read back off the negotiated host_caps). Spawned on DualSense-family arrivals that
+    // declare renderer bits, reaped on remove/teardown below.
+    let mut pad_streams = PadAudioSlots::new();
     // Motion-cadence observability (debug level): inter-arrival percentiles per 5 s window,
     // the measurement a "gyro feels floaty" report needs. Bounded: 5 s at even a 1 kHz pad
     // is 5000 u32s.
@@ -854,16 +953,53 @@ pub(super) fn input_thread(
                             &mut rumble_seen[idx],
                             &mut rumble_stop_burst[idx],
                         );
+                        // The unplugged pad's 0xD1 streamer goes with it (seq-gated like the
+                        // rest of this arm, so a reordered stale removal can't kill the
+                        // stream of a re-plugged pad). A re-plug re-arrives and re-spawns.
+                        pad_streams.stop(idx);
                     }
                 }
                 InputKind::GamepadArrival => {
                     // Per-pad controller kind declaration (mixed types): route this pad's future
-                    // frames to a backend of the declared kind. `code` = the GamepadPref wire byte,
-                    // `flags` = pad index. Applied before the pad's first frame (the client sends it
-                    // on slot open), so the device is built as the right type from the start.
-                    let idx = ev.flags as usize;
+                    // frames to a backend of the declared kind. `code` = the GamepadPref wire
+                    // byte, `flags` = pad index in the LOW BYTE — bits 8/9 carry the pad's
+                    // audio-render caps (haptics/speaker) from a pad-audio-capable client, so
+                    // the index MUST come from `decode_gamepad_arrival`, never the whole word.
+                    // Applied before the pad's first frame (the client sends it on slot open),
+                    // so the device is built as the right type from the start. The audio caps
+                    // are surfaced here for the 0xD1 capture path (which emits pad audio only
+                    // toward pads that declared a renderer).
+                    let (pad, audio_caps) = punktfunk_core::input::decode_gamepad_arrival(ev.flags);
+                    let idx = pad as usize;
                     let kind = GamepadPref::from_u8(ev.code as u8);
+                    if audio_caps != 0 {
+                        tracing::debug!(
+                            pad = idx,
+                            haptics = audio_caps & 0x01 != 0,
+                            speaker = audio_caps & 0x02 != 0,
+                            "pad-audio render caps declared (arrival flags bits 8/9)"
+                        );
+                    }
                     pads.set_kind(idx, kind);
+                    // Pad audio (0xD1): stream toward DualSense-family pads that declared a
+                    // renderer, only on a session that negotiated the cap. Idempotent across
+                    // the arrival re-sends (same kinds keeps the running streamer); a
+                    // re-declare without bits — or as a kind with no pad audio — stops it.
+                    if pad_audio_on {
+                        let want = if matches!(
+                            kind,
+                            GamepadPref::DualSense | GamepadPref::DualSenseEdge
+                        ) {
+                            audio_caps
+                        } else {
+                            0
+                        };
+                        if want != 0 {
+                            pad_streams.ensure(&conn, pad, want);
+                        } else {
+                            pad_streams.stop(idx);
+                        }
+                    }
                 }
                 _ => {
                     // Track press/release so a mid-press disconnect can be undone below.
@@ -1019,6 +1155,9 @@ pub(super) fn input_thread(
             flags: 0,
         });
     }
+    // Reap the per-pad 0xD1 streamers with the session (after the instant release sends above
+    // — this can block on a quiet pad's capturer timeout, see PadAudioSlots::stop_all).
+    pad_streams.stop_all();
 }
 
 #[cfg(test)]

