fix(host/pads): a centred stick reads centred, and a delayed effect waits its turn

Four encoder faults, plus a note on a fifth that turned out not to be one.

A centred stick did not encode as centre on the Y axes. The mapper inverted the
already-quantised byte, and 0..255 has no exact midpoint: the forward map puts
centre at 0x80, so mirroring the output lands on 0x7F — one below the 0x80 that
DsState::neutral and the pad's own resting report use. Games idle-poll a
centred stick constantly, so a DualSense, Edge or DS4 sat under a permanent
sub-deadzone tilt. Inverting in i16 space instead maps centre to centre by
construction and keeps both extremes exact; the only cost is i16::MIN and
-32767 sharing a code, one LSB at the very end of the travel.

Force-feedback ignored replay.delay. It was decoded on upload and never read:
an effect started the moment it was played and ended replay.length later, so
anything scheduling a delayed effect — DirectInput under Wine does this
routinely — fired early AND finished early by the same amount. The delay now
shifts the whole window, with length measured from the end of the delay rather
than eaten into by it. Writing the test for that surfaced a second bug in the
same path: a waiting effect was still a candidate for the abandoned-effect
force-off, and since the play command is itself the last FF activity, an
infinite effect with a delay longer than the idle window would be killed on its
first contributing tick after sitting silent the whole time it waited. Being
abandoned now requires the effect to have been audible for the window too.

An empty serial panicked the service thread. The reply builder clamped the
length to at least 1 and then sliced that many bytes out of the string, which
asks a zero-byte slice for one byte. The kernel already has a graceful answer
for a length it rejects, so report the true one and let it fall back.

Deck triggers could not reach full pull. Scaling by 128 tops out at 32640 of a
declared 32767, leaving the last 127 counts unreachable, so no game could ever
see the axis bottom out. One multiply gets both ends exact.

The idle watchdog is left alone. It does cut finite multi-second effects that
the uinput path exempts, but only the uinput path is handed an explicit
duration; the protocols behind the watchdog are level-triggered with no
duration field anywhere in a report, so there is nothing at that layer to
exempt. The choice is between cutting a long effect and letting an abandoned
one drone forever, and only the latter has field evidence behind it. Recorded
at the constant so the next reader sees the cost rather than rediscovering it.

clients/android/native/src/audio.rs

@@ -12,10 +12,14 @@
 //! realtime callback and makes us own the buffer. So this client diverges deliberately to stop the
 //! Android-only crackle: (1) the callback is allocation/free-free — decoded buffers are recycled to
 //! the producer via a free-list instead of being freed on the audio thread (Android's Scudo `free`
-//! has unbounded tail latency); (2) the jitter ring is deeper (~40 ms prime / ~150 ms hard cap) and
-//! decoupled from the tiny LowLatency burst size, with de-prime hysteresis so a transient drain
-//! doesn't manufacture a silence; (3) the AAudio HW buffer is primed above its 2-burst default and
-//! grown on XRuns (Google's anti-glitch technique).
+//! has unbounded tail latency); (2) the jitter ring is deeper than the other clients' and decoupled
+//! from the tiny LowLatency burst size, with de-prime hysteresis so a transient drain doesn't
+//! manufacture a silence; (3) the AAudio HW buffer is primed above its 2-burst default and grown on
+//! XRuns (Google's anti-glitch technique).
+//!
+//! (2) is now the SHARED `punktfunk_core::audio::JitterPolicy` at `JitterTuning::AAUDIO`, which also
+//! fixed what this ring was missing: it had a hard cap but nothing that walked the depth back down,
+//! so drift and arrival bursts raised latency permanently and Android settled on its ceiling.
 
 use ndk::audio::{
     AudioCallbackResult, AudioContentType, AudioDirection, AudioFormat, AudioPerformanceMode,
@@ -34,26 +38,18 @@ const SAMPLE_RATE: i32 = 48_000;
 /// Decoded-chunk hand-off depth: 64 × 5 ms = 320 ms slack (matches the core's AUDIO_QUEUE).
 const RING_CHUNKS: usize = 64;
 
-// --- Jitter-ring depths, in MILLISECONDS (scaled to interleaved-f32 samples at runtime). --------
-// The channel count is negotiated, not a compile-time const, so these are kept in ms and multiplied
-// by `ms` (interleaved-f32 samples per millisecond at the resolved layout) inside `start`.
-// Unlike the Linux client (PipeWire adaptively rate-matches the stream to the graph clock, masking
-// host↔DAC drift + a shallow ring), AAudio hands us a raw callback and we own the buffer: drift and
-// WiFi power-save bunching land as underruns/overflows = crackle. So Android runs a deliberately
-// deeper, smoothly-managed ring than Linux — keep the two clients' depths intentionally divergent.
-/// Prime/target floor: fill to ~40 ms before playing (and after a sustained drain). Deep enough to
-/// ride out WiFi arrival jitter + clock drift; the dominant Android-only anti-crackle lever.
-const PRIME_FLOOR_MS: usize = 40;
-/// Ceiling for the burst-scaled target (so a large quantum can't push the prime depth too high).
-const PRIME_CEIL_MS: usize = 80;
-/// Drop-oldest headroom above the target before trimming — a ~80 ms band swallows an arrival burst
-/// without overflowing.
-const JITTER_HEADROOM_MS: usize = 80;
-/// Hard latency bound: never let the ring exceed ~150 ms (the only thing that caps added latency).
-const HARD_CAP_MS: usize = 150;
-/// Re-prime (go silent to refill) only after this many CONSECUTIVE empty callbacks, so one transient
-/// drain doesn't manufacture a fresh 40 ms silence (the old `if ring.is_empty()` re-primed instantly).
-const DEPRIME_AFTER_CALLBACKS: u32 = 5;
+// --- Jitter-ring depths now come from the SHARED policy (`punktfunk_core::audio::JitterTuning`). --
+// They used to be four Android-only constants here. The rationale for Android being DEEPER than the
+// other clients still holds and is preserved in `JitterTuning::AAUDIO`: unlike PipeWire, which
+// adaptively rate-matches the stream to the graph clock and masks host↔DAC drift, AAudio hands us a
+// raw callback and we own the buffer, so drift and Wi-Fi power-save bunching land as
+// underruns/overflows = crackle.
+//
+// Two things changed with the move. The prime floor drops 40 ms → 25 ms, because the policy GROWS
+// the target on the devices that actually underrun instead of every device pre-paying for the worst
+// one. And the ring finally sheds: it had a hard cap but nothing that walked the depth back down, so
+// any drift or burst raised latency permanently and Android converged on its 120 ms ceiling and
+// stayed there — the "audio latency is too high" report.
 /// Throttle the AAudio XRun-driven HW-buffer grow check (cheap, but no need to poll every quantum).
 const XRUN_CHECK_EVERY: u32 = 128;
 
@@ -104,6 +100,7 @@ struct Counters {
     pcm_written: AtomicU64,  // PCM frames copied out to AAudio (device clock is pulling)
     underruns: AtomicU64,    // callbacks that emitted silence (ring not primed / drained)
     ring_depth: AtomicU64,   // ring sample count at the last callback
+    target_ms: AtomicU64,    // the policy's LIVE target depth (it grows on this device's underruns)
 }
 
 /// Owned by [`crate::session::SessionHandle`]: the live AAudio stream + the decode thread.
@@ -126,10 +123,9 @@ impl AudioPlayback {
         // Interleaved f32 samples per millisecond at this layout (48 kHz × channels); the ms-
         // denominated jitter-ring depths scale by it.
         let ms = (SAMPLE_RATE as usize / 1000) * channels;
-        let prime_floor = PRIME_FLOOR_MS * ms;
-        let prime_ceil = PRIME_CEIL_MS * ms;
-        let jitter_headroom = JITTER_HEADROOM_MS * ms;
-        let hard_cap_max = HARD_CAP_MS * ms;
+        let tuning = punktfunk_core::audio::JitterTuning::AAUDIO;
+        // Worst transient the ring can hold before the policy trims it.
+        let hard_cap_max = tuning.hard_cap_ms as usize * ms;
         let counters = Arc::new(Counters::default());
 
         // One open attempt at a given sharing mode. Everything the realtime callback captures
@@ -157,8 +153,10 @@ impl AudioPlayback {
             // `decode_loop`.
             let mut ring: VecDeque<f32> =
                 VecDeque::with_capacity(hard_cap_max + RING_CHUNKS * 5 * ms);
-            let mut primed = false;
-            let mut empties: u32 = 0; // consecutive empty callbacks (de-prime hysteresis)
+            // Shared de-jitter policy — prime depth, drift correction, de-prime hysteresis. The
+            // hysteresis this replaces was Android-only; Linux and Windows carried the instant
+            // `if ring.is_empty()` re-prime until now.
+            let mut policy = punktfunk_core::audio::JitterPolicy::new(tuning, channels as u8);
             let mut cb_count: u32 = 0; // callbacks since open (throttles the XRun grow check)
             let mut last_xrun: i32 = 0; // last AAudio XRun count we grew the buffer for
             let callback = move |s: &AudioStream, data: *mut c_void, num_frames: i32| {
@@ -173,21 +171,25 @@ impl AudioPlayback {
                     ring.extend(chunk.drain(..));
                     let _ = free_tx.try_send(chunk);
                 }
-                // Jitter buffer: prime to ~40 ms (prime_floor) before playing and after a sustained
-                // drain; drop-oldest only above a wide ~120 ms band. Decoupled from the AAudio burst
-                // `want` (tiny on the LowLatency MMAP path) so the depth doesn't collapse to a single
-                // quantum.
-                let target = (3 * want).clamp(prime_floor, prime_ceil);
-                let hard_cap = (target + jitter_headroom).min(hard_cap_max);
-                while ring.len() > hard_cap {
-                    ring.pop_front();
+                // Jitter buffer: the shared policy decides prime/silence, trims a burst, and —
+                // new here — sheds ONE crossfaded 5 ms frame when the depth average has sat above
+                // target long enough to be drift rather than jitter. Without that shed this ring
+                // had no way back down: it clamped at 120 ms and stayed pinned there.
+                let step = policy.step(ring.len(), want);
+                if step.drop_front > 0 {
+                    punktfunk_core::audio::crossfade_drop(
+                        &mut ring,
+                        step.drop_front,
+                        step.crossfade,
+                    );
                 }
-                if !primed && ring.len() >= target {
-                    primed = true;
-                }
-                if primed {
+                let mut ran_short = false;
+                if !step.silence {
                     for slot in out.iter_mut() {
-                        *slot = ring.pop_front().unwrap_or(0.0);
+                        *slot = ring.pop_front().unwrap_or_else(|| {
+                            ran_short = true;
+                            0.0
+                        });
                     }
                     cb_counters
                         .pcm_written
@@ -196,20 +198,15 @@ impl AudioPlayback {
                     out.fill(0.0);
                     cb_counters.underruns.fetch_add(1, Ordering::Relaxed);
                 }
-                // Re-prime only after a RUN of empty callbacks, not a single transient one —
-                // otherwise every momentary drain costs a fresh 40 ms silence (the old behaviour,
-                // self-inflicted crackle on any jitter spike).
-                if ring.is_empty() {
-                    empties += 1;
-                    if empties >= DEPRIME_AFTER_CALLBACKS {
-                        primed = false;
-                    }
-                } else {
-                    empties = 0;
-                }
+                // No-op while un-primed, so a deliberate priming silence is never counted as an
+                // underrun (which would otherwise drive the adaptive floor up for no reason).
+                policy.note_read(ran_short);
                 cb_counters
                     .ring_depth
                     .store(ring.len() as u64, Ordering::Relaxed);
+                cb_counters
+                    .target_ms
+                    .store(policy.target_ms() as u64, Ordering::Relaxed);
                 // Google's AAudio anti-glitch technique: when the device reports new XRuns, grow the
                 // HW buffer by one burst (up to capacity). getXRunCount + setBufferSizeInFrames are
                 // both callback-safe / non-blocking, and set clamps to capacity so it self-limits.
@@ -408,10 +405,11 @@ fn decode_loop(
                         }
                         if count % 600 == 0 {
                             log::info!(
-                                "audio: opus={count} pcm_frames={} underruns={} ring={} peak={window_peak:.3}",
+                                "audio: opus={count} pcm_frames={} underruns={} buffer_ms={} target_ms={} peak={window_peak:.3}",
                                 counters.pcm_written.load(Ordering::Relaxed),
                                 counters.underruns.load(Ordering::Relaxed),
-                                counters.ring_depth.load(Ordering::Relaxed),
+                                counters.ring_depth.load(Ordering::Relaxed) / ms.max(1) as u64,
+                                counters.target_ms.load(Ordering::Relaxed),
                             );
                             window_peak = 0.0;
                         }

clients/apple/Sources/PunktfunkKit/Audio/AudioRing.swift

@@ -3,28 +3,66 @@ import os
 
 /// SPSC-ish jitter ring (interleaved float, `channels` per frame), drain thread → render
 /// callback. The unfair lock is held for microseconds; fine at render-callback rates. Priming:
-/// reads return silence until enough is buffered (at least `prefill`, and at least one
+/// reads return silence until enough is buffered (at least the target, and at least one
 /// packet more than the device's render quantum — large-buffer devices would otherwise
-/// chronically out-demand the prefill and oscillate prime → dropout → re-prime), and an
-/// underrun re-primes, concealing jitter as one short dip instead of sustained crackle.
+/// chronically out-demand the prefill and oscillate prime → dropout → re-prime).
 /// All counts stay whole frames (multiples of `channels`), so the interleave can never slip.
+///
+/// **Drift correction.** Both ends run at 48 kHz but on different crystals, so backlog from a
+/// network stall or plain host-vs-DAC skew never drains on its own: without correction one 300 ms
+/// hiccup leaves audio 300 ms behind video for the rest of the session. This used to be handled by
+/// a `highWater` shed that dropped a whole `2 × prefill` at once — its own comment called that "one
+/// audible blip". It is now the same two-stage scheme the Rust clients share
+/// (`punktfunk_core::audio::JitterPolicy`): a slow depth average that sits above target for a
+/// sustained window sheds ONE 5 ms frame with a crossfade, and the hard cap is only a backstop.
+/// Keep the constants here in step with `JitterTuning.COREAUDIO`.
 final class AudioRing: @unchecked Sendable {
+    /// Mirrors `JitterTuning::COREAUDIO` — see that type for the rationale.
+    private static let targetMS = 20
+    private static let headroomMS = 30
+    private static let hardCapMS = 90
+    private static let deprimeAfter = 4
+    /// The protocol's frame: the shed unit, and the slack added over a large device quantum.
+    private static let frameMS = 5
+    /// Depth average must exceed target by this before drift correction fires — the middle of the
+    /// headroom band, so the smooth shed always gets its chance BEFORE the hard cap trims.
+    private static let shedExcessMS = 15
+    /// …and must stay there for this much consumed audio. Long, because a shed is the only thing
+    /// here a listener could notice; it must never fire on a transient.
+    private static let shedSustainMS = 2_000
+    private static let crossfadeMS = 2
+    /// Time constant of the depth average.
+    private static let ewmaTauMS = 1_000
+
     private var buf: [Float]
     private var readIdx = 0
     private var writeIdx = 0
     private var primed = false
     private var renderQuantum = 0
-    private let prefill: Int
-    private let highWater: Int
+    private var emptyReads = 0
+    private var depthAvg: Double = 0
+    private var overRun = 0
+    /// Reported, not acted on: short reads that actually starved the callback, and smooth drift
+    /// corrections. A rising underrun count means the ring is being starved (network or CPU),
+    /// which is a different problem from the depth being wrong.
+    private var underrunCount = 0
+    private var shedCount = 0
     private let channels: Int
+    private let perMS: Int
     private let lock = OSAllocatedUnfairLock()
 
-    /// `capacity`/`prefill` in samples (interleaved — `channels` per frame, both whole frames).
-    init(capacity: Int, prefill: Int, channels: Int) {
+    /// `capacity` in samples (interleaved — `channels` per frame, a whole number of frames).
+    /// The de-jitter depth is the ring's own business (`targetMS`), not a caller's prefill.
+    init(capacity: Int, channels: Int) {
         buf = [Float](repeating: 0, count: capacity)
-        self.prefill = prefill
         self.channels = channels
-        highWater = prefill * 4
+        perMS = 48 * channels
+    }
+
+    /// Live target depth in interleaved samples, lifted so it can always serve one device quantum
+    /// plus a packet (a large-buffer device cannot sustain a target below its own quantum).
+    private var target: Int {
+        max(Self.targetMS * perMS, renderQuantum + Self.frameMS * perMS)
     }
 
     func write(_ samples: UnsafePointer<Float>, count: Int) {
@@ -42,12 +80,12 @@ final class AudioRing: @unchecked Sendable {
             buf[(writeIdx + i) % capacity] = samples[i]
         }
         writeIdx += count
-        // Latency clamp: both ends run at 48 kHz, so backlog from a network stall (or
-        // creeping host-vs-DAC clock skew) never drains on its own — without this, one
-        // 300 ms hiccup leaves audio 300 ms behind video for the rest of the session.
-        // Shedding down to 2× prefill costs one audible blip instead.
-        if writeIdx - readIdx > highWater {
-            readIdx = writeIdx - prefill * 2
+        // Backstop only: the smooth shed in `read` is what normally holds the depth down.
+        let cap = min(target + Self.headroomMS * perMS, Self.hardCapMS * perMS)
+        if writeIdx - readIdx > cap {
+            readIdx = writeIdx - cap
+            depthAvg = Double(cap)
+            overRun = 0
         }
     }
 
@@ -57,16 +95,37 @@ final class AudioRing: @unchecked Sendable {
         defer { lock.unlock() }
         renderQuantum = max(renderQuantum, count)
         let available = writeIdx - readIdx
+
+        // Depth average, weighted by the callback size so its time constant is independent of the
+        // device quantum.
+        let alpha = min(1.0, Double(count) / Double(Self.ewmaTauMS * perMS))
+        depthAvg += (Double(available) - depthAvg) * alpha
+
         if !primed {
-            // One 5 ms host packet (240 frames × channels) of slack beyond the device's demand.
-            if available >= max(prefill, renderQuantum + 240 * channels) {
+            if available >= target {
                 primed = true
+                emptyReads = 0
             } else {
                 for i in 0..<count { out[i] = 0 }
                 return
             }
         }
-        let n = min(available, count)
+
+        // Drift correction: shed exactly one frame, crossfaded, once the AVERAGE has sat above
+        // the threshold for the sustain window. Anything shorter is jitter and must be left alone.
+        if depthAvg > Double(target + Self.shedExcessMS * perMS) {
+            overRun += count
+            if overRun >= Self.shedSustainMS * perMS {
+                overRun = 0
+                shedOneFrame()
+                shedCount += 1
+                depthAvg = Double(writeIdx - readIdx)
+            }
+        } else {
+            overRun = 0
+        }
+
+        let n = min(writeIdx - readIdx, count)
         let capacity = buf.count
         for i in 0..<n {
             out[i] = buf[(readIdx + i) % capacity]
@@ -74,9 +133,63 @@ final class AudioRing: @unchecked Sendable {
         readIdx += n
         if n < count {
             for i in n..<count { out[i] = 0 }
-            primed = false // underrun — re-prime before resuming
+            // De-prime only after a RUN of short reads: a single transient drain must not
+            // manufacture a whole target's worth of fresh silence.
+            emptyReads += 1
+            underrunCount += 1
+            if emptyReads >= Self.deprimeAfter { primed = false }
+        } else {
+            emptyReads = 0
         }
     }
+
+    /// Drop one protocol frame from the front, linearly crossfading the seam so the correction is
+    /// inaudible rather than a click. Mirrors `punktfunk_core::audio::crossfade_drop`; caller holds
+    /// the lock.
+    private func shedOneFrame() {
+        let drop = Self.frameMS * perMS
+        let available = writeIdx - readIdx
+        guard available > drop else { return }
+        let fade = min(Self.crossfadeMS * perMS, min(drop, available - drop))
+        let capacity = buf.count
+        if fade > 0 {
+            // The tail of what we discard fades out into the head of what survives.
+            for i in 0..<fade {
+                let old = buf[(readIdx + drop - fade + i) % capacity]
+                let new = buf[(readIdx + drop + i) % capacity]
+                let t = Float(i + 1) / Float(fade + 1)
+                buf[(readIdx + drop + i) % capacity] = old * (1 - t) + new * t
+            }
+        }
+        readIdx += drop
+    }
+
+    /// Current buffered depth in milliseconds — for the stats overlay and the drain thread's
+    /// periodic log.
+    var bufferedMS: Int {
+        lock.lock()
+        defer { lock.unlock() }
+        return (writeIdx - readIdx) / max(perMS, 1)
+    }
+
+    /// One consistent snapshot of the ring's vitals, taken under a single lock so the numbers in
+    /// a log line describe the same instant. Mirrors what the three Rust clients report.
+    struct Stats {
+        let bufferedMS: Int
+        let targetMS: Int
+        let underruns: Int
+        let sheds: Int
+    }
+
+    var stats: Stats {
+        lock.lock()
+        defer { lock.unlock() }
+        return Stats(
+            bufferedMS: (writeIdx - readIdx) / max(perMS, 1),
+            targetMS: target / max(perMS, 1),
+            underruns: underrunCount,
+            sheds: shedCount)
+    }
 }
 
 /// CoreAudio channel layout for the canonical wire order FL FR FC LFE RL RR [SL SR]. nil for

clients/apple/Sources/PunktfunkKit/Audio/SessionAudio.swift

@@ -317,10 +317,10 @@ public final class SessionAudio {
         // Build the playback layout from the host-RESOLVED channel count (never the request):
         // 2 = stereo / 6 = 5.1 / 8 = 7.1, canonical wire order FL FR FC LFE RL RR SL SR.
         let channels = Int(connection.resolvedAudioChannels)
-        // 1 s interleaved capacity, ~20 ms prefill (four 5 ms host packets of jitter absorption
-        // before the first sample plays), both scaled by the channel count.
-        let ring = self.ring ?? AudioRing(
-            capacity: 48_000 * channels, prefill: 960 * channels, channels: channels)
+        // 1 s interleaved capacity, scaled by the channel count. The de-jitter depth itself is
+        // the ring's own business now (`AudioRing.targetMS`, mirroring `JitterTuning::COREAUDIO`)
+        // rather than a prefill passed in here.
+        let ring = self.ring ?? AudioRing(capacity: 48_000 * channels, channels: channels)
         self.ring = ring
 
         // Engine-native deinterleaved float; the render block deinterleaves from the ring. Surround
@@ -403,6 +403,7 @@ public final class SessionAudio {
         stateLock.unlock()
         let thread = Thread { [connection, flag, drainDone] in
             defer { drainDone.signal() }
+            var drained = 0
             // Decode happens IN-CORE (libopus multistream) — AudioToolbox's Opus path is
             // stereo-only — and is handed back as interleaved f32 PCM in wire channel order.
             // Per-iteration autorelease pool: no runloop on this thread (see Stage2Pipeline).
@@ -421,6 +422,17 @@ public final class SessionAudio {
                         ring.write(base, count: pcm.frameCount * pcm.channels)
                     }
                 }
+                // Periodic vitals (~10 s at the protocol's 5 ms frames). The other three clients
+                // log buffer depth and underruns; without this an Apple audio report — latency or
+                // dropout — arrives with no numbers at all, which is the position every platform
+                // was in before the 2026-08 audio work.
+                drained += 1
+                if drained % 2_000 == 0 {
+                    let s = ring.stats
+                    log.info(
+                        "audio: buffer_ms=\(s.bufferedMS) target_ms=\(s.targetMS) underruns=\(s.underruns) drift_sheds=\(s.sheds)"
+                    )
+                }
                 return true
                 }
             }

clients/apple/Sources/PunktfunkKit/Views/StreamViewIOS.swift

@@ -186,6 +186,16 @@ public final class StreamViewController: StreamViewControllerBase {
     // pointer back to iPadOS, so an unwanted drop is re-requested below. The DELIBERATE releases
     // (⌘⎋, ⌃⌥⇧Q, the Stream menu, backgrounding) all clear `captured` first, so `wantsPointerLock`
     // is already false when their drop is observed and none of them are fought here.
+    //
+    // Recovery is TWO-STAGE, because either stage alone leaves a hole:
+    //   1. the burst below, fired the instant the drop is observed — wins back a lock the system
+    //      is willing to return immediately (a transient drop that wasn't Escape at all);
+    //   2. a CLICK into the video while still captured (`onPointerButton`) — the fallback for the
+    //      Escape case proper, where the platform declines during the moment right after its own
+    //      release gesture and the burst therefore expires having achieved nothing.
+    // Stage 2 is what keeps a lost burst from being permanent: `captured` is still true, so no
+    // other path would ever ask again, and the capture would spend the rest of its life on the
+    // absolute pointer — clicking correctly, aiming not at all.
     /// Whether this capture ever actually held the lock. Only a lock we HELD is worth winning back
     /// — never having been granted one means the scene doesn't qualify, not that Esc took it.
     /// Cleared when capture ends, so each capture starts from a clean slate.
@@ -446,6 +456,31 @@ public final class StreamViewController: StreamViewControllerBase {
             }
             guard self.inputCapture?.gcMouseForwarding == false else { return }
             self.inputCapture?.sendMouseButton(button, pressed: down)
+            // …and if we're captured but NOT locked, this click is also the recovery gesture for an
+            // Escape-drop the burst lost. iPadOS refuses to re-lock in the moment right after its
+            // own "let me out" gesture, so the burst fired at the drop can spend its whole budget
+            // and give up while the capture is still wanted. Nothing else would ever re-ask —
+            // setCaptured is the only other requester and a bare Esc never clears `captured` — so
+            // without this the session stays on the absolute path for the rest of the capture:
+            // clicks still land where you aim (absolute positions keep forwarding) but the game
+            // gets no relative deltas, so camera look is dead. A click is a real user gesture,
+            // which is exactly what the platform wants before it will hand the lock back.
+            //
+            // On the button UP, so the click has fully forwarded on ONE transport first: asking on
+            // the DOWN can flip `gcMouseForwarding` mid-click and strand the release on the GCMouse
+            // path. Gated on `pointerLockWasEngaged` exactly as the drop path is, so a scene that
+            // never qualifies (Stage Manager, Split View) is never bursted at, and on a burst not
+            // already being in flight — a pending burst mutes absolute motion, so re-arming one on
+            // every click of a menu the user is still aiming around would freeze the cursor between
+            // clicks. Only once it has settled does a further click buy a fresh budget (clearing the
+            // attempt counter, so a gesture isn't refused inside the 2 s window the drop's own burst
+            // may have just spent).
+            if !down, self.wantsPointerLock, self.pointerLockWasEngaged,
+                !self.pointerRelockPending, self.pointerLockEngaged() != true {
+                self.pointerRelockAttempt = 0
+                self.updatePointerLockChain() // a reparent since the drop would break the walk to us
+                self.requestPointerRelock()
+            }
         }
         // Scroll is the ONE indirect channel that is NOT gated on the lock. The scroll pan keeps
         // firing while the scene is pointer-locked (it is the only way trackpad two-finger scrolling

clients/apple/Tests/PunktfunkKitTests/AudioRingDriftTests.swift

@@ -0,0 +1,95 @@
+// The Apple half of the shared de-jitter policy (`punktfunk_core::audio::JitterPolicy`, whose
+// constants `AudioRing` mirrors). These pin the two behaviours a listener actually notices, in the
+// one client where the policy is hand-written in a second language rather than shared as code — so
+// a divergence from the Rust side shows up here rather than as a field report.
+//
+// The defect being pinned: the ring primed *up* to a target and clamped at a ceiling, with nothing
+// walking the depth back *down*. Host-vs-DAC clock skew of a few dozen ppm therefore added latency
+// permanently, and the only correction was a `highWater` shed that dropped `2 x prefill` at once —
+// its own comment called that "one audible blip".
+
+#if !os(tvOS)
+import XCTest
+
+@testable import PunktfunkKit
+
+final class AudioRingDriftTests: XCTestCase {
+    private let channels = 2
+    private var perMS: Int { 48 * channels }
+
+    /// Run `ms` of audio through the ring at a `quantumMS` device where the producer delivers
+    /// `driftPPM` more than the consumer takes. Returns `(final ms, peak ms, silent callbacks)`.
+    private func simulate(ms: Int, quantumMS: Int, driftPPM: Int) -> (Int, Int, Int) {
+        let ring = AudioRing(capacity: 48_000 * channels, channels: channels)
+        let want = quantumMS * perMS
+        var scratch = [Float](repeating: 0, count: want)
+        // Non-zero so a silent callback is distinguishable from real audio.
+        let producer = [Float](repeating: 0.25, count: want + 8)
+        var carry = 0, peak = 0, final = 0, silent = 0
+
+        for i in 0..<(ms / quantumMS) {
+            carry += want * driftPPM
+            let extra = carry / 1_000_000
+            carry -= extra * 1_000_000
+            producer.withUnsafeBufferPointer { ring.write($0.baseAddress!, count: want + extra) }
+
+            scratch.withUnsafeMutableBufferPointer { ring.read(into: $0.baseAddress!, count: want) }
+            // Skip the priming window at the very start.
+            if i > 20, scratch.allSatisfy({ $0 == 0 }) { silent += 1 }
+            peak = max(peak, ring.bufferedMS)
+            final = ring.bufferedMS
+        }
+        return (final, peak, silent)
+    }
+
+    /// THE regression: with the host clock running fast, buffered latency must return to target
+    /// instead of climbing to the hard cap and staying pinned there. +200 ppm is deliberately
+    /// harsher than real hardware (tens of ppm).
+    func testDriftDoesNotRatchetLatencyToTheCeiling() {
+        let (final, peak, silent) = simulate(ms: 5 * 60 * 1_000, quantumMS: 5, driftPPM: 200)
+        // Must settle inside the headroom band (target 20 + headroom 30), never near the 90 ms cap.
+        XCTAssertLessThanOrEqual(final, 50, "settled at \(final) ms — that is the ratchet")
+        XCTAssertLessThanOrEqual(peak, 50, "peaked at \(peak) ms")
+        XCTAssertEqual(silent, 0, "drift correction must never starve the callback")
+    }
+
+    /// The mirror case: a host clock running SLOW must keep audio flowing rather than being
+    /// "corrected" into a stutter.
+    func testNegativeDriftKeepsPlaying() {
+        let (_, _, silent) = simulate(ms: 2 * 60 * 1_000, quantumMS: 5, driftPPM: -200)
+        XCTAssertEqual(silent, 0, "a draining ring must re-prime, not chatter")
+    }
+
+    /// A device that pulls a large quantum cannot sustain a target below it — the ring must lift
+    /// its target rather than oscillating prime → dropout → re-prime forever.
+    func testLargeDeviceQuantumStillPlays() {
+        let (_, _, silent) = simulate(ms: 60 * 1_000, quantumMS: 40, driftPPM: 0)
+        XCTAssertEqual(silent, 0, "a 40 ms quantum must not starve a 20 ms target")
+    }
+
+    /// One transient drain must not manufacture a whole target's worth of fresh silence: the ring
+    /// de-primes only after a RUN of short reads.
+    func testSingleShortReadDoesNotDeprime() {
+        let ring = AudioRing(capacity: 48_000 * channels, channels: channels)
+        let want = 5 * perMS
+        var scratch = [Float](repeating: 0, count: want)
+        // Prime well past target.
+        let big = [Float](repeating: 0.5, count: 60 * perMS)
+        big.withUnsafeBufferPointer { ring.write($0.baseAddress!, count: big.count) }
+        scratch.withUnsafeMutableBufferPointer { ring.read(into: $0.baseAddress!, count: want) }
+        XCTAssertTrue(scratch.contains { $0 != 0 }, "should be playing after priming")
+
+        // Drain it dry with one oversized read, then feed a normal quantum again. The length comes
+        // off the buffer pointer, not off `huge`: touching the array inside the closure that is
+        // already holding it exclusively is an exclusivity violation.
+        var huge = [Float](repeating: 0, count: 200 * perMS)
+        huge.withUnsafeMutableBufferPointer { ring.read(into: $0.baseAddress!, count: $0.count) }
+        let feed = [Float](repeating: 0.5, count: want)
+        feed.withUnsafeBufferPointer { ring.write($0.baseAddress!, count: want) }
+        scratch.withUnsafeMutableBufferPointer { ring.read(into: $0.baseAddress!, count: want) }
+        XCTAssertTrue(
+            scratch.contains { $0 != 0 },
+            "a single short read must not force a full re-prime")
+    }
+}
+#endif

clients/windows/src/app/settings.rs

@@ -1045,13 +1045,18 @@ pub(crate) fn settings_page(
         let ss = set_screen.clone();
         button("Third-party licenses").on_click(move || ss.call(Screen::Licenses))
     };
-    // The client log's home (%LOCALAPPDATA%\punktfunk\logs) — the file every "check the
-    // client log" message means, which until this row had no way in from the UI at all.
-    // The folder rather than the file so the rotated `.old` generation is in reach too.
-    // Best-effort, like the log itself: a missing dir or a failed spawn stays silent.
+    // The client log's home — the file every "check the client log" message means, which until
+    // this row had no way in from the UI at all. The folder rather than the file so the rotated
+    // `.old` generation is in reach too.
+    //
+    // `real_dir` (not the literal %LOCALAPPDATA% path) because Explorer lives outside our MSIX
+    // container: handed a path the package redirection keeps from ever existing, it silently
+    // opens the user's Documents folder instead of failing, which is precisely what this button
+    // shipped doing. The `is_dir` guard keeps that fallback unreachable — if the resolve ever
+    // comes back wrong, the click does nothing rather than landing somewhere misleading.
+    // Best-effort otherwise, like the log itself: a failed spawn stays silent.
     let logs_button = button("Open log folder").on_click(|| {
-        if let Some(dir) = crate::logfile::log_dir() {
-            let _ = std::fs::create_dir_all(&dir);
+        if let Some(dir) = crate::logfile::real_dir().filter(|d| d.is_dir()) {
             let _ = std::process::Command::new("explorer.exe").arg(&dir).spawn();
         }
     });

clients/windows/src/logfile.rs

@@ -10,6 +10,10 @@
 //! Mirrors the host's convention (`%ProgramData%\punktfunk\logs`, size-capped): a file over
 //! 10 MB is rotated to `.old` at the next client start, one generation kept. Everything is
 //! best-effort — a missing/locked directory degrades to plain stderr, never a startup failure.
+//!
+//! Two paths, deliberately: [`log_dir`] is what we open files through, [`real_dir`] is where
+//! they actually land. Under MSIX those differ, and only the second one is fit to show a user
+//! or hand to Explorer.
 
 use std::fs::{File, OpenOptions};
 use std::io::{self, BufRead, Write};
@@ -21,14 +25,74 @@ const ROTATE_BYTES: u64 = 10 * 1024 * 1024;
 
 static SINK: OnceLock<Option<Arc<Mutex<File>>>> = OnceLock::new();
 
-/// The log directory — Settings ▸ About's "Open log folder" opens it in Explorer.
-pub(crate) fn log_dir() -> Option<PathBuf> {
+/// The log directory we WRITE through: `%LOCALAPPDATA%\punktfunk\logs`.
+///
+/// Correct to open files under, but NOT necessarily where the bytes land — see [`real_dir`].
+/// Anything shown to a user or handed to another process wants that one instead.
+fn log_dir() -> Option<PathBuf> {
     Some(PathBuf::from(std::env::var_os("LOCALAPPDATA")?).join(r"punktfunk\logs"))
 }
 
+/// The log directory as it exists ON DISK — Settings ▸ About's "Open log folder" opens this in
+/// Explorer, and [`path`] names it in the startup line and the failed-spawn banner.
+///
+/// The shipping client is a full-trust MSIX package, and Windows redirects a packaged app's
+/// `%LOCALAPPDATA%` writes into its private `…\Packages\<family>\LocalCache\Local\…`. We create
+/// and append through that redirection without ever seeing it, so [`log_dir`] is the right path
+/// to WRITE to yet names a directory that never exists on disk. Explorer runs OUTSIDE the
+/// container: it resolves the literal path, finds nothing, and silently falls back to the user's
+/// Documents folder — which is exactly what "Open log folder" did in every packaged install, and
+/// what the two "check <path>" messages pointed at. An unpackaged dev run creates the literal
+/// directory for real, which is why this only ever showed up in the field.
+///
+/// Canonicalizing the directory we just created resolves through the redirection on a packaged
+/// run and changes nothing on an unpackaged one, so there is no package identity to detect.
+pub(crate) fn real_dir() -> Option<PathBuf> {
+    let dir = log_dir()?;
+    std::fs::create_dir_all(&dir).ok()?;
+    Some(std::fs::canonicalize(&dir).map_or(dir, strip_verbatim))
+}
+
+/// Undo the `\\?\` that [`std::fs::canonicalize`] always prefixes. Explorer refuses a verbatim
+/// path — it would take the very same silent Documents fallback [`real_dir`] exists to avoid —
+/// and it is noise in a line a user is meant to read and act on.
+fn strip_verbatim(p: PathBuf) -> PathBuf {
+    use std::path::{Component, Prefix};
+
+    // Scoped so the borrow ends before the `return p` below can move it.
+    let head = match p.components().next() {
+        Some(Component::Prefix(pre)) => match pre.kind() {
+            // `\\?\C:\…` → `C:\…`
+            Prefix::VerbatimDisk(drive) => Some(PathBuf::from(format!(r"{}:\", drive as char))),
+            // `\\?\UNC\server\share\…` → `\\server\share\…` (a roaming profile on a share).
+            // Built through `OsString`, which appends verbatim — `PathBuf::push` would apply
+            // separator logic to the bare `\\` and mangle it.
+            Prefix::VerbatimUNC(server, share) => {
+                let mut unc = std::ffi::OsString::from(r"\\");
+                unc.push(server);
+                unc.push(r"\");
+                unc.push(share);
+                Some(PathBuf::from(unc))
+            }
+            // Already a plain path — nothing to undo.
+            _ => None,
+        },
+        _ => None,
+    };
+    let Some(mut out) = head else { return p };
+    // `skip(1)` drops the prefix; the `RootDir` that follows it is already in `head`.
+    out.extend(
+        p.components()
+            .skip(1)
+            .filter(|c| !matches!(c, Component::RootDir)),
+    );
+    out
+}
+
 /// The log file's path, for the "logs land here" startup line and the failed-spawn banner.
+/// Resolved like [`real_dir`] — a path a user is told to check has to be the one on disk.
 pub(crate) fn path() -> Option<PathBuf> {
-    Some(log_dir()?.join("client.log"))
+    Some(real_dir()?.join("client.log"))
 }
 
 /// Open (rotating first) and cache the sink. Called once at startup, before the tracing
@@ -97,3 +161,67 @@ pub(crate) fn forward_child_stderr(stderr: impl io::Read + Send + 'static) {
             }
         });
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The shape `canonicalize` actually returns for a local profile. Explorer treats a `\\?\`
+    /// path as unresolvable and opens Documents instead, so the prefix has to come off.
+    #[test]
+    fn verbatim_disk_prefix_comes_off() {
+        let p = PathBuf::from(r"\\?\C:\Users\ada\AppData\Local\punktfunk\logs");
+        assert_eq!(
+            strip_verbatim(p),
+            PathBuf::from(r"C:\Users\ada\AppData\Local\punktfunk\logs")
+        );
+    }
+
+    /// The MSIX-redirected form is what the fix is for: same treatment, longer path.
+    #[test]
+    fn verbatim_disk_prefix_comes_off_for_the_package_local_cache() {
+        let p = PathBuf::from(
+            r"\\?\C:\Users\ada\AppData\Local\Packages\unom.Punktfunk_8wekyb3d8bbwe\LocalCache\Local\punktfunk\logs",
+        );
+        assert_eq!(
+            strip_verbatim(p),
+            PathBuf::from(
+                r"C:\Users\ada\AppData\Local\Packages\unom.Punktfunk_8wekyb3d8bbwe\LocalCache\Local\punktfunk\logs"
+            )
+        );
+    }
+
+    /// A roaming profile on a share canonicalizes to `\\?\UNC\…`; the plain UNC form is what
+    /// Explorer takes. `\\server\share` must survive intact — dropping either half, or letting
+    /// `PathBuf::push`'s separator logic at the bare `\\`, yields a path that opens nothing.
+    #[test]
+    fn verbatim_unc_prefix_becomes_a_plain_unc_path() {
+        let p = PathBuf::from(r"\\?\UNC\fileserv\profiles\ada\AppData\Local\punktfunk\logs");
+        assert_eq!(
+            strip_verbatim(p),
+            PathBuf::from(r"\\fileserv\profiles\ada\AppData\Local\punktfunk\logs")
+        );
+    }
+
+    /// An unpackaged dev run resolves to a path that was never verbatim — leave it alone.
+    #[test]
+    fn plain_path_is_untouched() {
+        let p = PathBuf::from(r"C:\Users\ada\AppData\Local\punktfunk\logs");
+        assert_eq!(strip_verbatim(p.clone()), p);
+    }
+
+    /// Whatever the run, the resolved directory is one Explorer can open: it exists, and it
+    /// carries no verbatim prefix. This is the button's actual precondition.
+    #[test]
+    fn real_dir_is_an_openable_directory() {
+        let Some(dir) = real_dir() else {
+            return; // no LOCALAPPDATA (not a normal user session) — nothing to assert
+        };
+        assert!(dir.is_dir(), "{} is not a directory", dir.display());
+        assert!(
+            !dir.to_string_lossy().starts_with(r"\\?\"),
+            "{} kept its verbatim prefix",
+            dir.display()
+        );
+    }
+}