crates/punktfunk-host/src/native/pad_audio.rs

@@ -0,0 +1,662 @@
+//! Per-pad DualSense audio (the 0xD1 pad-audio plane): WASAPI loopback of a pre-provisioned pad
+//! endpoint ([`crate::audio::pad_endpoint`]) → 4-ch de-interleave into the speaker (front) and
+//! voice-coil haptics (back) pairs → per-kind silence gate → stereo Opus (48 kHz, CBR, LowDelay)
+//! → [`PAD_AUDIO_MAGIC`](punktfunk_core::quic::PAD_AUDIO_MAGIC) datagrams. One thread per
+//! arriving pad, spawned/reaped by the input thread ([`super::input`]) as arrivals declare
+//! renderers and pads leave. Modeled on the session audio thread ([`super::audio`]): the same
+//! reopen-with-backoff on capture death, the same monotonic-seq-kept-across-reopens discipline,
+//! the same power-of-two encode-warn throttle.
+
+use super::*;
+
+/// `kinds` bit for the haptics stream (bit N = wire kind N — the same packing the arrival's
+/// audio-caps bits use, see [`punktfunk_core::input::decode_gamepad_arrival`]).
+#[cfg(any(target_os = "windows", test))]
+pub(super) const KIND_BIT_HAPTICS: u8 = 1 << punktfunk_core::quic::PAD_AUDIO_KIND_HAPTICS;
+/// `kinds` bit for the speaker stream.
+#[cfg(any(target_os = "windows", test))]
+pub(super) const KIND_BIT_SPEAKER: u8 = 1 << punktfunk_core::quic::PAD_AUDIO_KIND_SPEAKER;
+
+/// Haptics frames are 5 ms (the session-audio cadence — haptics are felt latency); speaker
+/// frames are 10 ms (speaker content tolerates the buffering for the coding efficiency). Both
+/// are the wire contract's cadences (`punktfunk_core::quic::PAD_AUDIO_KIND_*`).
+#[cfg(any(target_os = "windows", test))]
+const HAPTICS_FRAME_MS: u32 = 5;
+#[cfg(any(target_os = "windows", test))]
+const SPEAKER_FRAME_MS: u32 = 10;
+/// Samples per frame (per channel) at 48 kHz: 240 / 480.
+#[cfg(any(target_os = "windows", test))]
+const HAPTICS_FRAME_SAMPLES: usize =
+    crate::audio::SAMPLE_RATE as usize * HAPTICS_FRAME_MS as usize / 1000;
+#[cfg(any(target_os = "windows", test))]
+const SPEAKER_FRAME_SAMPLES: usize =
+    crate::audio::SAMPLE_RATE as usize * SPEAKER_FRAME_MS as usize / 1000;
+/// The capture's channel count — the pad endpoint is stamped quad (FL FR BL BR: front pair =
+/// speaker, back pair = voice coils). Mirrors `pad_endpoint::PAD_CHANNELS` (Windows-gated, so
+/// the pure splitter logic keeps its own copy).
+#[cfg(any(target_os = "windows", test))]
+const CAP_CHANNELS: usize = 4;
+
+/// Peak (absolute sample) at or above which a frame counts as signal — the gate OPENS on that
+/// very frame (haptics are felt latency; the first active frame must ship). ≈ −60 dBFS.
+#[cfg(any(target_os = "windows", test))]
+const GATE_OPEN_PEAK: f32 = 1e-3;
+/// How long the gate keeps sending after the last signal frame before it CLOSES (hangover):
+/// long enough that a decaying haptic tail (and the client decoder's own tail) is never
+/// clipped, short enough that an idle pad costs nothing in steady state.
+#[cfg(any(target_os = "windows", test))]
+const GATE_HANGOVER_MS: u32 = 250;
+
+/// Per-kind Opus bitrate — a stereo voice-coil / pad-speaker pair needs far less than the
+/// session plane's 128 kbps; 64 kbps CBR keeps every frame comfortably under one MTU.
+#[cfg(target_os = "windows")]
+const PAD_AUDIO_BITRATE: i32 = 64_000;
+
+/// The per-kind silence gate — the steady-state-cost feature: an idle pad endpoint (games
+/// rarely render pad audio) must cost ZERO encodes and ZERO datagrams, not a permanent 200 Hz
+/// stream of coded silence. Opens the instant a frame carries signal ([`GATE_OPEN_PEAK`]);
+/// closes only after [`GATE_HANGOVER_MS`] of continuous sub-threshold frames. Pure logic,
+/// unit-tested below.
+#[cfg(any(target_os = "windows", test))]
+struct SilenceGate {
+    /// Consecutive sub-threshold frames that close the gate ([`GATE_HANGOVER_MS`] ÷ frame ms).
+    hangover_frames: u32,
+    /// Consecutive sub-threshold frames seen so far while open.
+    quiet: u32,
+    /// Starts closed: a pad no game ever renders into never opens (and never sends).
+    open: bool,
+}
+
+#[cfg(any(target_os = "windows", test))]
+impl SilenceGate {
+    fn new(frame_ms: u32) -> SilenceGate {
+        SilenceGate {
+            hangover_frames: (GATE_HANGOVER_MS / frame_ms).max(1),
+            quiet: 0,
+            open: false,
+        }
+    }
+
+    /// Feed one frame; `true` = encode + send it. Signal opens the gate on THIS frame; the
+    /// frame that completes the hangover closes it and is itself suppressed (the client
+    /// already has ~250 ms of ramped-out silence by then).
+    fn feed(&mut self, frame: &[f32]) -> bool {
+        if frame.iter().any(|s| s.abs() >= GATE_OPEN_PEAK) {
+            self.open = true;
+            self.quiet = 0;
+        } else if self.open {
+            self.quiet += 1;
+            if self.quiet >= self.hangover_frames {
+                self.open = false;
+                self.quiet = 0;
+            }
+        }
+        self.open
+    }
+}
+
+/// One kind's send-admission + seq bookkeeping (pure logic — the capture thread wraps it with
+/// the encoder and the datagram send). `seq` is monotonic per (pad, kind) and NEVER advances
+/// while the gate is closed: frozen-seq = deliberate silence — the client tells silence from
+/// loss by seq continuity (the mic-mute discipline, pf-client-core/src/audio.rs). It is also
+/// kept across capture reopens (the session audio thread's discipline, audio.rs): the client
+/// sees a gap, not a restart.
+#[cfg(any(target_os = "windows", test))]
+struct LaneCtl {
+    gate: SilenceGate,
+    seq: u32,
+}
+
+#[cfg(any(target_os = "windows", test))]
+impl LaneCtl {
+    fn new(frame_ms: u32) -> LaneCtl {
+        LaneCtl {
+            gate: SilenceGate::new(frame_ms),
+            seq: 0,
+        }
+    }
+
+    /// Admit one frame: `Some(seq)` = encode + send it with this seq (advanced for the next);
+    /// `None` = gated — do not send, do not advance. An encode failure AFTER admission leaves a
+    /// one-frame seq gap, which the client conceals exactly like datagram loss.
+    fn admit(&mut self, frame: &[f32]) -> Option<u32> {
+        if !self.gate.feed(frame) {
+            return None;
+        }
+        let seq = self.seq;
+        self.seq = self.seq.wrapping_add(1);
+        Some(seq)
+    }
+}
+
+/// De-interleave one 4-ch block (FL FR BL BR) into its stereo pairs: `(front, back)` — front =
+/// speaker (channels 0/1), back = voice-coil haptics (channels 2/3). A ragged tail (not a
+/// multiple of 4 — the capturer only ever delivers whole frames) is dropped, never smeared
+/// across channels.
+#[cfg(any(target_os = "windows", test))]
+fn split_quad(block: &[f32]) -> (Vec<f32>, Vec<f32>) {
+    let mut front = Vec::with_capacity(block.len() / 2);
+    let mut back = Vec::with_capacity(block.len() / 2);
+    for s in block.chunks_exact(CAP_CHANNELS) {
+        front.extend_from_slice(&s[..2]);
+        back.extend_from_slice(&s[2..4]);
+    }
+    (front, back)
+}
+
+/// Accumulates interleaved 4-ch capture and cuts it into the wire contract's per-kind stereo
+/// frames — haptics every 5 ms from the back pair, speaker every 10 ms from the front pair —
+/// emitting ONLY the kinds enabled in `kinds` (a disabled kind is never even split out, so it
+/// can never reach an encoder). Pure logic, unit-tested; the capture thread wraps it.
+#[cfg(any(target_os = "windows", test))]
+struct PadFramer {
+    kinds: u8,
+    /// Raw interleaved 4-ch accumulation, drained in 5 ms blocks.
+    acc: Vec<f32>,
+    /// Front-pair stereo accumulation toward the next 10 ms speaker frame.
+    front: Vec<f32>,
+}
+
+#[cfg(any(target_os = "windows", test))]
+impl PadFramer {
+    fn new(kinds: u8) -> PadFramer {
+        PadFramer {
+            kinds,
+            acc: Vec::with_capacity(HAPTICS_FRAME_SAMPLES * CAP_CHANNELS * 4),
+            front: Vec::new(),
+        }
+    }
+
+    /// Feed one capture chunk; `emit(kind, stereo_frame)` fires for each completed frame
+    /// (haptics first — it is the latency-critical pair).
+    fn feed(&mut self, chunk: &[f32], mut emit: impl FnMut(u8, &[f32])) {
+        self.acc.extend_from_slice(chunk);
+        let block_len = HAPTICS_FRAME_SAMPLES * CAP_CHANNELS;
+        while self.acc.len() >= block_len {
+            let block: Vec<f32> = self.acc.drain(..block_len).collect();
+            let (front, back) = split_quad(&block);
+            if self.kinds & KIND_BIT_HAPTICS != 0 {
+                emit(punktfunk_core::quic::PAD_AUDIO_KIND_HAPTICS, &back);
+            }
+            if self.kinds & KIND_BIT_SPEAKER != 0 {
+                self.front.extend_from_slice(&front);
+                let frame_len = SPEAKER_FRAME_SAMPLES * 2;
+                while self.front.len() >= frame_len {
+                    let frame: Vec<f32> = self.front.drain(..frame_len).collect();
+                    emit(punktfunk_core::quic::PAD_AUDIO_KIND_SPEAKER, &frame);
+                }
+            }
+        }
+    }
+
+    /// Drop the partial frames straddling a capture gap (reopen). The seq/gate state is NOT
+    /// here — [`LaneCtl`] deliberately survives reopens, so the client sees a gap, not a
+    /// restart.
+    fn clear(&mut self) {
+        self.acc.clear();
+        self.front.clear();
+    }
+}
+
+/// A running per-pad streamer. [`stop`](PadAudioHandle::stop) (or drop) flags the thread and
+/// joins it; [`signal`](PadAudioHandle::signal) only flags — the input thread's teardown flags
+/// every pad first so the joins overlap instead of serializing the capturer's worst-case ~5 s