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

@@ -168,9 +168,18 @@ struct PlayerData {
     /// Drained chunk Vecs go back here for the decode side to refill (allocation pool).
     recycle: SyncSender<Vec<f32>>,
     ring: VecDeque<f32>,
-    primed: bool,
+    /// Shared ms-denominated de-jitter policy: prime depth, drift correction, de-prime
+    /// hysteresis. Replaces the old `3 × quantum` target, which meant 15 ms at a 5 ms graph
+    /// quantum and a silent 64 ms at a 20 ms one, and the `if ring.is_empty()` re-prime, where
+    /// one transient drain manufactured a whole target's worth of fresh silence.
+    policy: punktfunk_core::audio::JitterPolicy,
     /// Interleaved channel count this stream was opened with (2/6/8).
     channels: usize,
+    /// Diagnostics (WP0.3), logged ~every 10 s: the audio plane used to be entirely silent in a
+    /// client log, so a latency or dropout report had nothing to go on.
+    underruns: u64,
+    sheds: u64,
+    callbacks: u64,
 }
 
 fn pw_thread(
@@ -223,8 +232,14 @@ fn pw_thread(
         rx: pcm_rx,
         recycle: recycle_tx,
         ring: VecDeque::new(),
-        primed: false,
+        policy: punktfunk_core::audio::JitterPolicy::new(
+            punktfunk_core::audio::JitterTuning::PIPEWIRE,
+            channels as u8,
+        ),
         channels,
+        underruns: 0,
+        sheds: 0,
+        callbacks: 0,
     };
 
     let _listener = stream
@@ -252,23 +267,29 @@ fn pw_thread(
                 let want_frames = data.data().map(|s| s.len() / stride).unwrap_or(0);
                 let want = want_frames * ud.channels;
 
-                // Adaptive jitter buffer (same shape as the host's virtual mic): prime to
-                // ~3 quanta, cap at ~1 quantum of slack beyond that, re-prime after a
-                // genuine drain.
-                let target = (3 * want).clamp(720 * ud.channels, 9600 * ud.channels);
-                while ud.ring.len() > target.max(want) + want {
-                    ud.ring.pop_front();
-                }
-                if !ud.primed && ud.ring.len() >= target {
-                    ud.primed = true;
+                // Shared de-jitter policy: prime depth in MILLISECONDS, smooth drift correction
+                // (a crossfaded 5 ms shed) so latency returns to target instead of ratcheting,
+                // and a hard cap as the backstop.
+                let step = ud.policy.step(ud.ring.len(), want);
+                if step.drop_front > 0 {
+                    ud.sheds += 1;
+                    punktfunk_core::audio::crossfade_drop(
+                        &mut ud.ring,
+                        step.drop_front,
+                        step.crossfade,
+                    );
                 }
 
+                let mut ran_short = false;
                 let n_frames = if let Some(slice) = data.data() {
                     for k in 0..want {
-                        let s = if ud.primed {
-                            ud.ring.pop_front().unwrap_or(0.0)
-                        } else {
+                        let s = if step.silence {
                             0.0
+                        } else {
+                            ud.ring.pop_front().unwrap_or_else(|| {
+                                ran_short = true;
+                                0.0
+                            })
                         };
                         let off = k * 4;
                         slice[off..off + 4].copy_from_slice(&s.to_le_bytes());
@@ -277,8 +298,21 @@ fn pw_thread(
                 } else {
                     0
                 };
-                if ud.ring.is_empty() {
-                    ud.primed = false;
+                // No-op while un-primed (the policy ignores it), so a deliberate priming silence
+                // is never miscounted as an underrun.
+                ud.policy.note_read(ran_short);
+                ud.underruns += u64::from(ran_short);
+                ud.callbacks += 1;
+                // ~10 s at a 5 ms quantum; the exact cadence does not matter, only that the
+                // plane stops being invisible.
+                if ud.callbacks % 2_000 == 0 {
+                    tracing::debug!(
+                        buffer_ms = ud.policy.avg_depth_ms(),
+                        target_ms = ud.policy.target_ms(),
+                        underruns = ud.underruns,
+                        drift_sheds = ud.sheds,
+                        "audio playback"
+                    );
                 }
                 let chunk = data.chunk_mut();
                 *chunk.offset_mut() = 0;

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

@@ -3,14 +3,15 @@
 //!
 //! The WASAPI twin of `audio.rs` (PipeWire) — same public surface (`AudioPlayer::spawn`/
 //! `take_buffer`/`push`, `MicStreamer::spawn`), swapped in by lib.rs's `#[path]` so the
-//! session pump compiles against one `crate::audio` on both OSes. Adapted from
-//! `clients/windows/src/audio.rs` (which remains the WinUI shell's own copy until its
-//! built-in streaming path is deleted).
+//! session pump compiles against one `crate::audio` on both OSes. It began as a copy of the
+//! WinUI shell's own audio path; that shell's built-in streaming path has since been deleted,
+//! so this is now the only WASAPI client ring.
 //!
-//! Playback mirrors the host's virtual-mic producer's adaptive jitter buffer: the session
-//! pump pushes 5 ms Opus-decoded chunks on the network clock; the WASAPI render thread
-//! pulls whole event-driven quanta on the device clock. Prime to ~3 quanta before
-//! producing, cap the ring so latency stays bounded, re-prime after a real drain.
+//! Playback: the session pump pushes 5 ms Opus-decoded chunks on the network clock; the WASAPI
+//! render thread pulls whole event-driven quanta on the device clock. The depth policy between
+//! them is the SHARED `punktfunk_core::audio::JitterPolicy` (`JitterTuning::WASAPI`) — target in
+//! milliseconds, crossfaded drift correction, de-prime hysteresis — so all four clients behave
+//! the same way and none of them can ratchet latency upward.
 //!
 //! WASAPI objects are COM-apartment-bound and not `Send`, so they live on a dedicated
 //! thread (the same discipline as the host's `wasapi_cap`); only the channels + stop flag
@@ -250,10 +251,20 @@ fn render_thread(
         audio_client.start_stream().context("start render stream")?;
         let _ = ready.send(Ok(()));
 
-        // Adaptive jitter buffer, in f32-byte units (same shape as the host's virtual mic).
-        let mut ring: VecDeque<u8> = VecDeque::new();
-        let mut primed = false;
+        // De-jitter ring, in interleaved f32 SAMPLES (it used to be raw bytes, which made the
+        // depth arithmetic byte-vs-sample and kept it from sharing the policy and the crossfade
+        // helper with the other three clients).
+        let mut ring: VecDeque<f32> = VecDeque::new();
+        // Shared ms-denominated policy: prime depth, crossfaded drift correction so latency
+        // returns to target instead of ratcheting, and de-prime hysteresis — the last replacing
+        // the old `if ring.is_empty()`, where a single transient drain manufactured a whole
+        // target's worth of fresh silence.
+        let mut policy = punktfunk_core::audio::JitterPolicy::new(
+            punktfunk_core::audio::JitterTuning::WASAPI,
+            channels,
+        );
         let mut out = Vec::new(); // per-quantum scratch, reused across iterations
+        let (mut underruns, mut sheds, mut callbacks) = (0u64, 0u64, 0u64);
 
         while !stop.load(Ordering::Relaxed) {
             if h_event.wait_for_event(100).is_err() {
@@ -262,9 +273,7 @@ fn render_thread(
             // Drain everything the pump has queued into the ring, returning each drained
             // Vec to the pool (a full/closed pool drops it).
             while let Ok(mut chunk) = pcm_rx.try_recv() {
-                for s in chunk.iter() {
-                    ring.extend(s.to_le_bytes());
-                }
+                ring.extend(chunk.iter().copied());
                 chunk.clear();
                 let _ = recycle_tx.try_send(chunk);
             }
@@ -274,28 +283,40 @@ fn render_thread(
             if avail_frames == 0 {
                 continue;
             }
-            let want_bytes = avail_frames * block_align;
+            let want = avail_frames * channels as usize;
 
-            // Prime to ~3 quanta; cap at ~1 quantum of slack beyond that; re-prime on drain.
-            let target = (3 * want_bytes).clamp(720 * block_align, 9600 * block_align);
-            let cap = target.max(want_bytes) + want_bytes;
-            if ring.len() > cap {
-                ring.drain(..ring.len() - cap);
-            }
-            if !primed && ring.len() >= target {
-                primed = true;
+            let step = policy.step(ring.len(), want);
+            if step.drop_front > 0 {
+                sheds += 1;
+                punktfunk_core::audio::crossfade_drop(&mut ring, step.drop_front, step.crossfade);
             }
 
             out.clear();
-            out.resize(want_bytes, 0);
-            if primed {
-                let n = ring.len().min(want_bytes);
-                for (dst, b) in out.iter_mut().zip(ring.drain(..n)) {
-                    *dst = b;
+            out.resize(avail_frames * block_align, 0);
+            let mut ran_short = false;
+            if !step.silence {
+                // `out` is exactly `want` f32s wide (avail_frames × channels × 4 bytes).
+                for dst in out.chunks_exact_mut(4) {
+                    let s = ring.pop_front().unwrap_or_else(|| {
+                        ran_short = true;
+                        0.0
+                    });
+                    dst.copy_from_slice(&s.to_le_bytes());
                 }
             }
-            if ring.is_empty() {
-                primed = false;
+            // No-op while un-primed (the policy ignores it), so a deliberate priming silence is
+            // never miscounted as an underrun.
+            policy.note_read(ran_short);
+            underruns += u64::from(ran_short);
+            callbacks += 1;
+            if callbacks % 1_000 == 0 {
+                tracing::debug!(
+                    buffer_ms = policy.avg_depth_ms(),
+                    target_ms = policy.target_ms(),
+                    underruns,
+                    drift_sheds = sheds,
+                    "audio playback"
+                );
             }
             render_client
                 .write_to_device(avail_frames, &out, None)

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

@@ -285,21 +285,6 @@ fn set_valve_hidapi(enabled: bool) {
     sdl3::hint::set("SDL_JOYSTICK_HIDAPI_STEAM", v);
 }
 
-/// Disable the Valve HIDAPI drivers **before SDL exists** — call this alongside the other
-/// pre-`SDL_Init` hints, not after a subsystem is up.
-///
-/// The damage these drivers do happens at *enumeration*, which is part of initialising the
-/// joystick/gamepad subsystem. Setting the hint afterwards does detach the driver, but only after
-/// it has already sent the Deck its `ID_CLEAR_DIGITAL_MAPPINGS` + `TRACKPAD_NONE` — so the
-/// built-in trackpad-mouse dies system-wide and stays dead until the firmware watchdog restores
-/// lizard mode seconds later. The threaded worker ([`run`]) has always done this in the right
-/// order; the caller-pumped path could not, because by the time it receives a
-/// [`sdl3::GamepadSubsystem`] the enumeration has already happened. Hence a separate entry point
-/// its callers can put in the right place.
-pub fn preinit_disable_valve_hidapi() {
-    set_valve_hidapi(false);
-}
-
 /// Map the SDL-reported controller type to the virtual pad we'd ask the host to create.
 fn pref_for_type(t: sdl3::gamepad::GamepadType) -> GamepadPref {
     use sdl3::gamepad::GamepadType as T;
@@ -408,12 +393,9 @@ impl GamepadService {
     /// and calls [`GamepadPump::tick`] once per loop iteration (the threaded worker's
     /// per-wakeup work: ctl drain, chord-hold check, menu repeat, feedback).
     ///
-    /// The Valve HIDAPI drivers are held off here too, but this is **too late to be the only
-    /// place it happens**: the `subsystem` argument means enumeration is already done, and that
-    /// is when the Deck driver kills the trackpad-mouse. The caller must also call
-    /// [`preinit_disable_valve_hidapi`] with its other pre-`SDL_Init` hints. This call still
-    /// earns its place — it re-asserts "off" for a process that ran a session earlier — but on
-    /// its own it only detaches a driver that has already done the damage.
+    /// Like the threaded worker, this disables the Valve HIDAPI drivers up front (their
+    /// mere enumeration kills the Deck's trackpad-mouse system-wide); they are enabled
+    /// for the duration of an attached session only.
     pub fn pumped(subsystem: sdl3::GamepadSubsystem) -> (GamepadService, GamepadPump) {
         set_valve_hidapi(false);
         let pads = Arc::new(Mutex::new(Vec::new()));
@@ -574,38 +556,6 @@ impl GamepadPump {
         self.worker.menu_poll();
         self.worker.render_feedback();
     }
-
-    /// Close every forwarded slot — flush its held wire state, tell the host to remove the pad,
-    /// and physically silence it. Call once on the way out of the caller's event loop.
-    ///
-    /// [`GamepadService::detach`] only *posts* `Ctl::Detach`; the close — the flush, the host-side
-    /// `GamepadRemove`, and the explicit `set_rumble(0, 0)` backstop in `close_slot_at` — happens
-    /// when the pump next drains it. An exit path that detached and then left the loop without
-    /// another [`tick`](Self::tick) therefore skipped all of it, and nothing else would: the slots
-    /// hold no `Drop` that silences them. A pad left mid-buzz stayed buzzing.
-    ///
-    /// This closes the slots directly rather than draining the queued `Ctl::Detach` that would
-    /// have done it. Same physical outcome by a shorter path, and deliberately so: this also runs
-    /// from `Drop`, and `drain_ctl` reaches `Mutex::lock().unwrap()`, which on a poisoned lock
-    /// would panic — during an unwind that aborts the process. Closing a slot touches no lock.
-    ///
-    /// Idempotent, and safe with nothing attached.
-    pub fn shutdown(&mut self) {
-        self.worker.close_all_slots();
-    }
-}
-
-/// The silence backstop of last resort. A caller's loop can also leave by `?` on a fatal overlay
-/// or present error — several paths do — and those would skip an explicit
-/// [`shutdown`](GamepadPump::shutdown) entirely, leaving a forwarded pad buzzing on the way out.
-///
-/// Callers should still call `shutdown` at their normal exit rather than lean on this: the pad
-/// wants to go quiet *before* a long teardown (session join, `vkDeviceWaitIdle`), not after it.
-/// Doing both is free — `shutdown` is idempotent.
-impl Drop for GamepadPump {
-    fn drop(&mut self) {
-        self.shutdown();
-    }
 }
 
 /// The lowest wire pad index (0..[`MAX_PADS`](punktfunk_core::input::MAX_PADS)) not already held
@@ -1676,11 +1626,6 @@ impl Worker {
                 HidOutput::PlayerLeds { bits, .. } if is_ds => {
                     let _ = slot.pad.send_effect(&Ds5Feedback::player_packet(bits));
                 }
-                // Every other pad with player LEDs gets them through SDL, which owns the
-                // per-device pattern. This used to fall through and do nothing at all.
-                HidOutput::PlayerLeds { bits, .. } => {
-                    let _ = set_player_leds(&slot.pad, bits);
-                }
                 HidOutput::Trigger {
                     which, ref effect, ..
                 } if is_ds => {
@@ -1688,43 +1633,12 @@ impl Worker {
                         .pad
                         .send_effect(&Ds5Feedback::trigger_packet(which, effect));
                 }
-                // 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.
-                HidOutput::Trigger { .. }
-                | HidOutput::TrackpadHaptic { .. }
-                | HidOutput::HidRaw { .. } => {}
+                _ => {}
             }
         }
     }
 }
 
-/// The SDL player index for the wire's positional player-LED `bits`, or `None` for "no player".
-///
-/// The wire carries a bitmask — one bit per LED, low 5 — while SDL wants a player *index* and owns
-/// the per-device pattern. The count bridges them: every convention that reaches this wire spells
-/// "player N" as N lit LEDs, both the DualSense patterns (`0x04`, `0x0A`, `0x15`, `0x1B`, `0x1F`)
-/// and the Switch/XInput run of low bits (`0x01`, `0x03`, `0x07`, `0x0F`). SDL's index is 0-based,
-/// so player 1 is index 0; no lit LED means *no* player rather than player 0.
-///
-/// Split out from [`set_player_leds`] so the mapping is testable — an `sdl3::Gamepad` needs a real
-/// device, so nothing that takes one can be.
-fn player_index_from_bits(bits: u8) -> Option<u16> {
-    match (bits & 0x1F).count_ones() {
-        0 => None,
-        n => Some((n - 1) as u16),
-    }
-}
-
-/// Drive a non-DualSense pad's player LEDs from the wire's positional `bits`.
-fn set_player_leds(pad: &sdl3::gamepad::Gamepad, bits: u8) -> Result<(), sdl3::Error> {
-    match player_index_from_bits(bits) {
-        None => pad.unset_player_index(),
-        Some(i) => pad.set_player_index(i),
-    }
-}
-
 /// The wire pad index a [`HidOutput`] is addressed to (every variant carries `pad`).
 fn hidout_pad(h: &HidOutput) -> u8 {
     match h {
@@ -2094,43 +2008,3 @@ mod slot_tests {
         );
     }
 }
-
-#[cfg(test)]
-mod player_led_tests {
-    use super::*;
-
-    /// Both conventions that reach this wire spell "player N" as N lit LEDs, so the count is the
-    /// player number regardless of WHICH bits a given pad lights. Pinned because the mapping is
-    /// otherwise only obvious once you have seen both patterns side by side.
-    #[test]
-    fn player_index_counts_lit_leds_for_both_conventions() {
-        // DualSense / hid-playstation patterns — non-contiguous, symmetric about the centre LED.
-        assert_eq!(player_index_from_bits(0x04), Some(0)); // player 1
-        assert_eq!(player_index_from_bits(0x0A), Some(1)); // player 2
-        assert_eq!(player_index_from_bits(0x15), Some(2)); // player 3
-        assert_eq!(player_index_from_bits(0x1B), Some(3)); // player 4
-        assert_eq!(player_index_from_bits(0x1F), Some(4)); // player 5
-
-        // Switch/XInput style — a contiguous run of low bits, the same count each time.
-        assert_eq!(player_index_from_bits(0x01), Some(0));
-        assert_eq!(player_index_from_bits(0x03), Some(1));
-        assert_eq!(player_index_from_bits(0x07), Some(2));
-        assert_eq!(player_index_from_bits(0x0F), Some(3));
-    }
-
-    /// No lit LED is "no player", NOT player 0 — the difference between LEDs off and player 1 lit.
-    #[test]
-    fn no_lit_led_is_no_player() {
-        assert_eq!(player_index_from_bits(0x00), None);
-        // Only the low 5 bits are player LEDs; junk above them must not invent a player.
-        assert_eq!(player_index_from_bits(0xE0), None);
-    }
-
-    /// The mask is applied before counting, so out-of-range bits cannot inflate the index past
-    /// the 5 real LEDs.
-    #[test]
-    fn high_bits_are_masked_off_before_counting() {
-        assert_eq!(player_index_from_bits(0xFF), Some(4)); // 0x1F worth of LEDs, not 8
-        assert_eq!(player_index_from_bits(0xE4), Some(0)); // 0x04 with junk on top
-    }
-}

crates/pf-host-config/src/lib.rs

@@ -57,6 +57,82 @@ pub fn env_on(name: &str) -> Option<bool> {
     })
 }
 
+/// Where desktop audio should be audible — which decides the render endpoint the loopback captures.
+///
+/// Supersedes the two env-only knobs that used to encode this (`PUNKTFUNK_HOST_AUDIO`,
+/// `PUNKTFUNK_KEEP_DEFAULT`), which stay honoured as back-compat spellings so nobody's `host.env`
+/// breaks. Named modes exist because "which endpoint do we capture" is a routing decision an
+/// operator has to be able to make deliberately — the 2026-08-03 field report is what happens when
+/// the only way to express it is an undocumented environment variable.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum AudioOutputMode {
+    /// Default. Prefer a render endpoint that is silent on the host, so streamed audio does not
+    /// also play out of the host's speakers. Since 2026-08 a silent sink has to be able to carry
+    /// the mix without narrowing it — otherwise real hardware wins anyway.
+    #[default]
+    ClientOnly,
+    /// Prefer real hardware: audio plays on the host as well as the client. The old
+    /// `PUNKTFUNK_HOST_AUDIO=1`.
+    HostAndClient,
+    /// Touch nothing — capture whatever the operator's own default playback device is, and never
+    /// write the default-device policy. The old `PUNKTFUNK_KEEP_DEFAULT=1`.
+    FollowDefault,
+}
+
+impl AudioOutputMode {
+    /// `PUNKTFUNK_AUDIO_OUTPUT_MODE` wins; otherwise fall back to the legacy flags, `follow_default`
+    /// first (it is the more restrictive promise — "do not touch my devices" must not be overridden
+    /// by a stale `PUNKTFUNK_HOST_AUDIO` in the same `host.env`).
+    fn from_env() -> AudioOutputMode {
+        if let Ok(raw) = std::env::var("PUNKTFUNK_AUDIO_OUTPUT_MODE") {
+            if !raw.trim().is_empty() {
+                if let Some(m) = AudioOutputMode::parse(&raw) {
+                    return m;
+                }
+                // Never silently fall through to a different routing than the operator asked for.
+                eprintln!(
+                    "punktfunk: PUNKTFUNK_AUDIO_OUTPUT_MODE={raw:?} is not one of \
+                     client_only/host_and_client/follow_default — using client_only"
+                );
+            }
+        }
+        if std::env::var_os("PUNKTFUNK_KEEP_DEFAULT").is_some() {
+            return AudioOutputMode::FollowDefault;
+        }
+        if std::env::var_os("PUNKTFUNK_HOST_AUDIO").is_some() {
+            return AudioOutputMode::HostAndClient;
+        }
+        AudioOutputMode::ClientOnly
+    }
+
+    pub fn parse(s: &str) -> Option<AudioOutputMode> {
+        match s.trim().to_ascii_lowercase().replace('-', "_").as_str() {
+            "client_only" | "client" => Some(AudioOutputMode::ClientOnly),
+            "host_and_client" | "both" | "host" => Some(AudioOutputMode::HostAndClient),
+            "follow_default" | "follow" => Some(AudioOutputMode::FollowDefault),
+            _ => None,
+        }
+    }
+
+    pub fn as_str(self) -> &'static str {
+        match self {
+            AudioOutputMode::ClientOnly => "client_only",
+            AudioOutputMode::HostAndClient => "host_and_client",
+            AudioOutputMode::FollowDefault => "follow_default",
+        }
+    }
+
+    /// The loopback plan should prefer real hardware over a silent sink.
+    pub fn prefers_host_hardware(self) -> bool {
+        matches!(self, AudioOutputMode::HostAndClient)
+    }
+
+    /// Leave the operator's default playback/recording devices completely alone.
+    pub fn keeps_default(self) -> bool {
+        matches!(self, AudioOutputMode::FollowDefault)
+    }
+}
+
 /// Resolved host configuration. Holds the genuinely-constant operator/dispatch knobs (see module docs for
 /// what is deliberately excluded). Fields read on only one platform are kept alive cross-platform by the
 /// derived `Debug` impl, so the parser can stay a single platform-neutral function.
@@ -99,6 +175,24 @@ pub struct HostConfig {
     /// e.g. webOS TVs, whose GCM decrypt caps at ~100 Mbps); everyone else stays AES-128-GCM.
     /// `PUNKTFUNK_CHACHA20=0`/`false`/`off`/`no` disables.
     pub chacha20: bool,
+    /// `PUNKTFUNK_AUDIO_OUTPUT_MODE` — where desktop audio should be audible, and therefore which
+    /// render endpoint the loopback captures (`client_only` / `host_and_client` / `follow_default`).
+    ///
+    /// A first-class setting because the 2026-08-03 field report needed one: the default
+    /// client-only routing sent that box's whole desktop mix through Steam's voice-carrier virtual
+    /// endpoint for 25 sessions, and the only way to change it was an undocumented environment
+    /// variable. See [`AudioOutputMode`].
+    pub audio_output_mode: AudioOutputMode,
+    /// `PUNKTFUNK_AUDIO_QUALITY` — desktop-audio encode tier (`low` / `standard` / `high`; default
+    /// `high`). Kept as the raw string here because the tier table lives in `punktfunk-core`, and
+    /// this crate is deliberately dependency-free (see the crate doc). The audio thread resolves it
+    /// via `punktfunk_core::audio::AudioTier::parse` and warns on an unknown spelling rather than
+    /// silently downgrading someone's audio.
+    pub audio_quality: Option<String>,
+    /// `PUNKTFUNK_AUDIO_REDUNDANCY` — force the redundant `0xD2` audio plane on or off. `None`
+    /// (the default) = automatic: sent only to a client that asked for it, and only while the link
+    /// is actually losing packets.
+    pub audio_redundancy: Option<bool>,
     /// `PUNKTFUNK_PERF` — per-stage timing instrumentation.
     pub perf: bool,
     /// `PUNKTFUNK_VIDEO_SOURCE` — GameStream video source select. `virtual` (the default — a
@@ -246,6 +340,9 @@ impl HostConfig {
             // Default ON, explicit-off grammar (the client's VIDEO_CAP_CHACHA20 bit is the real
             // per-session switch; see the field doc).
             chacha20: env_on("PUNKTFUNK_CHACHA20").unwrap_or(true),
+            audio_output_mode: AudioOutputMode::from_env(),
+            audio_quality: val("PUNKTFUNK_AUDIO_QUALITY").map(|s| s.trim().to_lowercase()),
+            audio_redundancy: env_on("PUNKTFUNK_AUDIO_REDUNDANCY"),
             perf: flag("PUNKTFUNK_PERF"),
             // Default ON while the interval-stutter field program runs (see the field doc).
             stall_probes: env_on("PUNKTFUNK_STALL_PROBES").unwrap_or(true),
@@ -348,4 +445,50 @@ mod tests {
         // An invalid rate stays invalid rather than being laundered into a real one.
         assert_eq!(c.game_fps(0), 0);
     }
+
+    #[test]
+    fn audio_output_mode_parses_its_spellings() {
+        for (s, want) in [
+            ("client_only", AudioOutputMode::ClientOnly),
+            ("client-only", AudioOutputMode::ClientOnly),
+            ("  CLIENT  ", AudioOutputMode::ClientOnly),
+            ("host_and_client", AudioOutputMode::HostAndClient),
+            ("both", AudioOutputMode::HostAndClient),
+            ("follow_default", AudioOutputMode::FollowDefault),
+            ("follow", AudioOutputMode::FollowDefault),
+        ] {
+            assert_eq!(AudioOutputMode::parse(s), Some(want), "{s:?}");
+        }
+        // Unknown spellings are rejected so the caller can say so, not silently re-routed.
+        for s in ["", "silent", "off", "true"] {
+            assert_eq!(AudioOutputMode::parse(s), None, "{s:?}");
+        }
+        // Round-trip through the canonical spelling.
+        for m in [
+            AudioOutputMode::ClientOnly,
+            AudioOutputMode::HostAndClient,
+            AudioOutputMode::FollowDefault,
+        ] {
+            assert_eq!(AudioOutputMode::parse(m.as_str()), Some(m));
+        }
+    }
+
+    /// The two predicates are what the wiring plan and the capture loop actually branch on, and
+    /// they must stay mutually exclusive: "prefer host hardware" and "touch nothing" are different
+    /// promises, and conflating them would either silence the host or stomp the operator's devices.
+    #[test]
+    fn audio_output_mode_predicates_are_disjoint() {
+        assert_eq!(AudioOutputMode::default(), AudioOutputMode::ClientOnly);
+        for m in [
+            AudioOutputMode::ClientOnly,
+            AudioOutputMode::HostAndClient,
+            AudioOutputMode::FollowDefault,
+        ] {
+            assert!(!(m.prefers_host_hardware() && m.keeps_default()), "{m:?}");
+        }
+        assert!(AudioOutputMode::HostAndClient.prefers_host_hardware());
+        assert!(AudioOutputMode::FollowDefault.keeps_default());
+        assert!(!AudioOutputMode::ClientOnly.prefers_host_hardware());
+        assert!(!AudioOutputMode::ClientOnly.keeps_default());
+    }
 }

crates/pf-inject/src/inject/linux/gamepad.rs

@@ -254,13 +254,45 @@ fn ioctl_ptr<T>(fd: i32, req: libc::c_ulong, arg: *mut T, what: &str) -> Result<
     Ok(())
 }
 
+/// The window a played effect occupies: `replay.delay` of silence, then `replay.length` of rumble.
+#[derive(Clone, Copy)]
+struct Playback {
+    /// When the effect starts contributing — `play + replay.delay`. Until then it is armed but
+    /// silent, which is the whole point of the delay.
+    starts: Instant,
+    /// When it stops, or `None` for replay length 0 (until explicitly stopped).
+    ends: Option<Instant>,
+}
+
 /// One FF effect a game uploaded: rumble magnitudes + playback state.
 struct Effect {
     strong: u16,
     weak: u16,
-    /// `Some(deadline)` while playing (replay length 0 = until stopped).
-    playing: Option<Option<Instant>>,
+    /// `Some(window)` while playing.
+    playing: Option<Playback>,
     replay_ms: u16,
+    /// `replay.delay` — how long after the play command the effect stays silent. Decoded from the
+    /// upload since forever and, until now, never acted on: the effect started immediately and
+    /// ended `replay.length` later, so anything scheduling a delayed effect (DirectInput under
+    /// Wine does this routinely) fired early AND finished early by the same amount.
+    delay_ms: u16,
+}
+
+impl Effect {
+    /// The window a play command at `at` opens: silent for `replay.delay`, then `replay.length` of
+    /// rumble (or until stopped, when the length is 0).
+    ///
+    /// `replay.length` is measured from the END of the delay, not from the play command, so the
+    /// delay shifts the whole window instead of eating into it. Split out from the `EV_FF` handler
+    /// purely so this is testable — the handler itself needs a live uinput fd.
+    fn window(&self, at: Instant) -> Playback {
+        let starts = at + Duration::from_millis(self.delay_ms as u64);
+        Playback {
+            starts,
+            ends: (self.replay_ms > 0)
+                .then(|| starts + Duration::from_millis(self.replay_ms as u64)),
+        }
+    }
 }
 
 /// The force-feedback half of a virtual pad — the game-side effect table plus the mixdown policy
@@ -299,17 +331,29 @@ impl FfState {
     /// Mix: sum playing effects (expiring finished ones, force-stopping abandoned infinite ones),
     /// scale by gain. Returns the new `(low, high)` only when it changed since the last call.
     fn mix(&mut self, now: Instant, idle: Option<Duration>) -> Option<(u16, u16)> {
-        let stale = idle.is_some_and(|t| now.duration_since(self.last_activity) >= t);
+        let quiet_since = |t: Instant| idle.is_some_and(|d| now.duration_since(t) >= d);
+        let plane_stale = quiet_since(self.last_activity);
         let (mut strong, mut weak) = (0u32, 0u32);
         for e in self.effects.values_mut() {
-            let Some(deadline) = e.playing else { continue };
-            match deadline {
+            let Some(p) = e.playing else { continue };
+            // Still inside `replay.delay`: armed, silent, and NOT a candidate for expiry or the
+            // abandoned-effect force-off — it has not had its turn yet.
+            if now < p.starts {
+                continue;
+            }
+            match p.ends {
                 Some(d) if now >= d => e.playing = None,
                 // An infinite-replay effect the game stopped driving (no FF traffic for the whole
                 // idle window) — the alive-but-abandoned case the kernel's close-time auto-erase
                 // cannot see. Stop it once; a later EV_FF play re-arms it (and refreshes the
                 // clock). Mirrors the XUSB/UHID abandoned-rumble force-off.
-                None if stale => {
+                //
+                // "Abandoned" needs the effect to have been AUDIBLE for the window too, not just
+                // the plane quiet: the play command is itself the last activity, so an effect with
+                // a `replay.delay` longer than the window would otherwise be force-stopped the
+                // instant it finally started — silent the whole time it waited, then killed on its
+                // first contributing tick.
+                None if plane_stale && quiet_since(p.starts) => {
                     tracing::info!(
                         strong = e.strong,
                         weak = e.weak,
@@ -544,10 +588,12 @@ impl VirtualPad {
                                 weak: 0,
                                 playing: None,
                                 replay_ms: 0,
+                                delay_ms: 0,
                             });
                             slot.strong = strong;
                             slot.weak = weak;
                             slot.replay_ms = e.replay_length;
+                            slot.delay_ms = e.replay_delay;
                         }
                         up.effect.id = e.id; // hand the assigned slot back to the kernel
                         up.retval = 0;
@@ -574,14 +620,7 @@ impl VirtualPad {
                 (EV_FF, code) => {
                     self.ff.note_activity();
                     if let Some(e) = self.ff.effects.get_mut(&(code as i16)) {
-                        e.playing = if ev.value != 0 {
-                            Some((e.replay_ms > 0).then(|| {
-                                Instant::now()
-                                    + std::time::Duration::from_millis(e.replay_ms as u64)
-                            }))
-                        } else {
-                            None
-                        };
+                        e.playing = (ev.value != 0).then(|| e.window(Instant::now()));
                     }
                 }
                 _ => {}
@@ -802,15 +841,34 @@ mod ff_state_tests {
         ff
     }
 
+    /// Playing from `at`, no delay, until explicitly stopped.
+    fn playing(at: Instant) -> Option<Playback> {
+        Some(Playback {
+            starts: at,
+            ends: None,
+        })
+    }
+
+    /// Playing from `at`, no delay, for `len`.
+    fn playing_for(at: Instant, len: Duration) -> Option<Playback> {
+        Some(Playback {
+            starts: at,
+            ends: Some(at + len),
+        })
+    }
+
     #[test]
     fn abandoned_infinite_effect_is_forced_off_after_idle_window() {
+        let now = Instant::now();
         let mut ff = ff_with(Effect {
             strong: 0x8000,
             weak: 0,
-            playing: Some(None),
+            // Playing since before the window: "abandoned" means audible AND unattended, so an
+            // effect that only just started is not a candidate however stale the plane is.
+            playing: playing(now - Duration::from_millis(2600)),
             replay_ms: 0,
+            delay_ms: 0,
         });
-        let now = Instant::now();
         assert_eq!(ff.mix(now, IDLE), Some((scaled(0x8000), 0)));
         assert_eq!(ff.mix(now, IDLE), None); // unchanged level dedups, still playing
                                              // The game goes silent on the FF plane past the idle window: cut, exactly once.
@@ -825,8 +883,9 @@ mod ff_state_tests {
         let mut ff = ff_with(Effect {
             strong: 0x4000,
             weak: 0,
-            playing: Some(Some(now + Duration::from_secs(10))),
+            playing: playing_for(now, Duration::from_secs(10)),
             replay_ms: 10_000,
+            delay_ms: 0,
         });
         // FF plane long stale, but the effect declared a finite replay — the declared duration is
         // the contract (a real pad honors it too), so it keeps playing…
@@ -842,26 +901,135 @@ mod ff_state_tests {
         let mut ff = ff_with(Effect {
             strong: 0x8000,
             weak: 0,
-            playing: Some(None),
+            playing: playing(now - Duration::from_millis(3000)),
             replay_ms: 0,
+            delay_ms: 0,
         });
         assert_eq!(ff.mix(now, IDLE), Some((scaled(0x8000), 0)));
         ff.last_activity = now - Duration::from_millis(3000);
         assert_eq!(ff.mix(now, IDLE), Some((0, 0)));
         // The game plays the effect again — an FF event refreshes the clock and re-arms playback.
         ff.last_activity = now;
-        ff.effects.get_mut(&0).unwrap().playing = Some(None);
+        ff.effects.get_mut(&0).unwrap().playing = playing(now);
         assert_eq!(ff.mix(now, IDLE), Some((scaled(0x8000), 0)));
     }
 
+    /// `replay.delay` shifts the whole window: silent until it elapses, then the FULL
+    /// `replay.length`. Before this the delay was decoded and dropped, so a delayed effect both
+    /// started early and finished early — DirectInput under Wine schedules these routinely.
+    #[test]
+    fn replay_delay_holds_the_effect_off_then_gives_it_its_full_length() {
+        let now = Instant::now();
+        let starts = now + Duration::from_millis(500);
+        let mut ff = ff_with(Effect {
+            strong: 0x8000,
+            weak: 0,
+            playing: Some(Playback {
+                starts,
+                ends: Some(starts + Duration::from_secs(1)),
+            }),
+            replay_ms: 1000,
+            delay_ms: 500,
+        });
+        // Inside the delay: armed but silent.
+        assert_eq!(ff.mix(now, IDLE), None);
+        assert_eq!(ff.mix(now + Duration::from_millis(499), IDLE), None);
+        // Delay elapsed: it plays.
+        assert_eq!(
+            ff.mix(now + Duration::from_millis(501), IDLE),
+            Some((scaled(0x8000), 0))
+        );
+        // Still playing at 1400 ms — it gets its full second FROM the delay, not from the play.
+        assert_eq!(ff.mix(now + Duration::from_millis(1400), IDLE), None);
+        // And ends at delay + length, not at length.
+        assert_eq!(
+            ff.mix(now + Duration::from_millis(1600), IDLE),
+            Some((0, 0))
+        );
+    }
+
+    /// The window a play opens, straight from the uploaded fields — this is the half that reads
+    /// `replay.delay` at all. Pinned separately because the `EV_FF` handler that calls it needs a
+    /// live uinput fd, so a test driving `mix` alone would pass with the delay ignored entirely.
+    #[test]
+    fn window_offsets_the_whole_playback_by_replay_delay() {
+        let at = Instant::now();
+
+        let delayed = Effect {
+            strong: 0,
+            weak: 0,
+            playing: None,
+            replay_ms: 1000,
+            delay_ms: 500,
+        };
+        let w = delayed.window(at);
+        assert_eq!(
+            w.starts,
+            at + Duration::from_millis(500),
+            "delay defers the start"
+        );
+        assert_eq!(
+            w.ends,
+            Some(at + Duration::from_millis(1500)),
+            "length runs from the END of the delay, so the effect keeps its full second"
+        );
+
+        // No delay: starts immediately, unchanged from before.
+        let plain = Effect {
+            strong: 0,
+            weak: 0,
+            playing: None,
+            replay_ms: 1000,
+            delay_ms: 0,
+        };
+        let w = plain.window(at);
+        assert_eq!(w.starts, at);
+        assert_eq!(w.ends, Some(at + Duration::from_millis(1000)));
+
+        // Length 0 = until stopped, but the delay still applies.
+        let infinite = Effect {
+            strong: 0,
+            weak: 0,
+            playing: None,
+            replay_ms: 0,
+            delay_ms: 250,
+        };
+        let w = infinite.window(at);
+        assert_eq!(w.starts, at + Duration::from_millis(250));
+        assert_eq!(w.ends, None);
+    }
+
+    /// A delayed effect must not be force-stopped as "abandoned" while it is still waiting: it has
+    /// not had its turn, and the idle window is shorter than a delay can legitimately be.
+    #[test]
+    fn a_waiting_effect_is_not_cut_by_the_idle_watchdog() {
+        let now = Instant::now();
+        let starts = now + Duration::from_secs(5);
+        let mut ff = ff_with(Effect {
+            strong: 0x8000,
+            weak: 0,
+            playing: Some(Playback { starts, ends: None }),
+            replay_ms: 0,
+            delay_ms: 5000,
+        });
+        ff.last_activity = now - Duration::from_secs(60); // long stale
+        assert_eq!(ff.mix(now, IDLE), None); // silent, but NOT cut
+                                             // It still plays when its delay elapses.
+        assert_eq!(
+            ff.mix(now + Duration::from_millis(5001), IDLE),
+            Some((scaled(0x8000), 0))
+        );
+    }
+
     #[test]
     fn disabled_watchdog_never_cuts() {
         let now = Instant::now();
         let mut ff = ff_with(Effect {
             strong: 0x8000,
             weak: 0,
-            playing: Some(None),
+            playing: playing(now),
             replay_ms: 0,
+            delay_ms: 0,
         });
         ff.last_activity = now - Duration::from_secs(600);
         assert_eq!(ff.mix(now, None), Some((scaled(0x8000), 0)));