+/// quiet-endpoint recv timeout.
+pub(super) struct PadAudioHandle {
+    stop: Arc<AtomicBool>,
+    join: Option<std::thread::JoinHandle<()>>,
+}
+
+impl PadAudioHandle {
+    /// Flag the streamer to wind down without waiting for it.
+    pub(super) fn signal(&self) {
+        self.stop.store(true, Ordering::SeqCst);
+    }
+
+    /// Stop + reap. Bounded by the capturer's ~5 s quiet-endpoint recv timeout in the worst
+    /// case — the mid-session reap paths run this on a detached reaper thread for that reason
+    /// (`input.rs::PadAudioSlots::stop`); session teardown affords it inline (the 10 s
+    /// side-thread join grace covers it).
+    pub(super) fn stop(mut self) {
+        self.reap();
+    }
+
+    fn reap(&mut self) {
+        self.signal();
+        if let Some(join) = self.join.take() {
+            let _ = join.join();
+        }
+    }
+}
+
+/// A handle dropped without `stop()` (reaper-spawn failure) still winds its thread down.
+impl Drop for PadAudioHandle {
+    fn drop(&mut self) {
+        self.reap();
+    }
+}
+
+/// Whether this session's Welcome should advertise
+/// [`HOST_CAP_PAD_AUDIO`](punktfunk_core::quic::HOST_CAP_PAD_AUDIO): the client asked
+/// ([`CLIENT_CAP_PAD_AUDIO`](punktfunk_core::quic::CLIENT_CAP_PAD_AUDIO)), this is a Windows
+/// host with the feature on (`PUNKTFUNK_PAD_AUDIO` != "0"), and startup provisioning published
+/// at least one endpoint (`pad_endpoint::provision_at_startup`). Still-running provisioning
+/// reads as "none yet": a session racing host startup simply negotiates without pad audio and
+/// picks it up on its next connect.
+pub(super) fn host_cap(client_caps: u8) -> bool {
+    let asked = client_caps & punktfunk_core::quic::CLIENT_CAP_PAD_AUDIO != 0;
+    #[cfg(target_os = "windows")]
+    {
+        // R5: a startup attempt that failed transiently leaves nothing latched, so retry here —
+        // this is the first moment in a session's life that anyone asks whether pad audio exists.
+        if asked {
+            crate::audio::pad_endpoint::ensure_provisioned();
+        }
+        asked
+            && std::env::var_os("PUNKTFUNK_PAD_AUDIO").is_none_or(|v| v != "0")
+            && crate::audio::pad_endpoint::provisioned_endpoints()
+                .is_some_and(|eps| !eps.is_empty())
+    }
+    #[cfg(not(target_os = "windows"))]
+    {
+        // Only the Windows virtual DualSense exposes pad audio endpoints today.
+        let _ = asked;
+        false
+    }
+}
+
+/// Start the per-pad streamer toward `conn` for `pad`, streaming the kinds in `kinds` (bit 0 =
+/// haptics, bit 1 = speaker — the arrival's audio-caps packing). `stop` is this handle's own
+/// flag (fresh per spawn — pad streamers stop individually, not with the session). `None` when
+/// the slot has no provisioned endpoint (provisioning failed or still running, or the slot is
+/// past `PUNKTFUNK_PAD_AUDIO_SLOTS` — only 0..4 can ever have one) or the thread cannot spawn;
+/// the pad itself keeps working either way, just without audio.
+#[cfg(target_os = "windows")]
+pub(super) fn spawn(
+    conn: quinn::Connection,
+    pad: u8,
+    kinds: u8,
+    stop: Arc<AtomicBool>,
+) -> Option<PadAudioHandle> {
+    if kinds & (KIND_BIT_HAPTICS | KIND_BIT_SPEAKER) == 0 {
+        return None;
+    }
+    let Some(ep) = crate::audio::pad_endpoint::endpoint_for(pad) else {
+        tracing::debug!(
+            pad,
+            "pad-audio arrival for a slot without a provisioned endpoint — not streaming"
+        );
+        return None;
+    };
+    if ep.endpoint_id.is_empty() {
+        // The devnode-without-endpoint shape (`find`) — never in the provisioned set, but
+        // cheap to refuse rather than spin the open/backoff loop on an empty id.
+        return None;
+    }
+    if ep.needs_aeb_kick {
+        // R4: this flag was computed on every path and consulted nowhere past startup. It means
+        // the endpoint's stamps are STORED but not SERVED — the audio stack never picked up the
+        // DualSense identity — and startup's one restart did not fix it. Opening anyway is worse
+        // than refusing: `AUTOCONVERTPCM` makes a wrong-format endpoint initialize *successfully*,
+        // so the stream runs, the logs look healthy, and the haptics/speaker pair is mis-routed
+        // with nothing to point at. Decline, and say which reboot-shaped problem it is.
+        tracing::warn!(
+            pad,
+            endpoint = %ep.endpoint_id,
+            "pad endpoint stamps are stored but not served — the audio stack has not adopted the \
+             DualSense identity (a reboot, or a manual AudioEndpointBuilder+Audiosrv restart, \
+             clears it). Not streaming: the endpoint would open and mis-route."
+        );
+        return None;
+    }
+    let stop_t = stop.clone();
+    match std::thread::Builder::new()
+        .name(format!("punktfunk1-pad{pad}"))
+        .spawn(move || pad_audio_thread(conn, pad, kinds, ep.endpoint_id, stop_t))
+    {
+        Ok(join) => Some(PadAudioHandle {
+            stop,
+            join: Some(join),
+        }),
+        Err(e) => {
+            tracing::warn!(pad, error = %e, "pad-audio thread spawn failed — pad streams without audio");
+            None
+        }
+    }
+}
+
+/// Stub — pad endpoints exist only behind the Windows virtual DualSense; other hosts run pads
+/// without the audio side (and never advertise the cap, see [`host_cap`]).
+#[cfg(not(target_os = "windows"))]
+pub(super) fn spawn(
+    _conn: quinn::Connection,
+    _pad: u8,
+    _kinds: u8,
+    _stop: Arc<AtomicBool>,
+) -> Option<PadAudioHandle> {
+    None
+}
+
+/// One enabled kind's encoder lane: admission/seq control + its stereo Opus encoder + the
+/// power-of-two warn throttle (a stuck encoder would otherwise fail ~200 times a second).
+#[cfg(target_os = "windows")]
+struct Lane {
+    kind: u8,
+    ctl: LaneCtl,
+    enc: opus::Encoder,
+    encode_errs: u64,
+}
+
+/// Build one stereo encoder per enabled kind: 48 kHz LowDelay hard-CBR like the session audio
+/// plane ([`super::audio`]), at the pad plane's 64 kbps.
+#[cfg(target_os = "windows")]
+fn build_lanes(kinds: u8) -> Result<Vec<Lane>, opus::Error> {
+    let mut lanes = Vec::new();
+    for (bit, kind, frame_ms) in [
+        (
+            KIND_BIT_HAPTICS,
+            punktfunk_core::quic::PAD_AUDIO_KIND_HAPTICS,
+            HAPTICS_FRAME_MS,
+        ),
+        (
+            KIND_BIT_SPEAKER,
+            punktfunk_core::quic::PAD_AUDIO_KIND_SPEAKER,
+            SPEAKER_FRAME_MS,
+        ),
+    ] {
+        if kinds & bit == 0 {
+            continue;
+        }
+        let mut enc = opus::Encoder::new(
+            crate::audio::SAMPLE_RATE,
+            opus::Channels::Stereo,
+            opus::Application::LowDelay,
+        )?;
+        enc.set_bitrate(opus::Bitrate::Bits(PAD_AUDIO_BITRATE)).ok();
+        enc.set_vbr(false).ok();
+        lanes.push(Lane {
+            kind,
+            ctl: LaneCtl::new(frame_ms),
+            enc,
+            encode_errs: 0,
+        });
+    }
+    Ok(lanes)
+}
+
+/// The per-pad streaming thread: loopback capture → framer → per-kind gate/encode → 0xD1
+/// datagrams. Capture death reopens with the session-audio backoff ([`INJECTOR_REOPEN_BACKOFF`],
+/// encoders + seq kept); a send error ends the thread (the connection — the session — is gone).
+#[cfg(target_os = "windows")]
+fn pad_audio_thread(
+    conn: quinn::Connection,
+    pad: u8,
+    kinds: u8,
+    endpoint_id: String,
+    stop: Arc<AtomicBool>,
+) {
+    use crate::audio::AudioCapturer as _;
+    let mut lanes = match build_lanes(kinds) {
+        Ok(l) => l,
+        Err(e) => {
+            tracing::warn!(pad, error = %e, "pad-audio opus encoder init failed — pad continues without audio");
+            return;
+        }
+    };
+    if lanes.is_empty() {
+        return; // spawn() refuses kinds == 0 — belt and braces
+    }
+    let mut framer = PadFramer::new(kinds);
+    // One Opus frame per datagram; 64 kbps CBR at ≤10 ms is ~80 bytes — sized with the session
+    // plane's slack.
+    let mut opus_buf = vec![0u8; 1500];
+    // Reopen-with-backoff (the audio.rs discipline): a capture death (endpoint invalidated,
+    // audio-engine restart) reopens instead of muting the pad for the rest of the session. The
+    // first open ALSO rides this loop, so an open lost to endpoint churn starts late, not never.
+    let mut capturer: Option<crate::audio::pad_endpoint::PadLoopbackCapturer> = None;
+    let mut last_failed: Option<std::time::Instant> = None;
+    tracing::info!(
+        pad,
+        haptics = kinds & KIND_BIT_HAPTICS != 0,
+        speaker = kinds & KIND_BIT_SPEAKER != 0,
+        "pad audio streaming (0xD1, Opus 48 kHz, silence-gated)"
+    );
+    '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::pad_endpoint::PadLoopbackCapturer::open(&endpoint_id) {
+                Ok(c) => {
+                    if last_failed.take().is_some() {
+                        tracing::info!(pad, "pad-audio capture reopened");
+                    }
+                    capturer = Some(c);
+                    framer.clear(); // drop the partial frames straddling the gap
+                }
+                Err(e) => {
+                    tracing::debug!(pad, error = %format!("{e:#}"), "pad-audio open failed — will retry");
+                    last_failed = Some(std::time::Instant::now());
+                    std::thread::sleep(std::time::Duration::from_millis(200));
+                    continue;