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

@@ -250,11 +250,19 @@ impl DsState {
         use punktfunk_core::input::gamepad as gs;
         let to_u8 = |v: i16| (((v as i32) + 32768) >> 8) as u8;
         let on = |bit: u32| buttons & bit != 0;
+        // Invert in i16 space, BEFORE the quantisation, rather than as `255 - to_u8(v)`.
+        // 0..=255 has no exact midpoint: `to_u8` puts centre at 0x80, which leaves 128 codes below
+        // it and 127 above, so mirroring the *output* (`255 - 0x80` = 0x7F) lands a centred stick
+        // one LSB off the 0x80 that `DsState::neutral` — and the pad's own resting report — use.
+        // Games idle-poll a centred stick constantly, so that off-by-one showed up as a permanent
+        // sub-deadzone tilt on the Y axes only. Negating first maps centre to centre by
+        // construction and keeps both extremes exact (+32767 → 0, -32768 → 255); the only cost is
+        // that i16::MIN and -32767 share the 255 code, one LSB at the very end of the travel.
         let mut s = DsState {
             lx: to_u8(lx),
-            ly: 255 - to_u8(ly),
+            ly: to_u8(ly.saturating_neg()),
             rx: to_u8(rx),
-            ry: 255 - to_u8(ry),
+            ry: to_u8(ry.saturating_neg()),
             l2: lt,
             r2: rt,
             ..DsState::neutral()
@@ -783,6 +791,29 @@ mod tests {
         assert_eq!(r[53], 0x0A);
     }
 
+    /// A centred stick must encode as the pad's own neutral on BOTH axes. Inverting the quantised
+    /// byte (`255 - v`) put Y one LSB below it, which games idle-poll constantly — a permanent
+    /// sub-deadzone tilt. Extremes must stay exact either way.
+    #[test]
+    fn centred_sticks_encode_as_neutral_on_every_axis() {
+        let n = DsState::neutral();
+        let s = DsState::from_gamepad(0, 0, 0, 0, 0, 0, 0);
+        assert_eq!((s.lx, s.ly), (n.lx, n.ly), "left stick centre");
+        assert_eq!((s.rx, s.ry), (n.rx, n.ry), "right stick centre");
+
+        // Y is still inverted (XInput +y = up, DualSense 0 = up) and both ends stay exact.
+        let up = DsState::from_gamepad(0, 0, i16::MAX, 0, i16::MAX, 0, 0);
+        assert_eq!((up.ly, up.ry), (0, 0), "full up = 0");
+        let down = DsState::from_gamepad(0, 0, i16::MIN, 0, i16::MIN, 0, 0);
+        assert_eq!((down.ly, down.ry), (255, 255), "full down = 255");
+
+        // X keeps its existing mapping.
+        let right = DsState::from_gamepad(0, i16::MAX, 0, i16::MAX, 0, 0, 0);
+        assert_eq!((right.lx, right.rx), (255, 255));
+        let left = DsState::from_gamepad(0, i16::MIN, 0, i16::MIN, 0, 0, 0);
+        assert_eq!((left.lx, left.rx), (0, 0));
+    }
+
     /// The wire touchpad-click / guide / mute bits (Moonlight's extended positions) land in
     /// `buttons[2]`.
     #[test]

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

@@ -183,8 +183,9 @@ impl SteamState {
 
     /// Map an `XInput`/GameStream pad frame (button bitmask + i16 sticks + u8 triggers) into the Deck
     /// state. Sticks pass through (the kernel negates Y, which yields the conventional direction —
-    /// validated on-box); triggers scale u8 0..255 → u16 0..32640 and set the full-pull bit when
-    /// pressed. Trackpad + motion + the back grips arrive separately ([`apply_rich`], the M3 wire).
+    /// validated on-box); triggers scale u8 0..255 → u16 0..32767 ([`trigger_u16`]) and set the
+    /// full-pull bit when pressed. Trackpad + motion + the back grips arrive separately
+    /// ([`apply_rich`], the M3 wire).
     pub fn from_gamepad(
         buttons: u32,
         lx: i16,
@@ -200,8 +201,8 @@ impl SteamState {
             ly,
             rx,
             ry,
-            lt: (lt as u16) * 128,
-            rt: (rt as u16) * 128,
+            lt: trigger_u16(lt),
+            rt: trigger_u16(rt),
             ..SteamState::neutral()
         };
         let mut b = 0u64;
@@ -375,8 +376,8 @@ pub fn sc_from_gamepad(
         ly,
         rx: 0,
         ry: 0,
-        lt: (lt as u16) * 128,
-        rt: (rt as u16) * 128,
+        lt: trigger_u16(lt),
+        rt: trigger_u16(rt),
         // The wire right stick becomes a right-pad contact (see the doc above).
         rpad_x: rx,
         rpad_y: ry,
@@ -466,6 +467,18 @@ pub fn serialize_sc_state(r: &mut [u8; STEAM_REPORT_LEN], st: &SteamState, seq:
     r[38..40].copy_from_slice(&st.gyro[2].to_le_bytes());
 }
 
+/// Scale a wire trigger (u8 `0..=255`) onto the Deck's full axis (u16 `0..=32767`).
+///
+/// This was `v * 128`, which tops out at 32640 — a fully-pulled trigger reported 99.6% and the top
+/// 127 counts of the declared range were unreachable, so a game reading the axis could never see a
+/// true full pull. One multiply gets both ends exact (`0 → 0`, `255 → 32767`) and stays monotonic.
+///
+/// `serialize_report`'s inverse (`>> 7`, for the legacy u8 trigger bytes) still round-trips both
+/// ends against this: `32767 >> 7 == 255`.
+fn trigger_u16(v: u8) -> u16 {
+    ((v as u32 * 32767) / 255) as u16
+}
+
 /// Build the `steam_get_serial` GET_REPORT reply. The Steam feature path is report-id-0 with a
 /// leading report-id byte the kernel strips (`steam_recv_report` does `memcpy(data, buf+1, …)`), so
 /// the wire is `[0x00, 0xAE, len, 0x01, ascii…]`; the kernel then validates `reply[0]==0xAE`,
@@ -473,7 +486,12 @@ pub fn serialize_sc_state(r: &mut [u8; STEAM_REPORT_LEN], st: &SteamState, seq:
 pub fn serial_reply(serial: &str) -> [u8; STEAM_REPORT_LEN] {
     let mut buf = [0u8; STEAM_REPORT_LEN];
     let bytes = serial.as_bytes();
-    let len = bytes.len().clamp(1, 21);
+    // `min`, not `clamp(1, 21)`. Clamping the LOW end to 1 and then slicing `bytes[..len]` asks a
+    // zero-byte slice for one byte, which panics — on the service thread, for an input the kernel
+    // already has a graceful answer to. Reporting the true length lets its own validation
+    // (`1 <= reply[1] <= 21`) reject an empty serial and fall back to "XXXXXXXXXX", which is the
+    // documented behaviour for a reply it does not like.
+    let len = bytes.len().min(21);
     buf[0] = 0x00; // report id 0 — stripped by steam_recv_report
     buf[1] = ID_GET_STRING_ATTRIBUTE;
     buf[2] = len as u8;
@@ -704,7 +722,7 @@ mod tests {
         assert_ne!(s.buttons & btn::STEAM, 0);
         assert_ne!(s.buttons & btn::LB, 0);
         assert_ne!(s.buttons & btn::LT_FULL, 0); // lt=255 → full-pull bit
-        assert_eq!(s.lt, 255 * 128);
+        assert_eq!(s.lt, 32767); // full pull reaches the TOP of the declared range
         assert_eq!(s.lx, 1000);
         assert_eq!(s.ly, -2000);
 
@@ -730,6 +748,30 @@ mod tests {
         assert_eq!(s.accel, [16384, -8192, 0]);
     }
 
+    /// An empty serial must not panic. `clamp(1, 21)` asked a zero-byte slice for one byte, which
+    /// is an out-of-range slice index — on the service thread. The kernel rejects a zero length by
+    /// its own rule (`1 <= reply[1] <= 21`) and falls back, which is the graceful answer.
+    #[test]
+    fn empty_serial_reply_does_not_panic() {
+        let r = serial_reply("");
+        assert_eq!(r[1], ID_GET_STRING_ATTRIBUTE);
+        assert_eq!(
+            r[2], 0,
+            "length the kernel will reject, rather than a panic"
+        );
+
+        // Normal and over-long serials still behave.
+        let r = serial_reply("ABC123");
+        assert_eq!(r[2], 6);
+        assert_eq!(&r[4..10], b"ABC123");
+        let long = "X".repeat(40);
+        assert_eq!(
+            serial_reply(&long)[2],
+            21,
+            "clamped to the protocol maximum"
+        );
+    }
+
     /// M3: the wire back-button bits map to the four Deck grips + QAM, and `TouchpadEx` routes the
     /// left / right surfaces to the matching pad (x passes straight through; y flips from the
     /// wire's screen convention (+down) to the Deck's raw +up — the live-verified direction).

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

@@ -159,6 +159,22 @@ impl OverflowWarn {
 /// real firmware decays, and that re-assert is what keeps a legitimately-held long rumble alive
 /// here. The XUSB path shares this window via [`rumble_idle_timeout`] (every XUSB write IS a
 /// rumble write, so its any-activity keying is already rumble-keyed by construction).
+///
+/// KNOWN COST, deliberately accepted. That invariant only covers writers that re-assert. A game
+/// driving the pad through the kernel's *evdev* FF interface does not: `ff-memless` sends one
+/// output report when an effect starts and one when it stops, with nothing in between, so a finite
+/// effect longer than this window is cut in half here. The uinput path
+/// (`linux/gamepad.rs`) exempts exactly that case — but it can, because evdev FF hands it an
+/// explicit `replay.length`. Nothing equivalent reaches this layer: [`PadFeedback`] carries motor
+/// levels, and the protocols it speaks (DualSense / DS4 / Deck / Switch Pro) are all
+/// level-triggered with no duration field anywhere in a report. So the choice is between cutting a
+/// long finite effect and letting an abandoned residual drone forever, and the residual is the one
+/// with field evidence behind it (a stuck level resent every 500 ms for 5.5 minutes). Switch Pro is
+/// not affected either way — `hid-nintendo` re-sends rumble continuously, and a physical Pro's
+/// HD-rumble decays faster than this window regardless.
+///
+/// Do not "fix" this by widening or disabling the window without evidence about which failure real
+/// titles actually hit; the hatch below exists for exactly that experiment.
 const RUMBLE_IDLE_TIMEOUT: Duration = Duration::from_millis(2500);
 
 /// The abandoned-rumble force-off window, env-hatched: `PUNKTFUNK_RUMBLE_IDLE_MS` overrides

crates/pf-presenter/src/run.rs

@@ -466,13 +466,6 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
     #[cfg(windows)]
     crate::win32::set_app_user_model_id();
     sdl3::hint::set("SDL_JOYSTICK_THREAD", "1");
-    // Hold SDL's Valve HIDAPI drivers off BEFORE SDL_Init: the Deck driver clears the pad's
-    // digital mappings at *enumeration*, which is part of bringing the gamepad subsystem up, so a
-    // hint set after `sdl.gamepad()` — where this used to live, inside GamepadService::pumped —
-    // only detached a driver that had already killed the built-in trackpad-mouse system-wide. The
-    // symptom was the Deck losing its trackpad cursor at the start of every session until the
-    // firmware watchdog restored lizard mode. They are still enabled for an attached session.
-    pf_client_core::gamepad::preinit_disable_valve_hidapi();
     // A touchscreen (the Deck's glass) is forwarded as REAL touch passthrough below — so
     // suppress SDL's default synthesis of mouse events from touch. Left on, every touch
     // ALSO warps a synthetic mouse to the touch point, which under the stream's relative
@@ -1902,13 +1895,6 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
         }
     };
 
-    // Every exit from the loop above converges here, which is why the gamepad teardown belongs
-    // here and not on the individual `break`s. `gamepad.detach()` only queues the detach; the
-    // close — flush, host-side GamepadRemove, and the explicit rumble-stop backstop — runs when
-    // the pump drains it. Single mode broke out of the loop immediately after detaching and
-    // Event::Quit never detached at all, so both left forwarded pads unflushed and, if the game
-    // was rumbling at the time, still buzzing.
-    pump.shutdown();
     // Join the pump BEFORE the device-wide idle: its decode submissions on the shared
     // device would race vkDeviceWaitIdle otherwise.
     if let Some(st) = stream.take() {

crates/punktfunk-core/cbindgen.toml

@@ -48,6 +48,13 @@ exclude = ["MsghdrX", "recvmsg_x", "mmsghdr", "sendmmsg", "recvmmsg"]
 "AXIS_RT" = "PUNKTFUNK_AXIS_RT"
 "AUDIO_MAGIC" = "PUNKTFUNK_AUDIO_MAGIC"
 "RUMBLE_MAGIC" = "PUNKTFUNK_RUMBLE_MAGIC"
+"AUDIO_RED_MAGIC" = "PUNKTFUNK_AUDIO_RED_MAGIC"
+"AUDIO_RED_HEADER" = "PUNKTFUNK_AUDIO_RED_HEADER"
+# Same hazard as the BTN_* block above, one step worse: `FRAME_MS` and `SAMPLE_RATE_HZ` are
+# generic enough that an embedder is likely to have its own, and a clashing #define silently
+# takes the last definition rather than failing to compile.
+"FRAME_MS" = "PUNKTFUNK_AUDIO_FRAME_MS"
+"SAMPLE_RATE_HZ" = "PUNKTFUNK_AUDIO_SAMPLE_RATE_HZ"
 
 # QualifiedScreamingSnakeCase already qualifies each variant with the enum name
 # (PunktfunkStatus::Ok -> PUNKTFUNK_STATUS_OK); do NOT also set prefix_with_name or it doubles.

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

@@ -490,7 +490,13 @@ impl NativeClient {
                     video_codecs,
                     preferred_codec,
                     display_hdr,
-                    client_caps,
+                    // Redundant audio (`0xD2`) is advertised by CORE, not by the embedder: the
+                    // recovery happens on the demux side (`AudioRedRecovery` in the datagram
+                    // task) and re-inserts the rebuilt frame into the same queue, so every
+                    // embedder benefits without knowing the plane exists — and none of them can
+                    // forget to opt in. The bit is a pure "I can decode it"; the host still
+                    // decides whether to spend the extra ~1 %.
+                    client_caps: client_caps | crate::quic::CLIENT_CAP_AUDIO_RED,
                     frame_parts,
                     launch,
                     name,

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

@@ -23,6 +23,10 @@ pub(super) async fn run(
     // gate): a datagram the network reordered must not roll a stopped motor back on. Legacy v1
     // datagrams carry no seq and bypass it (an old host's own periodic re-send is the only heal).
     let mut rumble_last_seq: [Option<u8>; crate::input::MAX_PADS] = [None; crate::input::MAX_PADS];
+    // Redundant-audio-plane rebuild (`0xD2`). Recovery happens HERE rather than in the four
+    // client decoders: the recovered frame is re-inserted into this queue in order, so every
+    // embedder gets a complete stream without knowing the plane exists.
+    let mut audio_red = crate::audio::AudioRedRecovery::new();
     while let Ok(d) = conn.read_datagram().await {
         match d.first() {
             Some(&crate::quic::AUDIO_MAGIC) => {
@@ -34,6 +38,26 @@ pub(super) async fn run(
                     });
                 }
             }
+            Some(&crate::quic::AUDIO_RED_MAGIC) => {
+                if let Some((seq, pts_ns, opus, prev)) = crate::quic::decode_audio_red_datagram(&d)
+                {
+                    if audio_red.recover_before(seq, prev.is_some()) {
+                        // The copy is the frame BEFORE this one, so it carries the previous
+                        // sequence and presentation time — one protocol frame earlier.
+                        let _ = audio_tx.try_send(AudioPacket {
+                            seq: seq.wrapping_sub(1),
+                            pts_ns: pts_ns
+                                .saturating_sub(crate::audio::FRAME_MS as u64 * 1_000_000),
+                            data: prev.unwrap_or_default().to_vec(),
+                        });
+                    }
+                    let _ = audio_tx.try_send(AudioPacket {
+                        seq,
+                        pts_ns,
+                        data: opus.to_vec(),
+                    });
+                }
+            }
             Some(&crate::quic::RUMBLE_MAGIC) => {
                 if let Some(u) = crate::quic::decode_rumble_envelope(&d) {
                     // Gate v2 envelopes on their per-pad seq; forward v1 (envelope: None) as-is.

crates/punktfunk-core/src/config.rs

@@ -341,6 +341,50 @@ pub fn mtu1500_shard_payload_for(peer: core::net::IpAddr) -> usize {
     }
 }
 
+/// Floor for a negotiated `shard_payload` (even, well under every real path). A path whose UDP
+/// budget lands below this can't carry the QUIC control plane either (QUIC's own minimum is a
+/// 1200-byte UDP payload), so shrinking video shards further buys nothing — the clamp helpers
+/// bottom out here instead of producing degenerate confetti-sized shards.
+pub const MIN_SHARD_PAYLOAD: usize = 512;
+
+/// The sealed wire size of a video datagram carrying `shard_payload` bytes of shard — what
+/// actually leaves the socket as UDP payload (punktfunk header + shard + crypto overhead).
+pub const fn sealed_datagram_bytes(shard_payload: usize) -> usize {
+    HEADER_LEN + shard_payload + CRYPTO_OVERHEAD
+}
+
+/// The UDP-payload size a path must carry for full-size IPv4 video datagrams: the sealed size
+/// of the [`mtu1500_shard_payload`] default (= 1472, the exact 1500-MTU IPv4 ceiling). Doubles
+/// as the QUIC MTU-discovery probe ceiling (`quic/endpoint.rs`): with the ceiling set to
+/// exactly this value, a control connection whose discovery settles AT the ceiling has proven
+/// the path carries full-size video datagrams, and one that settles BELOW it has proven the
+/// path cannot — a discrimination quinn's stock 1452 ceiling can't make in either direction.
+pub const fn video_datagram_udp_ceiling() -> usize {
+    sealed_datagram_bytes(mtu1500_shard_payload())
+}
+
+/// Largest even shard payload whose sealed datagram fits in `udp_budget` bytes of UDP payload
+/// (the quantity QUIC MTU discovery measures — [`video_datagram_udp_ceiling`] is its probe
+/// ceiling). Clamped to the peer's family default ([`mtu1500_shard_payload_for`]) so a generous
+/// budget never grows packets past today's wire, and floored at [`MIN_SHARD_PAYLOAD`].
+pub fn shard_payload_for_udp_budget(udp_budget: usize, peer: core::net::IpAddr) -> usize {
+    let p = udp_budget.saturating_sub(HEADER_LEN + CRYPTO_OVERHEAD);
+    let p = p - p % 2; // FEC requires even shards
+    p.clamp(MIN_SHARD_PAYLOAD, mtu1500_shard_payload_for(peer))
+}
+
+/// [`shard_payload_for_udp_budget`] for an operator-supplied ON-WIRE IP MTU (the number
+/// `netsh interface ipv4 show subinterfaces` / `ip link` shows): subtracts the family's IP+UDP
+/// headers first — 28 for IPv4 (and IPv4-mapped), 48 for IPv6.
+pub fn shard_payload_for_wire_mtu(wire_mtu: usize, peer: core::net::IpAddr) -> usize {
+    let ip_udp = match peer {
+        core::net::IpAddr::V4(_) => 28,
+        core::net::IpAddr::V6(v6) if v6.to_ipv4_mapped().is_some() => 28,
+        core::net::IpAddr::V6(_) => 48,
+    };
+    shard_payload_for_udp_budget(wire_mtu.saturating_sub(ip_udp), peer)
+}
+
 /// Everything needed to construct a [`Session`](crate::session::Session).
 ///
 /// `Debug` is implemented by hand to redact `key`/`salt`, and `key`/`salt` are zeroized
@@ -514,6 +558,74 @@ mod tests {
         assert!(HEADER_LEN + (p + 2) + CRYPTO_OVERHEAD > 1452, "not maximal");
     }
 
+    /// The video-datagram ceiling IS the exact v4 sealed size — the QUIC MTU-discovery probe
+    /// ceiling (endpoint.rs) relies on this equality for its settled-at-vs-below verdict.
+    #[test]
+    fn video_datagram_ceiling_is_the_sealed_default() {
+        assert_eq!(
+            video_datagram_udp_ceiling(),
+            HEADER_LEN + mtu1500_shard_payload() + CRYPTO_OVERHEAD
+        );
+        assert_eq!(video_datagram_udp_ceiling(), 1472);
+    }
+
+    /// Budget-derived sizing: even, sealed-fits-the-budget, clamped to the family default
+    /// above and [`MIN_SHARD_PAYLOAD`] below.
+    #[test]
+    fn shard_payload_for_udp_budget_math() {
+        use core::net::IpAddr;
+        let v4: IpAddr = "192.168.1.50".parse().unwrap();
+        let v6: IpAddr = "fd00::50".parse().unwrap();
+        // The full ceiling reproduces the default exactly.
+        assert_eq!(
+            shard_payload_for_udp_budget(video_datagram_udp_ceiling(), v4),
+            mtu1500_shard_payload()
+        );
+        // A WARP/Tailscale-shaped 1280 budget: sealed result must fit the budget, stay even.
+        let p = shard_payload_for_udp_budget(1280, v4);
+        assert_eq!(p % 2, 0);
+        assert!(sealed_datagram_bytes(p) <= 1280);
+        assert!(sealed_datagram_bytes(p + 2) > 1280, "not maximal");
+        // Odd budgets round down to even shards.
+        assert_eq!(shard_payload_for_udp_budget(1281, v4) % 2, 0);
+        // A generous budget never grows past the family default (either family).
+        assert_eq!(
+            shard_payload_for_udp_budget(9000, v4),
+            mtu1500_shard_payload()
+        );
+        assert_eq!(
+            shard_payload_for_udp_budget(9000, v6),
+            mtu1500_shard_payload_v6()
+        );
+        // Degenerate budgets bottom out at the floor instead of confetti.
+        assert_eq!(shard_payload_for_udp_budget(100, v4), MIN_SHARD_PAYLOAD);
+    }
+
+    /// Operator-facing wire-MTU sizing subtracts the right IP+UDP header per family, and 1500
+    /// reproduces today's defaults exactly.
+    #[test]
+    fn shard_payload_for_wire_mtu_math() {
+        use core::net::IpAddr;
+        let v4: IpAddr = "192.168.1.50".parse().unwrap();
+        let v6: IpAddr = "fd00::50".parse().unwrap();
+        let mapped: IpAddr = "::ffff:192.168.1.50".parse().unwrap();
+        assert_eq!(
+            shard_payload_for_wire_mtu(1500, v4),
+            mtu1500_shard_payload()
+        );
+        assert_eq!(
+            shard_payload_for_wire_mtu(1500, mapped),
+            mtu1500_shard_payload()
+        );
+        assert_eq!(
+            shard_payload_for_wire_mtu(1500, v6),
+            mtu1500_shard_payload_v6()
+        );
+        // 1280 wire − 28 − 64 = 1188 (v4); − 48 − 64 = 1168 (v6).
+        assert_eq!(shard_payload_for_wire_mtu(1280, v4), 1188);
+        assert_eq!(shard_payload_for_wire_mtu(1280, v6), 1168);
+    }
+
     /// Family selection: genuine v6 remotes get the v6 size; v4 — including the IPv4-mapped v6
     /// form a dual-stack `[::]` socket reports for a v4 client — keeps the v4 size.
     #[test]

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

@@ -111,6 +111,17 @@ pub const CLIENT_CAP_CURSOR: u8 = 0x01;
 /// simply ignored — no behavior change in either direction.
 pub const CLIENT_CAP_PHASE_LOCK: u8 = 0x02;
 
+/// `Hello.client_caps` bit: this client can decode the redundant desktop-audio plane
+/// ([`AUDIO_RED_MAGIC`](super::datagram::AUDIO_RED_MAGIC), `0xD2`), where every datagram also
+/// carries a copy of the previous frame so a single lost packet is reconstructed instead of
+/// papered over with packet-loss concealment.
+///
+/// Active only when the host answers with [`HOST_CAP_AUDIO_RED`] (capable-and-agreed, the
+/// cursor/clipboard precedent). Toward an older host, or a host that declines because the link is
+/// 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;
+
 /// [`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,
 /// whose capture carries no cursor, and NOT Windows yet, where DWM composites into the IDD
@@ -132,6 +143,18 @@ pub const HOST_CAP_CURSOR: u8 = 0x08;
 /// [`HOST_CAP_TEXT_INPUT`], `0x01`/`0x02` are gamepad-state / clipboard.
 pub const HOST_CAP_PEN: u8 = 0x10;
 
+/// [`Welcome::host_caps`] bit: the host is sending the REDUNDANT desktop-audio plane
+/// ([`AUDIO_RED_MAGIC`](super::datagram::AUDIO_RED_MAGIC), `0xD2`) instead of plain `0xC9` — each
+/// datagram carries its own frame plus a copy of the previous one.
+///
+/// Set only when the client asked via [`CLIENT_CAP_AUDIO_RED`]. It is a statement about the WIRE,
+/// not a negotiation the client can decline: with the bit set the client must decode `0xD2`, and
+/// without it `0xC9`. The host may also drop back to `0xC9` mid-session (the redundancy is
+/// loss-gated — a clean LAN shouldn't pay for it), which is why clients decode BOTH tags
+/// 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;
+
 /// [`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
 /// advertise this.

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

@@ -42,6 +42,80 @@ pub fn decode_audio_datagram(b: &[u8]) -> Option<(u32, u64, &[u8])> {
     Some((seq, pts_ns, &b[13..]))
 }
 
+/// Redundant audio datagram, host → client: the [`AUDIO_MAGIC`] plane plus a copy of the PREVIOUS
+/// frame, so a single lost datagram is *reconstructed* rather than concealed.
+///
+/// `[0xD2][u32 seq LE][u64 pts_ns LE][u16 primary_len LE][primary opus][previous opus]`
+///
+/// **Why this and not Opus in-band FEC.** LBRR is a SILK-layer feature: the desktop-audio encoder
+/// runs `RESTRICTED_LOWDELAY` (CELT-only) at 5 ms frames, which is below SILK's 10 ms minimum, so
+/// `set_inband_fec(true)` on that encoder is a no-op. Nothing in libopus can protect this plane —
+/// the redundancy has to be at the application layer. (The mic uplink is a different encoder, VoIP
+/// mode at 10 ms, and *does* use real in-band FEC.)
+///
+/// **Why it costs no latency.** The copy rides the SUCCESSOR of the frame it protects, and the
+/// client is already holding 15–90 ms of de-jitter buffer — far more than the 5 ms the successor
+/// takes to arrive. So the recovery happens inside slack that already exists.
+///
+/// The previous frame's sequence is implicitly `seq - 1`; a host with nothing to duplicate yet
+/// (the first frame of a session, or straight after a capture reopen) simply sends an empty tail,
+/// which decodes to `None`.
+///
+/// Sent ONLY when the client advertised [`CLIENT_CAP_AUDIO_RED`](super::caps::CLIENT_CAP_AUDIO_RED)
+/// and the host answered [`HOST_CAP_AUDIO_RED`](super::caps::HOST_CAP_AUDIO_RED) — the
+/// capable-and-agreed handshake the cursor and 4:4:4 planes already use. Every other session keeps
+/// the plain [`AUDIO_MAGIC`] wire byte-for-byte.
+///
+/// NB `0xD1` is deliberately skipped: the DualSense pad-audio program has reserved it for the
+/// per-pad audio plane.
+pub const AUDIO_RED_MAGIC: u8 = 0xD2;
+
+/// Fixed header length of an [`AUDIO_RED_MAGIC`] datagram (tag + seq + pts + primary length).
+pub const AUDIO_RED_HEADER: usize = 1 + 4 + 8 + 2;
+
+/// Encode a redundant audio datagram. `prev` is the immediately-preceding frame's Opus payload
+/// (empty when there is none yet).
+pub fn encode_audio_red_datagram(seq: u32, pts_ns: u64, opus: &[u8], prev: &[u8]) -> Vec<u8> {
+    let mut b = Vec::with_capacity(AUDIO_RED_HEADER + opus.len() + prev.len());
+    b.push(AUDIO_RED_MAGIC);
+    b.extend_from_slice(&seq.to_le_bytes());
+    b.extend_from_slice(&pts_ns.to_le_bytes());
+    // A frame longer than u16::MAX cannot occur (5 ms of Opus is tens of bytes; the buffer the
+    // encoder writes into is 4 KiB) — but truncating silently would desync the split, so clamp
+    // the redundancy off instead of the primary.
+    let primary_len = u16::try_from(opus.len()).unwrap_or(u16::MAX);
+    b.extend_from_slice(&primary_len.to_le_bytes());
+    b.extend_from_slice(opus);
+    if opus.len() == primary_len as usize {
+        b.extend_from_slice(prev);
+    }
+    b
+}
+
+/// Parse a redundant audio datagram → `(seq, pts_ns, primary, previous)`. `previous` is `None`
+/// when the host had nothing to duplicate. `None` overall on bad tag/length, including a
+/// `primary_len` that overruns the datagram (a truncated or hostile packet must not panic).
+///
+/// The tuple shape deliberately mirrors [`decode_audio_datagram`] (one extra slot for the
+/// redundant copy) so the two planes read the same at every call site; a named struct here would
+/// be the odd one out on this module's decode surface, and cbindgen would then have to be taught
+/// to skip it.
+#[allow(clippy::type_complexity)]
+pub fn decode_audio_red_datagram(b: &[u8]) -> Option<(u32, u64, &[u8], Option<&[u8]>)> {
+    if b.len() < AUDIO_RED_HEADER || b[0] != AUDIO_RED_MAGIC {
+        return None;
+    }
+    let seq = u32::from_le_bytes(b[1..5].try_into().unwrap());
+    let pts_ns = u64::from_le_bytes(b[5..13].try_into().unwrap());
+    let primary_len = u16::from_le_bytes(b[13..15].try_into().unwrap()) as usize;
+    let rest = &b[AUDIO_RED_HEADER..];
+    if primary_len > rest.len() {
+        return None; // truncated: the split point is outside the datagram
+    }
+    let (primary, prev) = rest.split_at(primary_len);
+    Some((seq, pts_ns, primary, (!prev.is_empty()).then_some(prev)))
+}
+
 /// Legacy rumble datagram (v1), host → client: `[0xCA][u16 pad LE][u16 low LE][u16 high LE]`.
 /// Force-feedback state for pad `pad` (0xFFFF amplitudes, 0/0 = stop) as *level-triggered* state
 /// — it persists until superseded, which is why the host re-sends it periodically as its loss
@@ -806,6 +880,8 @@ mod tests {
     #[test]
     fn audio_datagram_roundtrip() {
         let opus = [0x42u8; 97];
+        let d = encode_audio_red_datagram(7, 42, &opus, &[]);
+        assert_eq!(d[0], AUDIO_RED_MAGIC);
         let d = encode_audio_datagram(7, 1_000_000_123, &opus);
         assert_eq!(d[0], AUDIO_MAGIC);
         let (seq, pts, payload) = decode_audio_datagram(&d).unwrap();
@@ -820,6 +896,83 @@ mod tests {
         assert!(empty.is_empty());
     }
 
+    #[test]
+    fn audio_red_datagram_roundtrip() {
+        let cur = [0x42u8; 97];
+        let prev = [0x37u8; 88];
+        let d = encode_audio_red_datagram(7, 1_000_000_123, &cur, &prev);
+        assert_eq!(d[0], AUDIO_RED_MAGIC);
+        let (seq, pts, primary, previous) = decode_audio_red_datagram(&d).unwrap();
+        assert_eq!((seq, pts), (7, 1_000_000_123));
+        assert_eq!(primary, cur);
+        assert_eq!(previous, Some(&prev[..]));
+
+        // No predecessor yet (first frame of a session / after a capture reopen).
+        let d = encode_audio_red_datagram(0, 5, &cur, &[]);
+        let (_, _, primary, previous) = decode_audio_red_datagram(&d).unwrap();
+        assert_eq!(primary, cur);
+        assert_eq!(
+            previous, None,
+            "an empty tail must decode as absent, not as a zero-length frame"
+        );
+
+        // Frames of equal length must still split at the right place — the length prefix is the
+        // only thing that can tell them apart.
+        let a = [1u8; 64];
+        let b = [2u8; 64];
+        let d = encode_audio_red_datagram(9, 0, &a, &b);
+        let (_, _, primary, previous) = decode_audio_red_datagram(&d).unwrap();
+        assert_eq!(primary, a);
+        assert_eq!(previous, Some(&b[..]));
+    }
+
+    /// A truncated or hostile `0xD2` must be rejected, never panic — the split point comes off
+    /// the wire, so an over-long `primary_len` is the obvious attack on `split_at`.
+    #[test]
+    fn audio_red_datagram_rejects_bad_input() {
+        let d = encode_audio_red_datagram(1, 2, &[0xAAu8; 30], &[0xBBu8; 20]);
+        for n in 0..AUDIO_RED_HEADER {
+            assert!(decode_audio_red_datagram(&d[..n]).is_none(), "len {n}");
+        }
+        // primary_len larger than the datagram: must be refused, not sliced.
+        let mut bad = d.clone();
+        bad[13..15].copy_from_slice(&u16::MAX.to_le_bytes());
+        assert!(decode_audio_red_datagram(&bad).is_none());
+        // Wrong tag.
+        let mut wrong = d.clone();
+        wrong[0] = AUDIO_MAGIC;
+        assert!(decode_audio_red_datagram(&wrong).is_none());
+    }
+
+    /// The two audio planes must not alias each other or any neighbouring plane: a client
+    /// demultiplexes purely on the first byte.
+    #[test]
+    fn audio_red_tag_is_disjoint() {
+        for other in [
+            AUDIO_MAGIC,
+            RUMBLE_MAGIC,
+            MIC_MAGIC,
+            RICH_INPUT_MAGIC,
+            HIDOUT_MAGIC,
+            HDR_META_MAGIC,
+            HOST_TIMING_MAGIC,
+            CURSOR_STATE_MAGIC,
+            crate::input::INPUT_MAGIC,
+        ] {
+            assert_ne!(AUDIO_RED_MAGIC, other);
+        }
+        let red = encode_audio_red_datagram(1, 2, &[9u8; 40], &[8u8; 40]);
+        assert!(
+            decode_audio_datagram(&red).is_none(),
+            "0xC9 must not accept a 0xD2"
+        );
+        let plain = encode_audio_datagram(1, 2, &[9u8; 40]);
+        assert!(
+            decode_audio_red_datagram(&plain).is_none(),
+            "0xD2 must not accept a 0xC9"
+        );
+    }
+
     #[test]
     fn rumble_datagram_roundtrip() {
         let d = encode_rumble_datagram(1, 0x1234, 0xFFFF);

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

@@ -47,6 +47,20 @@ fn stream_transport_idle(idle: std::time::Duration) -> Arc<quinn::TransportConfi
     // plane latest-wins at the source — ~200 ms of stereo Opus (proportionally less at
     // surround bitrates), so sustained congestion costs concealable drops, never lag.
     t.datagram_send_buffer_size(4 * 1024);
+    // MTU discovery probes up to EXACTLY the sealed size of a full IPv4 video datagram (1472)
+    // instead of quinn's stock 1452. Two reasons: (a) on a clean 1500-MTU path QUIC gets the
+    // last 20 bytes per packet; (b) the ceiling turns discovery into a video-path verdict the
+    // host's wire-MTU watcher reads (`punktfunk-host` `native/wire_mtu.rs`) — settled == ceiling
+    // proves the path carries full-size video datagrams, settled BELOW it proves it cannot (a
+    // VPN/overlay adapter at MTU ~1280 blackholes every video packet while all the small flows
+    // pass: the "connects fine, black screen forever" field shape). With the stock 1452 ceiling
+    // a healthy path and a constrained one are indistinguishable at the top. This is the ONLY
+    // behavioral change on healthy paths, and it's confined to discovery: probes are padded
+    // PINGs quinn already expects to lose above a constrained hop — a lost probe settles the
+    // search lower, exactly as it did before.
+    let mut mtud = quinn::MtuDiscoveryConfig::default();
+    mtud.upper_bound(crate::config::video_datagram_udp_ceiling() as u16);
+    t.mtu_discovery_config(Some(mtud));
     Arc::new(t)
 }
 

crates/punktfunk-host/src/audio.rs

@@ -192,6 +192,11 @@ mod wasapi_mic;
 #[cfg_attr(not(target_os = "windows"), allow(dead_code))]
 #[path = "audio/wiring_plan.rs"]
 pub(crate) mod wiring_plan;
+// Pure capture-loop policy, split out for the same reason `wiring_plan` is: it encodes field
+// behaviour, so its tests must run on every platform's CI, not only Windows.
+#[cfg_attr(not(target_os = "windows"), allow(dead_code))]
+#[path = "audio/capture_policy.rs"]
+pub(crate) mod capture_policy;
 
 mod mic_jitter;
 mod mic_pump;