+                }
+            }
+        }
+        // An empty chunk is a QUIET endpoint (the capturer's idle timeout), not a death — keep
+        // it; only a genuine Err (capture thread ended) drops the capturer for reopen.
+        let chunk = match capturer.as_mut().unwrap().next_chunk() {
+            Ok(c) => c,
+            Err(e) => {
+                tracing::warn!(pad, error = %format!("{e:#}"), "pad-audio capture lost — reopening");
+                capturer = None;
+                last_failed = Some(std::time::Instant::now());
+                continue;
+            }
+        };
+        let mut session_gone = false;
+        framer.feed(&chunk, |kind, frame| {
+            if session_gone {
+                return;
+            }
+            let Some(lane) = lanes.iter_mut().find(|l| l.kind == kind) else {
+                return; // framer emits only enabled kinds — unreachable, but never panic here
+            };
+            // Gated = deliberate silence: no datagram AND a frozen seq (the client tells
+            // silence from loss by seq continuity).
+            let Some(seq) = lane.ctl.admit(frame) else {
+                return;
+            };
+            let pts_ns = now_ns();
+            match lane.enc.encode_float(frame, &mut opus_buf) {
+                Ok(n) => {
+                    let d = punktfunk_core::quic::encode_pad_audio_datagram(
+                        pad,
+                        kind,
+                        seq,
+                        pts_ns,
+                        &opus_buf[..n],
+                    );
+                    if conn.send_datagram(d.into()).is_err() {
+                        session_gone = true; // connection gone — the session is over
+                    }
+                }
+                Err(e) => {
+                    lane.encode_errs += 1;
+                    if lane.encode_errs.is_power_of_two() {
+                        tracing::warn!(
+                            pad,
+                            kind,
+                            error = %e,
+                            count = lane.encode_errs,
+                            "pad-audio opus encode failed — dropping frame"
+                        );
+                    }
+                }
+            }
+        });
+        if session_gone {
+            break 'session;
+        }
+    }
+    // Dropping the capturer stops its WASAPI thread. Nothing to park: pad capture is per-pad,
+    // per-session by design (unlike the session audio slot there is no cross-session reuse).
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use punktfunk_core::quic::{PAD_AUDIO_KIND_HAPTICS, PAD_AUDIO_KIND_SPEAKER};
+
+    /// A stereo frame of `n` samples at a constant level.
+    fn frame(level: f32, n: usize) -> Vec<f32> {
+        vec![level; n * 2]
+    }
+
+    #[test]
+    fn gate_opens_immediately_and_closes_after_hangover() {
+        let mut g = SilenceGate::new(HAPTICS_FRAME_MS);
+        // 250 ms of 5 ms frames.
+        assert_eq!(g.hangover_frames, 50);
+        // Closed from birth: an idle pad never sends.
+        assert!(!g.feed(&frame(0.0, HAPTICS_FRAME_SAMPLES)));
+        // A peak at exactly the threshold opens on THIS frame (haptics are felt latency).
+        assert!(g.feed(&frame(GATE_OPEN_PEAK, HAPTICS_FRAME_SAMPLES)));
+        // 49 quiet frames ride the hangover; the 50th completes 250 ms and is suppressed.
+        for _ in 0..49 {
+            assert!(g.feed(&frame(0.0, HAPTICS_FRAME_SAMPLES)));
+        }
+        assert!(!g.feed(&frame(0.0, HAPTICS_FRAME_SAMPLES)));
+        // ... and stays closed.
+        assert!(!g.feed(&frame(0.0, HAPTICS_FRAME_SAMPLES)));
+        // Sub-threshold wiggle does not reopen; real signal does (negative peaks count).
+        assert!(!g.feed(&frame(9e-4, HAPTICS_FRAME_SAMPLES)));
+        assert!(g.feed(&frame(-0.5, HAPTICS_FRAME_SAMPLES)));
+        // A loud frame mid-hangover rearms the full 250 ms.
+        for _ in 0..49 {
+            assert!(g.feed(&frame(0.0, HAPTICS_FRAME_SAMPLES)));
+        }
+        assert!(g.feed(&frame(0.02, HAPTICS_FRAME_SAMPLES)));
+        for _ in 0..49 {
+            assert!(g.feed(&frame(0.0, HAPTICS_FRAME_SAMPLES)));
+        }
+        assert!(!g.feed(&frame(0.0, HAPTICS_FRAME_SAMPLES)));
+    }
+
+    #[test]
+    fn gate_hangover_scales_with_frame_ms() {
+        let mut g = SilenceGate::new(SPEAKER_FRAME_MS);
+        assert_eq!(g.hangover_frames, 25); // 250 ms of 10 ms frames
+        assert!(g.feed(&frame(0.1, SPEAKER_FRAME_SAMPLES)));
+        for _ in 0..24 {
+            assert!(g.feed(&frame(0.0, SPEAKER_FRAME_SAMPLES)));
+        }
+        assert!(!g.feed(&frame(0.0, SPEAKER_FRAME_SAMPLES)));
+    }
+
+    #[test]
+    fn seq_freezes_while_gated_and_survives_reopen() {
+        let mut lane = LaneCtl::new(HAPTICS_FRAME_MS);
+        // Two audible frames: seq 0, 1.
+        assert_eq!(lane.admit(&frame(0.5, HAPTICS_FRAME_SAMPLES)), Some(0));
+        assert_eq!(lane.admit(&frame(0.5, HAPTICS_FRAME_SAMPLES)), Some(1));
+        // The hangover is still sent (seq advances), then the gate closes and seq FREEZES —
+        // deliberate silence the client tells from loss by continuity.
+        for i in 0..49u32 {
+            assert_eq!(lane.admit(&frame(0.0, HAPTICS_FRAME_SAMPLES)), Some(2 + i));
+        }
+        for _ in 0..500 {
+            assert_eq!(lane.admit(&frame(0.0, HAPTICS_FRAME_SAMPLES)), None);
+        }
+        // A capture reopen resets ONLY the framer (PadFramer::clear) — LaneCtl is deliberately
+        // untouched, so the next audible frame CONTINUES the sequence (gap, not restart).
+        assert_eq!(lane.admit(&frame(0.9, HAPTICS_FRAME_SAMPLES)), Some(51));
+    }
+
+    #[test]
+    fn splitter_exact_pairs() {
+        // Interleave [FL FR BL BR] × 2 frames with distinct values everywhere.
+        let quad = [0.0, 1.0, 2.0, 3.0, 10.0, 11.0, 12.0, 13.0];
+        let (front, back) = split_quad(&quad);
+        assert_eq!(front, [0.0, 1.0, 10.0, 11.0]);
+        assert_eq!(back, [2.0, 3.0, 12.0, 13.0]);
+        // A ragged tail (never produced by the capturer) is dropped, not smeared.
+        let (front, back) = split_quad(&quad[..7]);
+        assert_eq!((front.len(), back.len()), (2, 2));
+    }
+
+    #[test]
+    fn framer_cuts_the_wire_cadence() {
+        let mut f = PadFramer::new(KIND_BIT_HAPTICS | KIND_BIT_SPEAKER);
+        let mut got: Vec<(u8, usize, f32)> = Vec::new();
+        // 10 ms of capture (480 samples), fed in ragged chunks: exactly two 5 ms haptics
+        // frames from the back pair, then one 10 ms speaker frame from the front pair.
+        let mut quad = Vec::new();
+        for _ in 0..2 * HAPTICS_FRAME_SAMPLES {
+            quad.extend_from_slice(&[0.25, 0.25, -0.5, -0.5]);
+        }
+        for chunk in quad.chunks(101) {
+            f.feed(chunk, |kind, frame| got.push((kind, frame.len(), frame[0])));
+        }
+        assert_eq!(
+            got,
+            vec![
+                (PAD_AUDIO_KIND_HAPTICS, 2 * HAPTICS_FRAME_SAMPLES, -0.5),
+                (PAD_AUDIO_KIND_HAPTICS, 2 * HAPTICS_FRAME_SAMPLES, -0.5),
+                (PAD_AUDIO_KIND_SPEAKER, 2 * SPEAKER_FRAME_SAMPLES, 0.25),
+            ]
+        );
+    }
+
+    #[test]
+    fn framer_masks_disabled_kinds() {
+        // 20 ms of all-ones capture: 4 potential haptics frames, 2 potential speaker frames.
+        let quad = vec![1.0f32; 4 * HAPTICS_FRAME_SAMPLES * CAP_CHANNELS];
+        let mut kinds_seen = Vec::new();
+        // Haptics-only: the front pair is never split out, let alone encoded.
+        let mut f = PadFramer::new(KIND_BIT_HAPTICS);
+        f.feed(&quad, |kind, _| kinds_seen.push(kind));
+        assert_eq!(kinds_seen, vec![PAD_AUDIO_KIND_HAPTICS; 4]);
+        // Speaker-only: no haptics frames.
+        let mut f = PadFramer::new(KIND_BIT_SPEAKER);
+        kinds_seen.clear();
+        f.feed(&quad, |kind, _| kinds_seen.push(kind));
+        assert_eq!(kinds_seen, vec![PAD_AUDIO_KIND_SPEAKER; 2]);
+        // kinds = 0 is never spawned, but the framer must still be total: nothing comes out.
+        let mut f = PadFramer::new(0);
+        kinds_seen.clear();
+        f.feed(&quad, |kind, _| kinds_seen.push(kind));
+        assert!(kinds_seen.is_empty());
+    }
+
+    #[test]
+    fn framer_clear_drops_partials_only() {
+        let mut f = PadFramer::new(KIND_BIT_HAPTICS | KIND_BIT_SPEAKER);
+        let mut emitted = 0;
+        // 100 samples: no frame boundary reached yet.
+        f.feed(&vec![0.1; 100 * CAP_CHANNELS], |_, _| emitted += 1);
+        assert_eq!(emitted, 0);
+        f.clear();
+        // After the gap: exactly one haptics frame from 240 fresh samples — the 100 stale
+        // samples are gone (they would skew every later frame boundary).
+        f.feed(
+            &vec![0.2; HAPTICS_FRAME_SAMPLES * CAP_CHANNELS],
+            |kind, frame| {
+                emitted += 1;
+                assert_eq!(
+                    (kind, frame.len()),
+                    (PAD_AUDIO_KIND_HAPTICS, 2 * HAPTICS_FRAME_SAMPLES)
+                );
+            },
+        );
+        assert_eq!(emitted, 1);
+    }
+
+    #[test]
+    fn host_cap_requires_the_client_bit() {
+        // Without CLIENT_CAP_PAD_AUDIO the answer is no on EVERY platform (on Windows the
+        // env + provisioning legs are environment-dependent — not unit-tested here).
+        assert!(!host_cap(0));
+        assert!(!host_cap(punktfunk_core::quic::CLIENT_CAP_CURSOR));
+    }
+}

docs/releases/v0.25.0.md

@@ -0,0 +1,81 @@
+Wire-compatible with 0.24.x — everything you have already paired keeps working, and you can update one side at a time. Nothing here changes how a host and a client agree on what to send each other, so an old client on a new host (or the other way round) streams exactly as it does today; the parts that are new switch themselves on only once both ends have them.
+
+The headline is that a **DualSense plugged in by USB can now play a game's fine-grained haptics — the textured detail in the grips, not just the rumble motors — and its own speaker, streamed from the host**. That needs a Windows host with Steam installed and either the Android app or the desktop session client; everywhere else, nothing changes.
+
+Behind it, three fronts. **Controllers** were swept end to end: rumble that faded on a Steam Deck, died for good after one hiccup on a phone, or kept buzzing after you quit; adaptive triggers and lightbars left stuck in a game's last state on your desk after the stream ended; player-number lights that never lit on anything but a DualSense — more than twenty separate faults, across every client and both hosts. **Sound** got the same treatment: desktop audio is encoded at roughly double the bitrate, hosts stopped routing the entire game mix through Steam's voice channel on PCs that had it installed, and audio that drifts behind the picture now pulls itself back instead of staying late for the rest of the session. And the **black screen** that some people hit on VPN-shaped networks — a session that connects, reports every gauge healthy, and then shows nothing at all, forever — is finally diagnosed, explained in the log, and healed on its own. Alongside those: the Steam Deck plugin is rebuilt as a launcher into the app, holding Select on any controller presses the host's Guide button, and saved settings profiles can be pinned to hosts without a mouse.
+
+## New
+
+- **Your DualSense's own haptics, carried from the host.** Games that drive the DualSense's fine-grained voice coils — the detailed, textured feedback in the grips, as distinct from the coarse rumble motors — now carry that across the stream to the controller in your hands, and the pad's built-in speaker can be carried with it. It needs all of: a DualSense or DualSense Edge **plugged into your device by USB** (over Bluetooth the pad exposes no audio device to play into, so there is nothing this can do), a **Windows host with Steam installed** (the per-controller audio device is built on Valve's Remote Play streaming-speakers driver), and either the Android app or the desktop session client. Anywhere else — a Linux host, the iPhone/iPad/Mac app, the ordinary Windows or Linux desktop app, a Bluetooth pad — nothing changes at all. A game that uses only ordinary rumble keeps rumbling exactly as it does today. On Android the controls are **Controller haptics** and **Controller speaker** under Controllers, alongside a **Test haptics** button that checks your phone can drive the pad at all without needing a stream running. Expect a new playback device named "DualSense Wireless Controller" to appear in the host's Windows sound settings — that is this feature, it is how games find the controller's speaker, and it will not take over as your default output.
+- **Hold Select to press the host's Guide button.** Hold Select (Back / View) on its own for about a third of a second and the host sees its Guide button go down — and it stays down while you hold, so a longer hold reads as a long-press on the host, which is how a big-screen host opens its Quick Access Menu. A quick tap of Select still goes to the game, and Select as part of a combo — including the leave chord — passes through untouched. It is on by default on iPhone, iPad and Apple TV, where the system keeps the controller's own Home press for itself and this is the only reliable route to the host's overlay. Everywhere else the raw press already reaches the host, so the gesture stays off by default and Select keeps its exact timing. Update the client.
+- **Get onto a host by asking, instead of typing a PIN.** From the Steam Deck panel, tapping a locked host now offers **Request access**: the stream opens and waits while whoever is at the host approves your Deck in its console, then the picture comes up by itself. It gives up after about three minutes like any failed connection. Offered only for hosts visible on your network — one you saved by typing an address has no advertised identity to check against, so those still use a PIN, and the sheet says why. Update the plugin; hosts already knew how to approve.
+- **Pin a settings profile to a host from a controller.** Every controller-driven settings screen — the Deck and Linux console home, the Apple app's gamepad UI including Apple TV, and the Android app's controller UI including Android TV — gains a **Profiles** section showing each profile and where it is pinned ("Not pinned", "Pinned to 2 hosts"). Open one and press A on a host to pin or unpin. On Apple TV this is the only profile management there has ever been; on Android, pinning previously needed a touchscreen. Creating and editing profiles is still a desktop or phone job. Update the client.
+- **Pinned profiles appear as their own cards on the console home.** A pinned profile shows up as an extra card right after its host, subtitled with the profile's name, and one press connects using those settings. A host already bound to a profile now names it next to its address, so you can see which settings a plain press will use. Update the client.
+- **A lost audio packet is rebuilt exactly instead of being papered over.** Each audio packet can carry a copy of the one before it, so a single loss is reconstructed bit-for-bit rather than concealed with a synthesized approximation you can hear. It costs no extra delay — the copy rides on a packet that was already arriving in time. Needs 0.25.0 on both ends; with either side older, audio goes over the wire exactly as it did before.
+- **Audio quality is now budgeted against your connection.** The higher bitrate and the packet redundancy above are worth having on a roomy link and much too expensive on a narrow one, and audio is not managed by the Automatic bitrate control — whatever it takes comes off the top. The host now picks quality and redundancy together against the session's video bitrate: full quality plus redundancy where there is room, redundancy dropped first as the link narrows, then the quality tier, never below a floor. Update the host.
+- **Turn on Sony USB passthrough from a TV.** The DualSense / DualShock USB toggle only ever existed on the touch settings screen, so on an Android TV box there was no way to reach it at all. It now sits on the controller-driven screen beside the Steam Controller toggle. Update the client.
+- **Press the host's Steam and Quick Access buttons from the Deck panel.** While a stream is running the panel shows a **Host menus** section with **Steam menu on host** and **Quick access on host**; either one presses that button on the host and closes the Deck's own menu so the host's shows through. Update the plugin and the client.
+- **Two new command-line tools.** `punktfunk discover` lists the hosts on your network with their addresses, whether you have already saved them and whether you are paired, with a `--json` mode for scripts. `punktfunk launch <host> --request-access` is the Request access flow above from a terminal, for admitting a headless machine without a PIN. Update the client.
+- **Hosts on a jumbo-frame network can opt into much larger video packets.** On a LAN deliberately configured end to end for 9000-byte frames, the host can send roughly six times fewer packets per frame. It is off by default, is only applied after the host has proven the path really carries them and the client has agreed, and it reverts on its own if those packets start disappearing. This is not a general speed-up: on an ordinary network it does nothing.
+
+## Improved
+
+- **Desktop audio is encoded at roughly double the bitrate.** Streamed sound now runs at 256 kbps in stereo rather than 128 kbps, which costs about one percent of what the video is already using. Because Punktfunk sends very short audio frames to keep latency down, the old rate was leaving real quality on the table — most audibly on music. Update the host; every existing client already plays whatever arrives.
+- **The black screen now heals in seconds, mid-stream.** With 0.25.0 on both ends, a host that detects a constrained network path re-sizes the video packets of the session you are already in, a few seconds after diagnosing it — the picture simply appears, without you reconnecting. With a 0.25.0 host and an older client you still get the fix below: the session in progress stays black, but the next connection is sized correctly and works.
+- **Hosts you reach over a VPN show as online on the Steam Deck.** The panel's list merges what it finds on the network with the hosts you have saved and probes the saved ones directly, so a box that never advertises itself — over Tailscale, or on another subnet — reads as up instead of unreachable. Rows sort online first, then most recently streamed.
+- **Waking a sleeping host from the Deck waits for it properly.** The panel used to send the wake-up and then guess how long to wait before dialling. It now waits for the host to actually answer.
+- **Two new troubleshooting sections on audio.** One explains what the host actually captures and why streamed sound can be worse than what you hear on the host itself — naming the Steam Streaming Microphone trap explicitly and showing the log line that identifies it. The other covers audio that lags the picture, why it should now correct itself, and what to check when it does not.
+
+## Fixed
+
+- **A host on a network that carries smaller packets than usual no longer streams a permanent black screen.** Everything small got through — the connection, your input, your sound — while every single video packet was slightly too big for one hop and died silently. The result connected fine, reported zero packet loss on the client, showed every gauge green on the host, and displayed nothing at all, with nothing written to either log to say why. The host now measures what the path to each client can really carry, warns with the actual diagnosis when it cannot carry full-size video, and remembers the measurement so the next connection from that client is sized to fit. The usual cause, and the one the warning names, is a VPN or overlay network adapter claiming the route. Update the host — this works with every client already out there.
+- **Audio that falls behind the picture pulls itself back.** Every client kept a small buffer to absorb network jitter, and that buffer could only ever grow: one burst of Wi-Fi interference, one stutter on the host, or simply two devices' clocks running at fractionally different speeds pushed sound permanently behind the video, and the only cure was reconnecting. Android was worst, with no correction at all — it settled at its ceiling and stayed there for the whole session. All four clients now trim the buffer back a few milliseconds at a time under a crossfade, which is inaudible. Update the client; an older one keeps drifting no matter how new the host is.
+- **The host stopped pushing your whole desktop mix through Steam's voice channel.** On a PC with Steam installed, the host was capturing Steam's Streaming Microphone device because it is silent on the host — but that device exists to carry voice, and if Windows had it set to mono or below 48 kHz, the entire game mix was squeezed through it before encoding, where no amount of bitrate could bring it back. A silent device now has to prove it can carry full-quality sound before being preferred over real hardware, and if nothing better exists the host says so in the log. Update the host.
+- **Sound no longer cuts out over and over when something keeps changing your default playback device.** Some applications re-set the Windows default device every few seconds; each time, the host tore the whole capture down and rebuilt it, which is an audible dropout — one field log shows seven in sixteen seconds. The host now restores the default without dropping the stream, and if it happens repeatedly it stops fighting for a minute and says so once. Update the host.
+- **Audio the host dropped internally is no longer silently glued over.** When the encoder fell behind, captured sound was discarded with nothing recording it: you heard a click, and everything after it stayed permanently shifted. Those drops are now counted and warned about, so a quiet host, a broken device and a stream damaging itself no longer look identical. Update the host.
+- **Automatic bitrate stops sawtoothing when your device's decoder, not the network, is the limit.** The control loop has a mechanism for learning "this device cannot decode much past here, stop trying", and in practice it never once fired — one recording at 1440p120 shows it swinging between 220 and 450 Mb/s for nine solid minutes without ever learning the lesson. Three separate reasons it was unreachable are fixed, including one where a struggling decoder repeatedly asking for a fresh picture on an otherwise clean link was blamed on the network. Update the client.
+- **Rumble stops fading in and out on a Steam Deck.** The Deck's motors need a fresh instruction every 40 ms or the repeat is discarded, and renewals kept colliding with that, stretching the real gap between motor writes to two and a half times what it should be — so sustained rumble came through weak and uneven. Update the client.
+- **Rumble survives a hiccup instead of dying for the rest of the session.** On Android a single failure from the phone's vibration service silently killed the thread driving rumble, with nothing to notice or restart it, so rumble was gone until you restarted the app. And on every client, a stop instruction that never reached the controller used to be assumed to have worked — over USB there is no firmware timeout behind that, so a dropped stop left the motors running with nothing scheduled to try again. Update the client.
+- **Two DualSenses stop rumbling for each other.** With two connected to an iPhone, iPad, Mac or Apple TV, both could end up driving the same physical controller, so one player's rumble came out of the other player's pad and the two fought over it. Each now drives its own. Very light rumble also stopped vanishing on that path — anything under about half a percent was being rounded away to nothing. Update the client.
+- **A controller is handed back to you neutral when the stream ends.** Trigger resistance, lightbar colour and player lights live in the controller's own firmware, so they outlast the stream, the app, and even unplugging. Ending a session while a game held a weapon's trigger resistance left that trigger physically stiff on your desktop afterwards, with the lightbar still showing the game's last colour. Every client now releases both triggers, darkens the lightbar and clears the player lights on the way out — including on the exit paths that previously skipped it and left the pad buzzing after the stream was gone. Update the client.
+- **A dropped lightbar or trigger change repairs itself instead of sticking.** These were sent once, when they changed, over packets that can be lost — so one lost packet could strand a controller on the previous weapon's trigger effect, or the last scene's lightbar colour, potentially for the rest of the level. The host now re-sends the current state once a second to repair it. Update the host.
+- **A cut-off packet no longer cancels a trigger effect a game is holding.** A truncated adaptive-trigger packet decoded as an empty effect, and an empty effect is exactly what a controller reads as "let go" — so a weapon's resistance could silently vanish mid-fight. That shape is now rejected, while a genuine release still works.
+- **Player-number lights work on controllers that are not a DualSense.** Xbox pads, Switch Pro controllers and everything else with player lights ignored the host's player number completely, so nothing lit at all. Update the client.