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

@@ -0,0 +1,260 @@
+//! Desktop-audio capture POLICY — the parts of [`wasapi_cap`](super::wasapi_cap) that are pure
+//! decisions rather than WASAPI plumbing, split out for the same reason
+//! [`wiring_plan`](super::wiring_plan) is: so they compile and their unit tests RUN on every
+//! platform. Both of these encode field-report behaviour, and regressing either must fail CI on
+//! Linux too, not only on a Windows box.
+//!
+//! * [`FightDamper`] — how hard to fight another program for the default playback device.
+//! * [`CaptureStats`] — the audio plane's vitals, so a log can tell a quiet host from a broken
+//!   endpoint from one we are damaging ourselves.
+
+use std::time::{Duration, Instant};
+
+/// Default-playback re-assertions inside [`FIGHT_WINDOW`] before we stop fighting.
+pub(crate) const FIGHT_LIMIT: u32 = 4;
+pub(crate) const FIGHT_WINDOW: Duration = Duration::from_secs(20);
+/// How long to leave the default alone once another program has proven it will take it back.
+pub(crate) const FIGHT_BACKOFF: Duration = Duration::from_secs(60);
+
+/// Damping for the default-playback tug-of-war (WP2.4).
+///
+/// The 2026-08-03 field log recorded seven full re-assert cycles in sixteen seconds — something on
+/// that box re-set the default playback to CABLE Input every ~4 s and we snapped it back every
+/// time, each round a capture teardown plus a wiring pass with `IPolicyConfig` writes. Winning that
+/// argument is not possible and every round was an audible dropout, so: re-assert a few times
+/// (transient churn does settle), then concede for a minute and say so once.
+///
+/// Time is passed IN rather than read here, which keeps the policy pure and testable.
+pub(crate) struct FightDamper {
+    /// Re-assertions in the current window, and when the window opened.
+    count: u32,
+    window_started: Instant,
+    /// Set while we are deliberately not fighting.
+    paused_until: Option<Instant>,
+    /// One warning per fight burst, and one per concession.
+    warned_fighting: bool,
+    warned_giving_up: bool,
+    now: Instant,
+}
+
+impl FightDamper {
+    pub(crate) fn new(now: Instant) -> FightDamper {
+        FightDamper {
+            count: 0,
+            window_started: now,
+            paused_until: None,
+            warned_fighting: false,
+            warned_giving_up: false,
+            now,
+        }
+    }
+
+    /// A dud default-device change was observed at `now`.
+    pub(crate) fn observed_at(&mut self, now: Instant) {
+        self.now = now;
+        if now.duration_since(self.window_started) >= FIGHT_WINDOW {
+            self.window_started = now;
+            self.count = 0;
+            self.warned_fighting = false;
+        }
+        if self.paused_until.is_some_and(|t| now >= t) {
+            self.paused_until = None;
+            self.warned_giving_up = false;
+            self.count = 0;
+            self.window_started = now;
+        }
+    }
+
+    /// Should we put the default back? False while paused, or once this window's budget is spent.
+    pub(crate) fn should_reassert(&mut self) -> bool {
+        if self.paused_until.is_some() {
+            return false;
+        }
+        if self.count >= FIGHT_LIMIT {
+            self.paused_until = Some(self.now + FIGHT_BACKOFF);
+            return false;
+        }
+        self.count += 1;
+        true
+    }
+
+    /// Warn on the FIRST re-assert of a burst only (the rest are noise).
+    pub(crate) fn warn_now(&mut self) -> bool {
+        !std::mem::replace(&mut self.warned_fighting, true)
+    }
+
+    /// Warn once when we concede.
+    pub(crate) fn warn_giving_up(&mut self) -> bool {
+        self.paused_until.is_some() && !std::mem::replace(&mut self.warned_giving_up, true)
+    }
+
+    /// Currently conceding (test/diagnostic accessor).
+    pub(crate) fn is_paused(&self) -> bool {
+        self.paused_until.is_some()
+    }
+}
+
+/// How often the capture loop reports its vitals (WP0.2).
+pub(crate) const STATS_EVERY: Duration = Duration::from_secs(30);
+
+/// One reporting window's worth of capture vitals.
+///
+/// The point is to make three states that used to look identical in a log tell themselves apart: a
+/// genuinely quiet host (`peak` ~0, no drops), a working stream (`peak` > 0), and a stream we are
+/// damaging ourselves (`dropped_chunks` > 0). The 2026-08-03 field log — 3,600 lines, filed over an
+/// audio-quality complaint — could distinguish none of them, because the audio plane logged nothing
+/// at all between "capturing" and the session ending.
+#[derive(Default)]
+pub(crate) struct CaptureStats {
+    pub(crate) frames: u64,
+    /// Interleaved SAMPLES seen — the RMS denominator. Deliberately separate from `frames`:
+    /// dividing the sum of squares by the frame count instead inflates RMS by sqrt(channels),
+    /// which made a sine report an RMS equal to its own peak.
+    pub(crate) samples: u64,
+    /// Loudest |sample| in the window — tells a silent endpoint from a working one.
+    pub(crate) peak: f32,
+    /// Sum of squares, for the window's RMS: a level far below peak means a badly attenuated
+    /// endpoint (a parked device sitting at 20 % volume costs ~14 dB before Opus ever sees it).
+    pub(crate) sumsq: f64,
+    /// Chunks the encode thread was too slow to take. Silent data loss, previously uncounted:
+    /// the encoder simply concatenates across the hole, so it is a click AND a permanent shift of
+    /// everything after it.
+    pub(crate) dropped_chunks: u64,
+}
+
+impl CaptureStats {
+    pub(crate) fn observe(&mut self, samples: &[f32], channels: u32) {
+        self.frames += (samples.len() / channels.max(1) as usize) as u64;
+        self.samples += samples.len() as u64;
+        for &s in samples {
+            let a = s.abs();
+            if a > self.peak {
+                self.peak = a;
+            }
+            self.sumsq += (s as f64) * (s as f64);
+        }
+    }
+
+    /// `(peak dBFS, rms dBFS, delivered %)` for this window. Silence reports -120 dB rather than
+    /// -inf so the log line stays parseable.
+    pub(crate) fn summary(&self, elapsed: Duration, sample_rate: u32) -> (f64, f64, f64) {
+        let rms = (self.sumsq / (self.samples as f64).max(1.0)).sqrt();
+        let db = |v: f64| if v > 0.0 { 20.0 * v.log10() } else { -120.0 };
+        // Expected frames for the window — a shortfall means the endpoint is not delivering at
+        // real time (a stalling virtual device), which a peak/RMS alone cannot show.
+        let expected = elapsed.as_secs_f64() * sample_rate as f64;
+        (
+            db(self.peak as f64),
+            db(rms),
+            (self.frames as f64 / expected.max(1.0)) * 100.0,
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Replays the 2026-08-03 field shape: a dud default change every ~2 s, forever. We must put
+    /// the default back a few times, then concede — and warn exactly once for each.
+    #[test]
+    fn fight_damper_concedes_instead_of_looping_forever() {
+        let t0 = Instant::now();
+        let mut d = FightDamper::new(t0);
+        let mut reasserts = 0;
+        let (mut warns_fighting, mut warns_giving_up) = (0, 0);
+        for i in 0..8 {
+            d.observed_at(t0 + Duration::from_millis(i * 2_000));
+            if d.should_reassert() {
+                reasserts += 1;
+                if d.warn_now() {
+                    warns_fighting += 1;
+                }
+            } else if d.warn_giving_up() {
+                warns_giving_up += 1;
+            }
+        }
+        assert_eq!(
+            reasserts, FIGHT_LIMIT,
+            "must stop after the window's budget"
+        );
+        assert_eq!(warns_fighting, 1, "one warning per burst, not one per flip");
+        assert_eq!(warns_giving_up, 1, "concede exactly once");
+    }
+
+    /// Occasional, genuinely transient churn must ALWAYS be corrected — the damper must not
+    /// accumulate across widely-spaced events and quietly stop doing its job.
+    #[test]
+    fn fight_damper_always_fixes_isolated_changes() {
+        let t0 = Instant::now();
+        let mut d = FightDamper::new(t0);
+        let mut reasserts = 0;
+        for i in 1..=10 {
+            d.observed_at(t0 + FIGHT_WINDOW * i);
+            if d.should_reassert() {
+                reasserts += 1;
+            }
+        }
+        assert_eq!(reasserts, 10, "isolated changes must always be corrected");
+    }
+
+    /// After the backoff expires the damper re-arms, so a program that goes quiet and comes back
+    /// later is fought again rather than being conceded to for the rest of the session.
+    #[test]
+    fn fight_damper_rearms_after_the_backoff() {
+        let t0 = Instant::now();
+        let mut d = FightDamper::new(t0);
+        for i in 0..FIGHT_LIMIT + 2 {
+            d.observed_at(t0 + Duration::from_millis(i as u64 * 500));
+            d.should_reassert();
+        }
+        assert!(d.is_paused(), "should have conceded");
+        d.observed_at(t0 + FIGHT_BACKOFF + FIGHT_WINDOW * 2);
+        assert!(d.should_reassert(), "must re-arm once the backoff expires");
+    }
+
+    /// Peak/RMS must separate the states a log could not previously tell apart.
+    #[test]
+    fn capture_stats_separate_silence_from_signal() {
+        let mut quiet = CaptureStats::default();
+        quiet.observe(&[0.0; 480], 2);
+        let (peak, rms, _) = quiet.summary(Duration::from_secs(1), 48_000);
+        assert_eq!(peak, -120.0, "digital silence reports the floor, not -inf");
+        assert_eq!(rms, -120.0);
+
+        let mut loud = CaptureStats::default();
+        let tone: Vec<f32> = (0..480).map(|i| (i as f32 * 0.13).sin() * 0.5).collect();
+        loud.observe(&tone, 2);
+        assert_eq!(
+            loud.frames, 240,
+            "480 interleaved stereo samples = 240 frames"
+        );
+        let (peak, rms, _) = loud.summary(Duration::from_secs(1), 48_000);
+        assert!(
+            peak > -8.0 && peak <= 0.0,
+            "peak {peak} dBFS should track a 0.5 tone"
+        );
+        // A sine's RMS is its amplitude / sqrt(2) — about 3 dB below peak. Getting this equal to
+        // peak is exactly what a frames-vs-samples mix-up in the denominator looks like, so the
+        // margin is asserted rather than just the ordering.
+        assert!(
+            rms < peak - 2.0,
+            "RMS {rms} vs peak {peak}: a sine must sit ~3 dB below its peak"
+        );
+    }
+
+    /// The delivered-percentage is what shows an endpoint that has stopped feeding us in real
+    /// time — invisible in peak/RMS, and the shape a stalling virtual device makes.
+    #[test]
+    fn capture_stats_report_a_delivery_shortfall() {
+        let mut full = CaptureStats::default();
+        full.observe(&vec![0.1f32; 48_000 * 2], 2); // exactly 1 s of stereo
+        let (_, _, pct) = full.summary(Duration::from_secs(1), 48_000);
+        assert!((pct - 100.0).abs() < 1.0, "expected ~100 %, got {pct}");
+
+        let mut half = CaptureStats::default();
+        half.observe(&vec![0.1f32; 48_000], 2); // 0.5 s of stereo in a 1 s window
+        let (_, _, pct) = half.summary(Duration::from_secs(1), 48_000);
+        assert!((pct - 50.0).abs() < 1.0, "expected ~50 %, got {pct}");
+    }
+}

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

@@ -674,6 +674,8 @@ fn pw_thread(
             })
             .register();
 
+        // Which source the negotiated format below actually describes — see the note there.
+        let sink_mode = sink_name.is_some();
         let props = match &sink_name {
             // Stream-sink mode: this stream IS the sink (media.class + Direction::Input). Apps
             // play into it, PipeWire mixes them, process() receives the mix. Mirrors the
@@ -710,8 +712,25 @@ fn pw_thread(
         let stream = pw::stream::StreamBox::new(&core, "punktfunk-audio", props)
             .context("pw audio Stream")?;
 
+        // The capture callback's state: the hand-off channel plus this plane's vitals. Before
+        // this it was the bare `tx`, and the desktop-audio plane logged NOTHING between "capture
+        // started" and the session ending — no level, no cadence, and in particular no sign of
+        // the silent drop below. That is exactly what made the 2026-08-03 Windows field report
+        // un-triageable, and the Linux half kept it after the Windows half was fixed.
+        struct CapUd {
+            tx: std::sync::mpsc::SyncSender<Vec<f32>>,
+            channels: u32,
+            stats: crate::audio::capture_policy::CaptureStats,
+            last_stats: std::time::Instant,
+        }
+        let ud = CapUd {
+            tx,
+            channels,
+            stats: Default::default(),
+            last_stats: std::time::Instant::now(),
+        };
         let _listener = stream
-            .add_local_listener_with_user_data(tx)
+            .add_local_listener_with_user_data(ud)
             .state_changed({
                 let mainloop = mainloop.clone();
                 move |_s, _ud, old, new| {
@@ -723,22 +742,32 @@ fn pw_thread(
                     }
                 }
             })
-            .param_changed(|_stream, _tx, id, param| {
+            .param_changed(move |_stream, _tx, id, param| {
                 let Some(param) = param else { return };
                 if id != pw::spa::param::ParamType::Format.as_raw() {
                     return;
                 }
                 let mut info = AudioInfoRaw::default();
                 if info.parse(param).is_ok() {
+                    // `stream_sink` says WHICH source this format describes, and that changes how
+                    // much it is worth. In stream-sink mode the host owns the sink, so this IS the
+                    // format apps render into and the desktop mix cannot have been narrowed before
+                    // we saw it. In LEGACY monitor mode we are capturing someone else's sink
+                    // through PipeWire's resampler: a 16 kHz Bluetooth headset upstream would
+                    // still be reported here as a clean 48 kHz, exactly the way WASAPI's
+                    // autoconvert hid the same thing on Windows (the 2026-08-03 report). Reading
+                    // the monitored node's OWN rate needs a registry lookup this stream does not
+                    // do — recorded as an open gap rather than implied to be covered.
                     tracing::info!(
                         format = ?info.format(),
                         rate = info.rate(),
                         channels = info.channels(),
+                        stream_sink = sink_mode,
                         "audio format negotiated"
                     );
                 }
             })
-            .process(|stream, tx| {
+            .process(|stream, ud| {
                 let outcome = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
                     let Some(mut buffer) = stream.dequeue_buffer() else {
                         return;
@@ -774,7 +803,34 @@ fn pw_thread(
                         ];
                         samples.push(f32::from_le_bytes(b));
                     }
-                    let _ = tx.try_send(samples); // drop if the encoder is behind
+                    ud.stats.observe(&samples, ud.channels);
+                    // Non-blocking and lossy, as before — but COUNTED. A full channel means the
+                    // encode thread is not keeping up, and because the encoder simply
+                    // concatenates across the hole every dropped chunk is a click AND a
+                    // permanent shift of everything after it.
+                    if ud.tx.try_send(samples).is_err() {
+                        ud.stats.dropped_chunks += 1;
+                    }
+                    if ud.last_stats.elapsed() >= crate::audio::capture_policy::STATS_EVERY {
+                        let (peak_db, rms_db, delivered_pct) =
+                            ud.stats.summary(ud.last_stats.elapsed(), SAMPLE_RATE);
+                        if ud.stats.dropped_chunks > 0 {
+                            tracing::warn!(
+                                dropped_chunks = ud.stats.dropped_chunks,
+                                "the audio encode thread could not keep up — captured audio was \
+                                 DROPPED; the stream will click and everything after it shifts"
+                            );
+                        }
+                        tracing::info!(
+                            peak_db = format!("{peak_db:.1}"),
+                            rms_db = format!("{rms_db:.1}"),
+                            delivered_pct = format!("{delivered_pct:.0}"),
+                            dropped_chunks = ud.stats.dropped_chunks,
+                            "desktop audio capture"
+                        );
+                        ud.stats = Default::default();
+                        ud.last_stats = std::time::Instant::now();
+                    }
                 }));
                 if outcome.is_err() {
                     tracing::error!("panic in pipewire audio callback — chunk dropped");

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

@@ -14,8 +14,11 @@
 //! * default **PLAYBACK** → the plan's loopback endpoint, applied ONLY while a desktop-audio capture
 //!   is open (`set_playback` — the mic pump must never park the playback default while the host is
 //!   idle). By default that endpoint is the SILENT sink (Steam Streaming Microphone render side) so
-//!   audio plays on the client only; `PUNKTFUNK_HOST_AUDIO` prefers real hardware instead (audible on
-//!   both ends). **Never** the Steam Streaming Speakers, whose loopback is silent — validated live;
+//!   audio plays on the client only; `audio.output_mode = host_and_client` (formerly
+//!   `PUNKTFUNK_HOST_AUDIO`) prefers real hardware instead (audible on both ends). Since 2026-08 a
+//!   silent sink must also be able to CARRY the mix — one that narrows it (a voice-carrier endpoint
+//!   mixing mono or at 24 kHz) loses to real hardware; see [`super::wiring_plan`]. **Never** the
+//!   Steam Streaming Speakers, whose loopback is silent — validated live;
 //! * default **RECORDING** → the mic target's capture endpoint (VB-Cable "CABLE Output") so host apps
 //!   record the client's mic by default.
 //!
@@ -33,18 +36,44 @@
 //!
 //! Setting a default endpoint uses the undocumented `IPolicyConfig` COM interface (the only way to set
 //! a default device programmatically — neither the `windows` nor `wasapi` crate exposes it; it is the
-//! same call `mmsys.cpl` makes). Opt out with `PUNKTFUNK_KEEP_DEFAULT` to leave the user's chosen
-//! defaults untouched (the plan is still computed — the mic must still pick a target).
+//! same call `mmsys.cpl` makes). The `audio.output_mode = follow_default` setting (formerly
+//! `PUNKTFUNK_KEEP_DEFAULT`) leaves the user's chosen defaults untouched — the plan is still
+//! computed, since the mic must still pick a target.
 
 // Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it.
 #![deny(clippy::undocumented_unsafe_blocks)]
 
-use super::wiring_plan::{self, plan, Endpoint, Wiring};
+use super::wiring_plan::{self, plan, plan_with_formats, Endpoint, MixFormat, Wiring};
 use anyhow::{anyhow, bail, Result};
 use std::ffi::c_void;
 use std::sync::Mutex;
 use wasapi::Direction;
 
+/// A render endpoint's engine mix format, or `None` if it cannot be asked right now.
+///
+/// This is the number the 2026-08-03 field report needed and no log had: the capture side requests
+/// 48 kHz f32 with `autoconvert`, so WASAPI converts silently from whatever the endpoint really
+/// runs — and a voice-carrier endpoint (Steam's Streaming Microphone) narrowing the desktop mix to
+/// mono or 24 kHz was invisible. Reading it costs one `IAudioClient` activation per endpoint, done
+/// only during a wiring pass.
+///
+/// Deliberately total: EVERY failure maps to `None` ("assume it is fine"), because the wiring plan
+/// treats an unknown format as non-narrowing. A box where activation fails therefore plans exactly
+/// as it did before formats existed, instead of mis-demoting a perfectly good endpoint.
+fn mix_format_of(ep: &Endpoint) -> Option<MixFormat> {
+    let fmt = open_endpoint(ep)
+        .ok()?
+        .get_iaudioclient()
+        .ok()?
+        .get_mixformat()
+        .ok()?;
+    Some(MixFormat {
+        rate_hz: fmt.get_samplespersec(),
+        channels: fmt.get_nchannels(),
+        bits: fmt.get_bitspersample(),
+    })
+}
+
 /// `(friendly_name, endpoint_id)` for every ACTIVE endpoint in direction `dir`.
 fn list_endpoints(dir: Direction) -> Vec<Endpoint> {
     let mut out = Vec::new();
@@ -69,10 +98,22 @@ fn list_endpoints(dir: Direction) -> Vec<Endpoint> {
     out
 }
 
-/// `PUNKTFUNK_HOST_AUDIO`: the operator wants the stream audible on the host too — the loopback
-/// plan prefers real hardware over the silent sink (the pre-client-only-default behavior).
+/// The operator wants the stream audible on the host too — the loopback plan prefers real
+/// hardware over the silent sink (the pre-client-only-default behavior).
+///
+/// Now driven by the first-class `audio.output_mode` setting
+/// ([`AudioOutputMode`](pf_host_config::AudioOutputMode)), which still honours the older
+/// `PUNKTFUNK_HOST_AUDIO` spelling.
 pub(crate) fn host_audio_requested() -> bool {
-    std::env::var_os("PUNKTFUNK_HOST_AUDIO").is_some()
+    pf_host_config::config()
+        .audio_output_mode
+        .prefers_host_hardware()
+}
+
+/// The operator's default playback/recording devices must not be touched at all — the
+/// `follow_default` mode, formerly `PUNKTFUNK_KEEP_DEFAULT`.
+pub(crate) fn keep_default_devices() -> bool {
+    pf_host_config::config().audio_output_mode.keeps_default()
 }
 
 /// One wiring pass plus the inputs the desktop-audio capture loop's failure handling needs:
@@ -118,7 +159,27 @@ pub(crate) fn wire_now_full(set_playback: bool) -> WiredPlan {
     let want = std::env::var("PUNKTFUNK_MIC_DEVICE")
         .ok()
         .map(|s| s.to_lowercase());
-    let wiring = plan(&renders, &captures, want.as_deref(), host_audio_requested());
+    // 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
+    // activation per render endpoint on every pass.
+    let probe: &dyn Fn(&Endpoint) -> Option<MixFormat> = if set_playback {
+        &mix_format_of
+    } else {
+        &wiring_plan::no_formats
+    };
+    let wiring = plan_with_formats(
+        &renders,
+        &captures,
+        want.as_deref(),
+        host_audio_requested(),
+        probe,
+        // The loopback is opened at the session's negotiated channel count, but the wiring pass
+        // runs before (and outside) any session. Stereo is the floor every session uses and the
+        // only count a *narrowing* verdict can be made against without guessing: an endpoint that
+        // cannot carry stereo cannot carry 5.1 either.
+        2,
+    );
     let done = |wiring: Wiring| WiredPlan {
         wiring,
         fingerprint,
@@ -142,6 +203,18 @@ pub(crate) fn wire_now_full(set_playback: bool) -> WiredPlan {
             renders = ?renders.iter().map(|(n, _)| n.as_str()).collect::<Vec<_>>(),
             "audio wiring plan"
         );
+        // The quality warning the 2026-08-03 report had no way to produce. Says WHICH endpoint,
+        // WHY it is narrow, and the two things the operator can actually do about it.
+        if let (Some(why), Some((name, _))) = (&wiring.loopback_narrowing, &wiring.loopback_render)
+        {
+            tracing::warn!(
+                device = %name,
+                "the desktop-audio loopback endpoint {why} — streamed audio will sound worse \
+                 than it does on the host. Attach or select a 48 kHz stereo output device, or \
+                 set audio.output_mode = host_and_client (PUNKTFUNK_HOST_AUDIO=1) to prefer \
+                 real hardware"
+            );
+        }
         if wiring.mic_render.is_some() && wiring.loopback_unsatisfiable() {
             // Inventory + per-endpoint reasons + ONLY the remedies not already taken — the old
             // static advice here suggested installing the Steam pair to a field box that had it
@@ -153,10 +226,11 @@ pub(crate) fn wire_now_full(set_playback: bool) -> WiredPlan {
         }
     }
 
-    if std::env::var_os("PUNKTFUNK_KEEP_DEFAULT").is_some() {
+    if keep_default_devices() {
         if changed {
             tracing::info!(
-                "PUNKTFUNK_KEEP_DEFAULT set — leaving the audio default devices untouched"
+                mode = %pf_host_config::config().audio_output_mode.as_str(),
+                "audio output mode is follow_default — leaving the audio default devices untouched"
             );
         }
         return done(wiring);
@@ -317,6 +391,25 @@ fn park_default_playback(name: &str, id: &str, changed: bool, mic_id: Option<&st
     }
 }
 
+/// Put the default playback device back on the endpoint we are already capturing, WITHOUT a
+/// wiring pass (WP2.4).
+///
+/// The capture loop uses this when something else takes the default mid-stream: in Assert mode the
+/// capture is bound to the planned endpoint explicitly, so the only thing a hijacked default
+/// changes is where *apps* render — one `IPolicyConfig` write fixes that, where the old path tore
+/// the capture down and re-ran the whole wiring pass. Deliberately does not touch the [`PARKED`]
+/// memo: the endpoint is the one we already parked, so the operator's original default is
+/// unchanged and still owed back at stream end.
+pub(crate) fn reassert_default_playback(id: &str) -> bool {
+    match set_default_endpoint(id) {
+        Ok(()) => true,
+        Err(e) => {
+            tracing::debug!(error = %format!("{e:#}"), "failed to re-assert the default playback device");
+            false
+        }
+    }
+}
+
 /// Put the operator's default playback device back after streaming — the inverse of
 /// [`park_default_playback`]. No-op if we never parked it, and a default the operator changed
 /// themselves mid-stream is left alone (their choice wins). Must run on a COM-initialized thread

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

@@ -27,6 +27,7 @@
 //! succeed). On thread exit (capturer dropped at stream end) the parked default playback
 //! device is restored.
 