+- **A centred stick reads as centred.** When the host presents your controller to games as a DualSense, DualSense Edge or DualShock 4, both sticks' vertical axes sat one step below true centre — a permanent, very slight downward pull, small enough to hide under most games' deadzones but plainly visible to any game reading the raw axis. Triggers on a controller presented as a Steam Deck pad also topped out just short of a full pull, so anything needing a genuine full press could never fire. Both are now exact. Update the host.
+- **Two virtual controllers stop corrupting each other's rumble on a Windows host.** When a game drove two pads hard enough for their updates to overlap, two rumble instructions could be written into the same slot and arrive as one garbled instruction, or one could be skipped outright — and a skipped *stop* is the one that hurts, leaving the pad buzzing until a safety timer noticed the game had gone quiet. Update the host.
+- **Delayed rumble effects fire at the right moment on a Linux host.** Games that schedule an effect to start after a short delay — routine for older Windows games running through Proton — had it start early and end early by the same amount, because the delay was read and then never applied. An effect still waiting its turn is also no longer cancelled by the idle safety-off before it has been felt. Update the host.
+- **A controller driver that failed to attach no longer stalls the stream while the host works out why.** The diagnosis ran a slow system lookup on the very thread feeding controller input and rumble — up to two seconds per affected pad, at exactly the moment a session was already going wrong. It now runs in the background, and because it is off the critical path it can afford to wait long enough to report what it actually found. Update the host.
+- **The Steam Deck keeps its trackpad mouse when a stream starts.** Starting a stream killed the built-in trackpad-as-mouse system-wide, and it only returned seconds later when the controller's own firmware watchdog restored it. Update the client.
+- **Controller settings you cannot use no longer look live.** With "Forward controllers" off, the rows beneath it have nothing to act on, but on the Windows app and both controller-driven settings screens they stayed fully interactive — so you could sit there changing settings that did nothing. They are now dimmed until forwarding is back on. On Apple devices, starting a stream with forwarding off also stopped claiming every button's system gesture (which took away your screenshot and Home presses) and stopped powering up the controller's motion sensors for a stream that was not forwarding anything. Update the client.
+- **A leftover folder from an uninstalled Sunshine or Apollo is no longer treated as a conflict.** Both uninstallers leave a settings folder behind, and Punktfunk counted any trace at all — a leftover folder, a file on disk, a registered but switched-off background service — as a live clash. Affected machines warned on every start and showed a red card in the web console reading that another streaming server was running, when nothing was. Only a server that is genuinely running, or set to start on its own, counts now; the console names exactly what it saw, and leftovers appear in the full report under a heading saying they need no action. Update the host.
+- **A crashed host gives you your screen back.** In Exclusive display mode the host switches your own monitors off for the length of a session and back on when it ends. If the host crashed or was killed mid-session that never happened — the desk simply stayed dark, no timeout brought it back, and the way out was Windows' own display shortcut or a reboot. The host now records which screens it is about to switch off before switching them off, and forces every connected display back on the next time it starts. Recovery happens at that next start, not on a timer: if the host stays down, the screen stays dark until it runs again. Update the host.
+- **Camera look survives pressing Escape on an iPad.** Pressing Escape mid-stream made iPadOS hand the pointer back to the system, and Punktfunk never took it back. Clicks kept landing exactly where you aimed, so input looked fine — but the game stopped receiving mouse movement, so camera look was dead for the rest of the session. Clicking back into the video now takes the pointer again, and if the system refuses the first time, the next click tries again. Update the client.
+- **"Open log folder" on Windows opens the log folder.** On installed builds it opened your Documents folder instead: the path the client handed to Explorer was correct to write to but did not exist as a real folder, and Explorer quietly fell back. The same wrong path appeared in the startup line naming the log file and in the message shown when a session fails to start. All three now point at the real folder. Update the client.
+
+## If you stream from a Steam Deck
+
+The Decky plugin has been rebuilt as a **launcher**. It no longer contains a second, separate streaming client; it is now a short list of your hosts plus one button into the Punktfunk app, which has the full controller-driven interface. This makes the plugin far smaller and means the Deck stops having two implementations of everything that could disagree with each other — but some things genuinely moved, and one was removed:
+
+- **Settings moved** to **Open Punktfunk → Settings**. Same rows, same saved values, still fully controller-navigable.
+- **Adding, renaming and forgetting hosts moved** to **Open Punktfunk → Add host**.
+- **Browsing a host's games moved** to **Open Punktfunk → Library**.
+- **Pinned Games has been removed, with nothing to migrate to yet.** The panel's one-tap "Stream *game*" rows are gone: pinning now works on a host and a settings profile rather than on a game. Your old pin file is deliberately left alone on disk so a later release can migrate it, but in 0.25.0 those rows do not appear.
+- **The Deck's Steam and `…` buttons now stay with the Deck.** One press used to open both menus at once, the Deck's own covering the stream, because SteamOS reacts to those buttons whatever the app does. Reach the host's menus with hold-Select, or the panel's new **Host menus** buttons. To restore the old behaviour, set **Open Punktfunk → Settings → Steam / guide button** to **Send to host**.
+- **The plugin needs the Punktfunk client on the Deck to be 0.22.0 or newer**, because it drives everything through the client. An older one is detected explicitly and the panel offers the update button that fixes it, rather than silently showing an empty list.
+
+## Under the hood (for developers)
+
+- **Wire protocol 2 — unchanged**, despite substantial growth, because every addition is optional or capability-gated. What grew without a bump: an optional trailing `max_shard_payload: u16` on `Hello` (absent/0 = legacy, and it doubles as both the renegotiation capability flag and the jumbo receive ceiling); two new control messages `ShardPayloadChanged` (`0x08`) and `ShardPayloadAck` (`0x09`); a redundant desktop-audio datagram tag `0xD2` alongside the plain `0xC9`; a controller-audio plane at `0xD1` (`[0xD1][u8 pad][u8 kind][u32 seq LE][u64 pts_ns LE][opus payload]`, which is why `0xD2` skipped that value); and `MAX_DATAGRAM_BYTES` 2048 → 9216.
+- **C ABI 14 → 16**, in two steps. **15** is unusual: no code changed and no symbol was added with it. It retroactively versions the shared rumble policy engine's C surface — `punktfunk_connection_next_rumble_cmd`, `punktfunk_connection_set_rumble_quirks` and the `PUNKTFUNK_RUMBLE_QUIRK_*` bits — which shipped while the constant still read 7 and never got one, so every core since has exported those symbols while advertising a version that did not promise them. That cannot be fixed retroactively, so 15 is declared as the **floor that guarantees** the surface: at or above 15 it is present, below it an embedder must probe for the symbol. **16** adds the controller-audio surface and mirrors its two capability bits into the C ABI.
+- **Breaking for C embedders: 149 unprefixed macros are now `PUNKTFUNK_`-prefixed** (139 `#define`s renamed in the checked-in header). Names as generic as `MAX_PADS`, `TAG_LEN`, `ABI_VERSION`, `WIRE_VERSION`, `INPUT_MAGIC` and the whole `BTN_*` / `AXIS_*` family were landing in the namespace of every program that included the header. Fixing it is mechanical — add the prefix, the values are identical — and there is **no silent breakage**: the old spellings cease to exist, so it is always an undeclared-identifier error, never a wrong value. That is precisely the failure it removes, since a colliding `#define` does not fail to compile; the preprocessor silently takes the last definition, so an embedder whose own header defined `MAX_PADS` previously got a wrong value at runtime. Associated constants are untouched — the generator already qualifies those with their type name. Scheduled for a release boundary deliberately; nothing in-tree used the old spellings but one Swift test, updated in the same commit.
+- **Four new capability bits, and the video-caps byte did not overflow.** In the handshake's client/host capability bytes: client `0x04` / host `0x20` for the redundant desktop-audio plane ("can decode it" / "is sending it"), and client `0x08` / host `0x40` for controller audio (`CLIENT_CAP_PAD_AUDIO` / `HOST_CAP_PAD_AUDIO`, mirrored into the C ABI as `PUNKTFUNK_*` and asserted equal to their wire twins). The video-caps byte still carries exactly the eight bits it carried at 0.24.0 — no ninth cap, no second byte, so nothing forced an ABI bump from that direction.
+- **Unchanged:** virtual-display driver protocol 6 (minimum accepted 3) and the Windows virtual-gamepad channel 3 — `crates/pf-driver-proto` is byte-for-byte identical to v0.24.0.
+- **Adaptive-trigger effects are now length-bounded** on both encode and decode against one shared constant, with the header emitting `uint8_t effect[PUNKTFUNK_HID_EFFECT_MAX]` in place of a literal `11` (same value, so the struct layout is byte-identical). A zero-length effect body is rejected rather than decoding as an empty — that is, a release — effect. Out-of-range pad indices are now dropped before either rumble consumer sees them; the reorder gate bounds-checked and the legacy queue did not, so an embedder draining it could be handed an index it would use to subscript its own array. The client also clamps the host's rumble lease receive-side at 5 s, where the existing ceiling was sender-side only.
+- **Windows pad drivers publish their sequence counters with release ordering** (the host was already loading with acquire and pairing with nothing) and serialize the output-ring publish. The `/dev/uhid` event ABI, previously transcribed verbatim into all five Linux gamepad backends, is consolidated into one module with tests on the two accessors that had already drifted.
+- **The controller-audio plane in detail.** `0xD1` carries one Opus frame per datagram behind a 15-byte header, with `PAD_AUDIO_KIND_HAPTICS = 0` (the pad's BACK channel pair — the voice coils — at 5 ms frames) and `PAD_AUDIO_KIND_SPEAKER = 1` (the FRONT pair, 10 ms). Best-effort like every audio plane: loss shows up as a sequence gap concealed by the gap tracker, and silence is a frozen sequence under the same mic-mute discipline, with the host gating at −60 dBFS on a 250 ms hangover. Alongside it, `HidOutput::AudioCtl` is a new `0xCD` kind `0x06` carrying the DualSense output report's volume/routing bytes, change-only and value-deduped — an older client drops it as an unknown kind. A client advertises per-pad intent through two new arrival flags (`1 << 8` haptics, `1 << 9` speaker), sent only toward a `HOST_CAP_PAD_AUDIO` host. **Capability-byte pressure is now worth watching:** `client_caps` has four bits free, but `host_caps` is down to its last one (`0x80`), and `video_caps` remains full from 0.23.0 — the standing "next video cap needs a second byte and an ABI bump" note still stands.
+- **The controller-audio host gate is Windows-only and Steam-dependent.** `host_cap()` returns false unconditionally off Windows, and on Windows it still requires provisioning to have published at least one endpoint, which requires Valve's driver. The client only advertises its capability if a setting would actually render something, so a user with both toggles off never causes the host to provision anything. Note a real inconsistency to reconcile: the desktop session client defaults `pad_speaker` to `"pad"` (on) while Android defaults its speaker toggle to off, and the desktop side exposes these as serde-defaulted JSON keys with no settings UI at all. `pad_speaker = "mix"` is a declared TODO that logs once and behaves as `off`. The GameStream/Moonlight path always reports no pad-audio capability.
+- **New host environment settings.** Controller audio: `PUNKTFUNK_PAD_AUDIO` (on unless set to `0`), `PUNKTFUNK_PAD_AUDIO_SLOTS` (default 1, max 4 — multi-pad needs an operator to raise it), and `PUNKTFUNK_PAD_AUDIO_STAMPS` (debug bisect hook), plus a `punktfunk-host pad-endpoint ensure|remove|status` devtest command. Audio: `PUNKTFUNK_AUDIO_QUALITY` (`low`/`standard`/`high`, default `high` = stereo 256 kbps; `standard` reproduces the pre-0.25 encoder exactly for an A/B, and a typo warns once rather than silently downgrading), `PUNKTFUNK_AUDIO_REDUNDANCY`, and `PUNKTFUNK_AUDIO_OUTPUT_MODE` (`client_only`/`host_and_client`/`follow_default`, default `client_only`, **Windows host only**). The legacy `PUNKTFUNK_HOST_AUDIO=1` and `PUNKTFUNK_KEEP_DEFAULT=1` still work, mapping to `host_and_client` and `follow_default`; `follow_default` wins if both are set. Wire: `PUNKTFUNK_WIRE_MTU` (pins on-wire IP MTU for all sessions; a value above 1500 also enables jumbo) and `PUNKTFUNK_JUMBO=1` (fixed 9000-MTU profile). All are documented on the troubleshooting page, not yet in the configuration reference.
+- **Mid-session shard renegotiation is gated off for PyroWave sessions**, which parse the video stream in windows fixed at session start — re-sizing mid-stream would corrupt the parse. Those sessions get the next-session clamp only, and are excluded from jumbo. The decode-cap latch fix likewise does not apply to PyroWave, where adaptive bitrate is open-loop by design.
+- **The Deck plugin's Python backend is now four thin shells over the `punktfunk` CLI** (`discover`, `hosts list --probe --json`, `pair`, `hosts add`); it parses no client data files and re-implements no client rules, and an outdated client reports itself deterministically as exit 5 + `unknown command "<verb>"` rather than being inferred from GTK startup noise. Host identity is matched by fingerprint first and address second in exactly one place, so a host that changed DHCP lease still matches its record while a different box inheriting the address does not inherit its pairing. `KnownHosts::read()` was split out of `load()` so `discover` can annotate against the store without minting-and-saving ids, which two parallel invocations could otherwise race.
+- **The hold-Select gesture is one state machine** with unit tests in the shared client core, re-implemented to the same rules in the Apple capture layer and Android's router. A tapped Select is delivered on release with its release scheduled 50 ms behind, because a back-to-back down+up can otherwise fold into a single sequenced snapshot and vanish. `punktfunk-session` gained a per-user Unix control socket (`$XDG_RUNTIME_DIR[/app/$FLATPAK_ID]/punktfunk-session-ctl.sock`) with two verbs, `guide` and `qam` — the one runtime path a flatpak and the outside-the-sandbox Decky backend see identically.
+- **Verification is build-level.** Clippy and test gates on Linux, the Windows runner and macOS; the desktop-audio, packet-sizing and iPad pointer work has not been confirmed on glass in these commits. **Controller audio in particular has never run on a real DualSense** — it is a hardware feature whose entire verification to date is unit tests and compile checks, and its rumble arbitration rests on an explicitly retracted assumption about whether the voice coils and the rumble motors are the same actuators (the evidence-based 500 ms idle window is correct either way, but the underlying exclusivity is unsettled). Android's arbiter is the evidence-based one; the desktop twin and the coil restore on Android's stop path are both still owed. Some Android OEM kernels also refuse the isochronous claim outright, which degrades to ordinary rumble and is reported by the self test.

docs/releases/whatsnew/v0.25.0.txt

@@ -0,0 +1,6 @@
+• New: a DualSense plugged in by USB can play the host's fine-grained haptics through the pad itself. Needs a Windows host with Steam installed.
+• Sound that drifts behind the picture now catches itself up instead of staying late all session.
+• Rumble no longer dies for the rest of the session after one glitch.
+• Automatic bitrate stops overshooting what your device can really decode.
+• Hold Select to reach the host's Guide menu.
+• Pin your settings profiles to hosts from the TV interface.

include/punktfunk_core.h

@@ -70,7 +70,13 @@
 // says what it says — so v15 is the floor that *guarantees* them: at or above it the surface is
 // present, below it an embedder must probe for the symbol. Purely a version statement; no code
 // changed with this bump, and no wire change, so [`WIRE_VERSION`] is unchanged.
-#define PUNKTFUNK_ABI_VERSION 15
+// v16: added the pad-audio client surface — `punktfunk_connection_next_pad_audio` (the 0xD1
+// per-gamepad DualSense haptics/speaker plane) + `punktfunk_connection_set_pad_audio_caps` and
+// the `PUNKTFUNK_CLIENT_CAP_PAD_AUDIO` / `PUNKTFUNK_HOST_CAP_PAD_AUDIO` mirrors. Additive and
+// capability-gated end to end: the wire grows a new datagram tag (0xD1) an old client never
+// receives (double-gated caps), a new 0xCD kind (0x06, dropped as unknown by old clients) and
+// arrival flag bits 8/9 sent only toward a capable host, so [`WIRE_VERSION`] is unchanged.
+#define PUNKTFUNK_ABI_VERSION 16
 
 // The punktfunk/1 **wire** version — what `Hello`/`Welcome` carry and hosts equality-check.
 // Deliberately its own constant: [`ABI_VERSION`] tracks the embeddable **C surface**
@@ -94,6 +100,13 @@
 // little-endian `u16`s with `effect_len = 6`. Clients without trackpad coils drop it.
 #define PUNKTFUNK_HIDOUT_TRACKPAD_HAPTIC 4
 
+// `PunktfunkHidOutput::kind` — the audio-control region of a DS5 output report (pad-audio
+// routing/volumes; the audio SAMPLES arrive via [`punktfunk_connection_next_pad_audio`]).
+// `which` = the condensed audio flags (bit0 = haptics-select, bits1..4 = the report's
+// audio-valid flags); `effect[0..6]` = bytes 5..=10 of the report verbatim
+// (headphone/speaker/mic volumes + routing) with `effect_len = 6`. Forwarded change-only.
+#define PUNKTFUNK_HIDOUT_AUDIO_CTL 5
+
 // Capacity of `PunktfunkHidOutput::effect` (the DualSense trigger parameter block).
 #define PUNKTFUNK_HID_EFFECT_MAX 11
 
@@ -278,6 +291,28 @@
 // design/pen-tablet-input.md.)
 #define PUNKTFUNK_HOST_CAP_PEN 16
 
+// Host-capability bit in [`punktfunk_connection_host_caps`]: the host can capture per-gamepad
+// audio (DualSense voice-coil haptics + speaker) and emit it on the 0xD1 plane toward pads
+// declared capable via [`punktfunk_connection_set_pad_audio_caps`]. Set only when the client
+// asked via [`PUNKTFUNK_CLIENT_CAP_PAD_AUDIO`]. (Mirrors `quic::HOST_CAP_PAD_AUDIO`.)
+#define PUNKTFUNK_HOST_CAP_PAD_AUDIO 64
+
+// Pad-audio `kind` ([`punktfunk_connection_next_pad_audio`]): the BACK channel pair — DualSense
+// voice-coil haptics, 5 ms Opus frames. (Mirrors `quic::PAD_AUDIO_KIND_HAPTICS`.)
+#define PUNKTFUNK_PAD_AUDIO_KIND_HAPTICS 0
+
+// Pad-audio `kind`: the FRONT channel pair — the controller's built-in speaker, 10 ms Opus
+// frames. (Mirrors `quic::PAD_AUDIO_KIND_SPEAKER`.)
+#define PUNKTFUNK_PAD_AUDIO_KIND_SPEAKER 1
+
+// [`punktfunk_connection_set_pad_audio_caps`] `audio_caps` bit: the pad renders the HAPTICS
+// stream (a real DualSense's voice coils).
+#define PUNKTFUNK_PAD_AUDIO_CAP_HAPTICS 1
+
+// [`punktfunk_connection_set_pad_audio_caps`] `audio_caps` bit: the pad renders the SPEAKER
+// stream.
+#define PUNKTFUNK_PAD_AUDIO_CAP_SPEAKER 2
+
 // [`punktfunk_connect_ex9`] `client_caps` bit: render the host cursor locally (the cursor
 // channel, `design/remote-desktop-sweep.md` M2).
 #define PUNKTFUNK_CLIENT_CAP_CURSOR 1
@@ -288,6 +323,13 @@
 // forward-compatible.
 #define PUNKTFUNK_CLIENT_CAP_PHASE_LOCK 2
 
+// [`punktfunk_connect_ex9`] `client_caps` bit: the client understands the pad-audio plane
+// (0xD1 — per-gamepad DualSense voice-coil haptics + speaker). The embedder MUST then drain
+// [`punktfunk_connection_next_pad_audio`] and declare each capable pad via
+// [`punktfunk_connection_set_pad_audio_caps`]; the host emits pad audio only when it answers
+// with [`PUNKTFUNK_HOST_CAP_PAD_AUDIO`]. (Mirrors `quic::CLIENT_CAP_PAD_AUDIO`.)
+#define PUNKTFUNK_CLIENT_CAP_PAD_AUDIO 8
+
 // `*ttl_ms` sentinel written by [`punktfunk_connection_next_rumble2`] for a legacy (v1) rumble
 // datagram — an old host that sent no self-termination lease. The client then falls back to its
 // own staleness heuristic for that update instead of a host-supplied deadline.
@@ -367,6 +409,19 @@
 // Fixed serialized size of an [`InputEvent`] on the wire (tag + fields).
 #define PUNKTFUNK_INPUT_WIRE_LEN (((((1 + 1) + 4) + 4) + 4) + 4)
 
+// [`InputKind::GamepadArrival`] `flags` bit: this pad renders pad-audio HAPTICS — it is (or
+// forwards to) a real DualSense whose voice-coil actuators can play the
+// [`PAD_AUDIO_KIND_HAPTICS`](crate::quic::PAD_AUDIO_KIND_HAPTICS) stream. Rides above the pad
+// index byte; sent only toward a [`HOST_CAP_PAD_AUDIO`](crate::quic::HOST_CAP_PAD_AUDIO) host
+// (an older host reads the whole `flags` word as the index, so unexpected high bits would make
+// it drop the declaration).
+#define ARRIVAL_FLAG_PAD_AUDIO_HAPTICS (1 << 8)
+
+// [`InputKind::GamepadArrival`] `flags` bit: this pad renders pad-audio SPEAKER — the
+// [`PAD_AUDIO_KIND_SPEAKER`](crate::quic::PAD_AUDIO_KIND_SPEAKER) stream. Same wire discipline
+// as [`ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`].
+#define ARRIVAL_FLAG_PAD_AUDIO_SPEAKER (1 << 9)
+
 // The number of gamepads addressable on the wire (`flags` pad index 0..15). Shared by the
 // client's snapshot fold and the host's per-pad accumulators.
 #define PUNKTFUNK_MAX_PADS 16
@@ -675,6 +730,18 @@
 #define PUNKTFUNK_CLIENT_CAP_AUDIO_RED 4
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`Hello::client_caps`] bit: the client understands the pad-audio plane
+// ([`PAD_AUDIO_MAGIC`](super::datagram::PAD_AUDIO_MAGIC), `0xD1`) — per-gamepad DualSense
+// voice-coil haptics + speaker Opus frames, plus the [`HidOutput::AudioCtl`]
+// (super::datagram::HidOutput) routing/volume events. Active only when the host answers with
+// [`HOST_CAP_PAD_AUDIO`] AND the pad's arrival declared a renderer for the kind
+// ([`crate::input::ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`]/`_SPEAKER`) — the capable-and-agreed
+// precedent, per pad; toward an older or incapable host nothing changes. `0x08` — `0x01` is [`CLIENT_CAP_CURSOR`],
+// `0x02` is [`CLIENT_CAP_PHASE_LOCK`], `0x04` is [`CLIENT_CAP_AUDIO_RED`].
+#define CLIENT_CAP_PAD_AUDIO 8
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // [`Welcome::host_caps`] bit: the host CAN forward the cursor out-of-band (it captures cursor
 // metadata separately from the frame — the Linux portal `SPA_META_Cursor` path; NOT gamescope,
@@ -714,6 +781,19 @@
 #define PUNKTFUNK_HOST_CAP_AUDIO_RED 32
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`Welcome::host_caps`] bit: the host can capture pad audio — its virtual DualSense exposes
+// the pad's audio endpoints (voice-coil haptics + speaker), so a game's per-pad audio can be
+// captured and shipped on the [`PAD_AUDIO_MAGIC`](super::datagram::PAD_AUDIO_MAGIC) plane.
+// Set only when the client asked via [`CLIENT_CAP_PAD_AUDIO`]; when both bits agree, a
+// capable client marks its pads' render capabilities on their arrivals
+// ([`crate::input::ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`]/`_SPEAKER`) and the host emits `0xD1`
+// toward exactly those pads. `0x40` — `0x20` is [`HOST_CAP_AUDIO_RED`], `0x10` is
+// [`HOST_CAP_PEN`], `0x08` is [`HOST_CAP_CURSOR`], `0x04` is [`HOST_CAP_TEXT_INPUT`],
+// `0x01`/`0x02` are gamepad-state / clipboard.
+#define HOST_CAP_PAD_AUDIO 64
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // [`Hello::video_codecs`] bit: the client can decode H.264 / AVC. The GPU-less **software**
 // encode path (openh264) emits H.264, so a client that wants to stream from a software host MUST
@@ -1011,7 +1091,9 @@
 // audio = [`AUDIO_MAGIC`] (0xC9, host→client), rumble = [`RUMBLE_MAGIC`] (0xCA, host→client),
 // mic = [`MIC_MAGIC`] (0xCB, client→host), rich-input = [`RICH_INPUT_MAGIC`] (0xCC, client→host),
 // HID-output = [`HIDOUT_MAGIC`] (0xCD, host→client), HDR metadata = [`HDR_META_MAGIC`]
-// (0xCE, host→client).
+// (0xCE, host→client), host timing = [`HOST_TIMING_MAGIC`] (0xCF, host→client), cursor state =
+// [`CURSOR_STATE_MAGIC`] (0xD0, host→client), pad audio = [`PAD_AUDIO_MAGIC`] (0xD1,
+// host→client).
 #define PUNKTFUNK_AUDIO_MAGIC 201
 #endif
 
@@ -1162,6 +1244,31 @@
 #define PUNKTFUNK_CURSOR_RELATIVE_HINT 2
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Pad-audio datagram tag, host → client: per-gamepad audio a game routed
+// to the host's virtual DualSense — voice-coil haptics and the built-in speaker — for the client
+// to render on the matching real controller. Next tag after [`CURSOR_STATE_MAGIC`]. The
+// per-pad AUDIO plane (Opus frames, the [`AUDIO_MAGIC`]/[`MIC_MAGIC`] shape plus pad + kind);
+// the routing/volume CONTROL side rides [`HidOutput::AudioCtl`]. Emitted only when the session
+// negotiated it ([`CLIENT_CAP_PAD_AUDIO`](super::caps::CLIENT_CAP_PAD_AUDIO) ∧
+// [`HOST_CAP_PAD_AUDIO`](super::caps::HOST_CAP_PAD_AUDIO)) and the pad's arrival declared a
+// renderer for the kind ([`crate::input::ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`]/`_SPEAKER`).
+// Best-effort like every audio datagram: a lost frame is a concealed gap, never state.
+#define PAD_AUDIO_MAGIC 209
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`PadAudioFrame::kind`]: the BACK channel pair — the DualSense voice-coil actuators (audio
+// haptics). 5 ms Opus frames, matching the [`AUDIO_MAGIC`] cadence: haptics are felt latency.
+#define PAD_AUDIO_KIND_HAPTICS 0
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`PadAudioFrame::kind`]: the FRONT channel pair — the controller's built-in speaker. 10 ms
+// Opus frames (speaker content tolerates the extra buffering for the better coding efficiency).
+#define PAD_AUDIO_KIND_SPEAKER 1
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // QUIC application error code a punktfunk/1 client closes the control connection with on a
 // **deliberate quit** (a user "stop", not a network drop). The host reads it off the connection's
@@ -1476,7 +1583,11 @@ enum PunktfunkInputKind
     PUNKTFUNK_INPUT_KIND_GAMEPAD_REMOVE = 13,
     // Declares which controller KIND a pad presents so a session can MIX types (pad 0 a
     // DualSense, pad 1 an Xbox pad). `code` = the [`GamepadPref`](crate::config::GamepadPref)
-    // wire byte, `flags` = pad index. Sent when the client opens a pad slot — before that pad's
+    // wire byte, `flags` = pad index in the low byte plus the pad's render capabilities in bits
+    // 8/9 ([`ARRIVAL_FLAG_PAD_AUDIO_HAPTICS`]/[`ARRIVAL_FLAG_PAD_AUDIO_SPEAKER`] — sent only
+    // toward a [`HOST_CAP_PAD_AUDIO`](crate::quic::HOST_CAP_PAD_AUDIO) host, so an older host
+    // keeps reading the whole word as the index; hosts decode via [`decode_gamepad_arrival`]).
+    // Sent when the client opens a pad slot — before that pad's
     // first input — and re-sent a few times against datagram loss (like [`GamepadRemove`]). The
     // host resolves the kind to a buildable backend and routes that pad's virtual device to it; a
     // pad the client never declares (an older client, or a fully-lost declaration) falls back to
@@ -2404,6 +2515,50 @@ PunktfunkStatus punktfunk_connection_next_audio_pcm(PunktfunkConnection *c,
                                                     uint32_t timeout_ms);
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Pull the next pad-audio frame (0xD1) — one Opus frame of DualSense voice-coil haptics
+// (`kind` = [`PUNKTFUNK_PAD_AUDIO_KIND_HAPTICS`], 5 ms) or built-in-speaker audio
+// ([`PUNKTFUNK_PAD_AUDIO_KIND_SPEAKER`], 10 ms) for gamepad `*out_pad` — waiting up to
+// `timeout_ms`. The payload is COPIED into `buf` (no borrow-until-next-call slot); the return
+// value is its length in bytes, `0` = nothing this poll (timeout — or a DTX/oversized frame,
+// both of which an embedder treats the same way), `-1` = the session ended (or an invalid
+// handle/buffer). All pads/kinds share one queue — fan out by `*out_pad`/`*out_kind` to
+// per-actuator Opus decoders. A frame larger than `buf_len` is dropped like the timeout case
+// (the plane is lossy by design; any real Opus frame fits a 1500-byte buffer). Only a session
+// connected with [`PUNKTFUNK_CLIENT_CAP_PAD_AUDIO`] against a
+// [`PUNKTFUNK_HOST_CAP_PAD_AUDIO`] host — with the pad declared via
+// [`punktfunk_connection_set_pad_audio_caps`] — ever receives any. Drain from a dedicated
+// thread (one puller, may run alongside the other planes' pullers).
+//
+// # Safety
+// `c` is a valid connection handle; the `out_*` pointers are writable (NULLs are skipped);
+// `buf` is writable for `buf_len` bytes.
+int32_t punktfunk_connection_next_pad_audio(PunktfunkConnection *c,
+                                            uint8_t *out_pad,
+                                            uint8_t *out_kind,
+                                            uint32_t *out_seq,
+                                            uint64_t *out_pts_ns,
+                                            uint8_t *buf,
+                                            uintptr_t buf_len,
+                                            uint32_t timeout_ms);
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Declare wire pad `pad`'s pad-audio render capabilities (`audio_caps`: OR of
+// [`PUNKTFUNK_PAD_AUDIO_CAP_HAPTICS`] / [`PUNKTFUNK_PAD_AUDIO_CAP_SPEAKER`]) — how a client
+// tells the host WHICH pads can actually play the 0xD1 streams. Call at controller attach,
+// BEFORE the pad's arrival event is sent (the [`punktfunk_connection_set_rumble_quirks`]
+// timing): the core folds the bits into the arrival's flags (bits 8/9), and only toward a
+// [`PUNKTFUNK_HOST_CAP_PAD_AUDIO`] host — never calling this leaves the wire bytes exactly as
+// before. Latest-wins per pad; unknown bits are masked off.
+//
+// # Safety
+// `c` is a valid connection handle. Callable from any thread.
+PunktfunkStatus punktfunk_connection_set_pad_audio_caps(PunktfunkConnection *c,
+                                                        uint8_t pad,
+                                                        uint8_t audio_caps);
+#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.