+use super::capture_policy::{CaptureStats, FightDamper, FIGHT_BACKOFF, STATS_EVERY};
 use super::{audio_control, wiring_plan, AudioCapturer, SAMPLE_RATE};
 use anyhow::{anyhow, Context, Result};
 use std::collections::VecDeque;
@@ -359,7 +360,7 @@ fn capture_once(
 ) -> Result<Next> {
     // Interleaved f32: channels * 4 bytes per frame.
     let block_align = channels as usize * 4;
-    let keep_default = std::env::var_os("PUNKTFUNK_KEEP_DEFAULT").is_some();
+    let keep_default = audio_control::keep_default_devices();
     // Assert-mode without KEEP_DEFAULT is the only shape that parks the playback default.
     let assert_plan = mode == TargetMode::Assert && !keep_default;
     let mut plan = audio_control::wire_now_full(assert_plan);
@@ -454,12 +455,25 @@ fn capture_once(
         channels as usize,
         Some(mask),
     );
-    let (default_period, _min_period) =
-        audio_client.get_device_period().context("device period")?;
+    // WP0.1 — the endpoint's ACTUAL engine mix format, read BEFORE we initialize. Everything the
+    // old log printed ("48 kHz f32 channels=2") was our REQUEST; with `autoconvert` WASAPI
+    // silently converts from whatever the endpoint really runs, so a voice-carrier endpoint
+    // narrowing the desktop mix to mono or 24 kHz was invisible in a 3,600-line field log. This
+    // line is what makes an audio-quality report triageable without a round trip.
+    let engine = audio_client.get_mixformat().ok();
+    // NB the plan's WP4.5 ("open the loopback at the MINIMUM device period, worth ~5–10 ms") is
+    // deliberately NOT done here, because its premise is wrong: in shared mode
+    // `IAudioClient::Initialize` cannot change the engine period at all — `hnsBufferDuration` sizes
+    // the buffer, and the callback still fires at the engine's fixed default period. Lowering it
+    // needs `IAudioClient3::InitializeSharedAudioStream`, which the `wasapi` crate does not wrap.
+    // Passing `min_period` here would therefore be a no-op at best and a new Initialize failure
+    // path at worst, on a device this tree cannot compile for, let alone test. Left as real work.
+    let (default_period, min_period) = audio_client.get_device_period().context("device period")?;
     let stream_mode = StreamMode::EventsShared {
         autoconvert: true,
         buffer_duration_hns: default_period,
     };
+    let used_period = default_period;
     audio_client
         .initialize_client(&desired, &Direction::Capture, &stream_mode)
         .context("initialize loopback client")?;
@@ -476,7 +490,17 @@ fn capture_once(
     tracing::info!(device = %dev_name,
         follow = matches!(mode, TargetMode::Follow) || keep_default,
         last_resort,
+        // The endpoint's own format — NOT the one we asked for.
+        engine_hz = engine.as_ref().map(|f| f.get_samplespersec()),
+        engine_ch = engine.as_ref().map(|f| f.get_nchannels()),
+        engine_bits = engine.as_ref().map(|f| f.get_bitspersample()),
+        buffer_ms = used_period as f32 / 10_000.0,
+        min_buffer_ms = min_period as f32 / 10_000.0,
         "audio loopback capturing");
+    if let Some(why) = &wiring.loopback_narrowing {
+        tracing::warn!(device = %dev_name,
+            "capturing an endpoint that {why} — the stream cannot sound better than this source");
+    }
 
     // Watchdog seed: the default as it stands right after our open. In Assert mode the plan just
     // parked the default on our endpoint — if it did NOT stick (IPolicyConfig denied) converge
@@ -514,6 +538,15 @@ fn capture_once(
     let opened_at = Instant::now();
     let mut saw_packets = false;
     let mut silence_noted = false;
+    // WP0.2 — the audio plane's own vitals, logged periodically. Before this, a host log said
+    // nothing whatsoever about audio between "capturing" and the session ending: no level, no
+    // cadence, and in particular no sign of the SILENT, uncounted drop below, where a stalled
+    // encode thread loses chunks and the encoder simply concatenates across the hole (a click,
+    // and a permanent A/V offset, with nothing in any log).
+    let mut stats = CaptureStats::default();
+    let mut last_stats = Instant::now();
+    // WP2.4 — damping for the default-playback tug-of-war.
+    let mut fight = FightDamper::new(Instant::now());
     loop {
         if stop.load(Ordering::Relaxed) {
             audio_client.stop_stream().ok();
@@ -556,7 +589,34 @@ fn capture_once(
             for c in raw.chunks_exact(4) {
                 samples.push(f32::from_le_bytes([c[0], c[1], c[2], c[3]]));
             }
-            let _ = tx.try_send(samples); // non-blocking, lossy — same discipline as PipeWire
+            stats.observe(&samples, channels);
+            // Non-blocking, lossy — same discipline as PipeWire. Now COUNTED: a full channel
+            // means the encode thread is not keeping up, and every dropped chunk is a click plus
+            // a permanent shift of everything after it.
+            if tx.try_send(samples).is_err() {
+                stats.dropped_chunks += 1;
+            }
+        }
+        if last_stats.elapsed() >= STATS_EVERY {
+            let (peak_db, rms_db, delivered_pct) = stats.summary(last_stats.elapsed(), SAMPLE_RATE);
+            if stats.dropped_chunks > 0 {
+                tracing::warn!(
+                    device = %dev_name,
+                    dropped_chunks = stats.dropped_chunks,
+                    "the audio encode thread could not keep up — captured audio was DROPPED; the \
+                     stream will click and everything after it shifts"
+                );
+            }
+            tracing::info!(
+                device = %dev_name,
+                peak_db = format!("{peak_db:.1}"),
+                rms_db = format!("{rms_db:.1}"),
+                delivered_pct = format!("{delivered_pct:.0}"),
+                dropped_chunks = stats.dropped_chunks,
+                "desktop audio capture"
+            );
+            last_stats = Instant::now();
+            stats = CaptureStats::default();
         }
 
         // Watchdog: react when the default render device CHANGES from what we last observed —
@@ -568,29 +628,68 @@ fn capture_once(
                 if seen_default.as_deref() != Some(nid.as_str()) {
                     seen_default = Some(nid.clone());
                     if nid != dev_id {
-                        audio_client.stop_stream().ok();
+                        // NB the stream is stopped per-branch below, NOT here: the WP2.4 Dud
+                        // path deliberately keeps capturing, and stopping first would have made
+                        // the "no teardown" fix silently useless.
                         if keep_default {
+                            audio_client.stop_stream().ok();
                             tracing::info!(
                                 "default render device changed (PUNKTFUNK_KEEP_DEFAULT) — \
                                  following it"
                             );
                             return Ok(Next::Reopen(TargetMode::Follow));
                         }
-                        return Ok(match judge_default(&en, wiring, &nid) {
+                        match judge_default(&en, wiring, &nid) {
                             DefaultKind::Capturable(name) => {
+                                audio_client.stop_stream().ok();
                                 tracing::info!(device = %name,
                                     "operator changed the output device mid-stream — following \
                                      it (audio now also plays on the host)");
-                                Next::Reopen(TargetMode::Follow)
+                                return Ok(Next::Reopen(TargetMode::Follow));
                             }
+                            // WP2.4 — a DUD default does not affect what we are capturing:
+                            // Assert mode binds the capture to the plan's endpoint EXPLICITLY,
+                            // not to whatever the default happens to be. Only where *apps*
+                            // render has moved. So put the default back and KEEP THE STREAM —
+                            // the old full reopen tore the capture down for nothing, and the
+                            // 2026-08-03 field log shows what that cost: something re-set the
+                            // default to CABLE Input every ~4 s and each round trip was a
+                            // teardown, a re-plan with IPolicyConfig writes, and an audible
+                            // dropout — seven of them in sixteen seconds, one ending in a 2 s
+                            // error backoff.
                             DefaultKind::Dud(name) => {
-                                tracing::warn!(device = %name,
-                                    "default playback moved to an endpoint whose loopback cannot \
-                                     work — re-asserting the audio wiring plan");
-                                Next::Reopen(TargetMode::Assert)
+                                if !assert_plan {
+                                    // Follow/KEEP_DEFAULT shapes still need the old behaviour:
+                                    // there the capture IS bound to the default.
+                                    audio_client.stop_stream().ok();
+                                    return Ok(Next::Reopen(TargetMode::Assert));
+                                }
+                                fight.observed_at(Instant::now());
+                                if fight.should_reassert() {
+                                    audio_control::reassert_default_playback(&dev_id);
+                                    // Believe our own write: the next watchdog tick sees the
+                                    // default back on our endpoint and stays quiet.
+                                    seen_default = Some(dev_id.clone());
+                                    if fight.warn_now() {
+                                        tracing::warn!(device = %name, planned = %dev_name,
+                                            "something keeps moving the default playback to an \
+                                             endpoint whose loopback cannot work — putting it \
+                                             back (the capture is unaffected)");
+                                    }
+                                } else if fight.warn_giving_up() {
+                                    tracing::warn!(device = %name, planned = %dev_name,
+                                        backoff_s = FIGHT_BACKOFF.as_secs(),
+                                        "another program is repeatedly taking the default \
+                                         playback device — backing off rather than fighting it. \
+                                         Desktop audio keeps streaming from the planned endpoint, \
+                                         but apps rendering to the other device will not be heard");
+                                }
                             }
-                            DefaultKind::Unknown => Next::Reopen(TargetMode::Assert),
-                        });
+                            DefaultKind::Unknown => {
+                                audio_client.stop_stream().ok();
+                                return Ok(Next::Reopen(TargetMode::Assert));
+                            }
+                        }
                     }
                 }
             }

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

@@ -43,6 +43,59 @@
 /// A `(friendly_name, endpoint_id)` pair as enumerated from WASAPI.
 pub(crate) type Endpoint = (String, String);
 
+/// A render endpoint's ENGINE MIX FORMAT, as `IAudioClient::GetMixFormat` reports it.
+///
+/// This is the number the 2026-08-03 field report needed and the log did not have. The capture
+/// side opens with `autoconvert: true` and asks for 48 kHz f32 in the wire layout, so WASAPI
+/// silently converts whatever the endpoint really runs — and the "48 kHz f32 channels=2" we
+/// logged was our REQUEST, not the source. An endpoint that mixes at 24 kHz mono therefore
+/// produced a 48 kHz stereo stream that had already been through a 24 kHz mono bottleneck, with
+/// nothing in any log to say so.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) struct MixFormat {
+    pub rate_hz: u32,
+    pub channels: u16,
+    pub bits: u16,
+}
+
+impl MixFormat {
+    /// Why this endpoint would NARROW a `want`-channel desktop mix, or `None` if it carries it
+    /// intact. Bit depth is deliberately not a criterion: 16-bit is ~96 dB of headroom, far below
+    /// Opus's own noise floor, whereas a lost channel or halved bandwidth is plainly audible.
+    pub(crate) fn narrowing(&self, want: u8) -> Option<String> {
+        if self.rate_hz < 48_000 && self.channels < want as u16 {
+            return Some(format!(
+                "mixes at {} Hz and only {} channel(s)",
+                self.rate_hz, self.channels
+            ));
+        }
+        if self.rate_hz < 48_000 {
+            return Some(format!(
+                "mixes at {} Hz, so the stream is band-limited to ~{} kHz before Opus sees it",
+                self.rate_hz,
+                self.rate_hz / 2000
+            ));
+        }
+        if self.channels < want as u16 {
+            return Some(format!(
+                "mixes {} channel(s), so a {want}-channel desktop mix is downmixed and re-expanded",
+                self.channels
+            ));
+        }
+        None
+    }
+}
+
+/// Looks up a render endpoint's mix format by endpoint id. `None` = unknown (enumeration failed,
+/// or the caller has no way to ask) — treated as "assume it is fine", so a probe failure can
+/// never make the plan worse than it was before formats existed.
+pub(crate) type FormatProbe<'a> = &'a dyn Fn(&Endpoint) -> Option<MixFormat>;
+
+/// A [`FormatProbe`] that knows nothing — the pre-WP2.1 behaviour.
+pub(crate) fn no_formats(_: &Endpoint) -> Option<MixFormat> {
+    None
+}
+
 /// The coherent endpoint assignment for one wiring pass. Computed fresh on every mic/capture
 /// (re)open — Windows endpoints churn (boot-time registration, hotplug, driver installs), so a
 /// once-per-process plan goes stale.
@@ -60,6 +113,11 @@ pub(crate) struct Wiring {
     /// the mic reservation. The capture side treats it as a stopgap: it warns when the silence
     /// materializes and re-plans on any endpoint-set change instead of riding it out.
     pub loopback_last_resort: bool,
+    /// Set when the chosen loopback endpoint's mix format NARROWS the desktop mix (see
+    /// [`MixFormat::narrowing`]) and the plan took it anyway because nothing better existed. Carries
+    /// the human-readable reason for the capture side to log — a quality risk the operator can act
+    /// on (attach a real output, or set the output mode to prefer hardware), not a failure.
+    pub loopback_narrowing: Option<String>,
 }
 
 impl Wiring {
@@ -137,6 +195,32 @@ pub(crate) fn plan(
     captures: &[Endpoint],
     mic_want: Option<&str>,
     host_audio: bool,
+) -> Wiring {
+    plan_with_formats(renders, captures, mic_want, host_audio, &no_formats, 2)
+}
+
+/// [`plan`] with knowledge of each render endpoint's engine mix format, and the channel count the
+/// session wants to carry.
+///
+/// **The 2026-08-03 field report is this function's reason to exist.** The default client-only
+/// preference takes the "silent sink" — Steam's Streaming *Microphone* render endpoint — over real
+/// hardware unconditionally, because it is silent on the host. But that endpoint exists to carry
+/// remote *voice*, and nothing checked whether it could carry music. On the reporter's box it won
+/// all 31 loopback opens across 25 sessions while a clean AMD HD Audio endpoint sat idle, and the
+/// whole desktop mix went through it before reaching Opus.
+///
+/// So a silent sink now has to EARN its preference: if its mix format narrows the mix (see
+/// [`MixFormat::narrowing`]) it drops below real hardware. It is still taken when nothing better
+/// exists — narrow audio beats no audio — but flagged in [`Wiring::loopback_narrowing`] so the
+/// capture side can say why. An unknown format (probe failed) counts as fine, so this can never
+/// make the plan worse than it was before formats existed.
+pub(crate) fn plan_with_formats(
+    renders: &[Endpoint],
+    captures: &[Endpoint],
+    mic_want: Option<&str>,
+    host_audio: bool,
+    format_of: FormatProbe,
+    want_channels: u8,
 ) -> Wiring {
     let find_render = |needle: &str| {
         renders
@@ -172,10 +256,18 @@ pub(crate) fn plan(
             not_mic(id) && !excluded_from_loopback(&ln) && !virtualish(&ln)
         })
     };
-    let silent = || {
-        renders
-            .iter()
-            .find(|(n, id)| not_mic(id) && silent_sink(&n.to_lowercase()))
+    // A silent sink splits in two: one that carries the mix intact, and one that narrows it. The
+    // first keeps the historical preference; the second falls BELOW real hardware.
+    let narrowing_of = |ep: &Endpoint| format_of(ep).and_then(|f| f.narrowing(want_channels));
+    let silent_intact = || {
+        renders.iter().find(|ep| {
+            not_mic(&ep.1) && silent_sink(&ep.0.to_lowercase()) && narrowing_of(ep).is_none()
+        })
+    };
+    let silent_narrow = || {
+        renders.iter().find(|ep| {
+            not_mic(&ep.1) && silent_sink(&ep.0.to_lowercase()) && narrowing_of(ep).is_some()
+        })
     };
     // LAST RESORT — the Steam Streaming Speakers, and ONLY them. Their loopback is known-silent
     // (validated live): a QUALITY risk, flagged so the capture side can warn when the silence
@@ -192,10 +284,13 @@ pub(crate) fn plan(
             .iter()
             .find(|(n, id)| not_mic(id) && n.to_lowercase().contains("steam streaming speakers"))
     };
+    // A narrowing silent sink sits below real hardware in BOTH modes: preferring silence on the
+    // host is a routing choice, but it must not silently cost audio quality when a clean endpoint
+    // is right there.
     let preferred = if host_audio {
-        real_hw().or_else(silent)
+        real_hw().or_else(silent_intact).or_else(silent_narrow)
     } else {
-        silent().or_else(real_hw)
+        silent_intact().or_else(real_hw).or_else(silent_narrow)
     };
     let (loopback_render, loopback_last_resort) = match preferred {
         Some(ep) => (Some(ep.clone()), false),
@@ -204,12 +299,16 @@ pub(crate) fn plan(
             None => (None, false),
         },
     };
+    // Report narrowing for whatever we actually chose — including real hardware, which can also
+    // be a 24 kHz mono endpoint (a headset's hands-free profile is exactly that).
+    let loopback_narrowing = loopback_render.as_ref().and_then(narrowing_of);
 
     Wiring {
         mic_render,
         mic_capture,
         loopback_render,
         loopback_last_resort,
+        loopback_narrowing,
     }
 }
 
@@ -550,6 +649,169 @@ mod tests {
         }
     }
 
+    // ---- format-aware loopback selection (WP2.1) -----------------------------------------
+
+    fn fmt(rate_hz: u32, channels: u16) -> MixFormat {
+        MixFormat {
+            rate_hz,
+            channels,
+            bits: 32,
+        }
+    }
+
+    /// Probe helper: give endpoints whose (lowercased) name contains a needle that format,
+    /// everything else unknown. Owns its table so call sites can pass a literal inline.
+    fn probe(table: Vec<(&'static str, MixFormat)>) -> impl Fn(&Endpoint) -> Option<MixFormat> {
+        move |ep: &Endpoint| {
+            let name = ep.0.to_lowercase();
+            table
+                .iter()
+                .find_map(|(needle, f)| name.contains(needle).then_some(*f))
+        }
+    }
+
+    /// THE 2026-08-03 field case, with formats. The reporter's exact endpoint inventory: the plan
+    /// took the Steam Streaming Microphone on all 31 opens while a clean AMD HD Audio endpoint sat
+    /// idle. Once we can see that the silent sink narrows the mix, real hardware must win.
+    #[test]
+    fn narrowing_silent_sink_loses_to_real_hardware() {
+        let renders = [
+            ep("CABLE In 16ch (VB-Audio Virtual Cable)"),
+            ep("Altavoces (Steam Streaming Speakers)"),
+            ep("Altavoces (Steam Streaming Microphone)"),
+            ep("CABLE Input (VB-Audio Virtual Cable)"),
+            ep("1 - Odyssey G60SD (AMD High Definition Audio Device)"),
+        ];
+        let captures = [
+            ep("CABLE Output (VB-Audio Virtual Cable)"),
+            ep("Microphone (Steam Streaming Microphone)"),
+        ];
+        // A voice-carrier endpoint: 24 kHz mono.
+        let p = probe(vec![
+            ("steam streaming microphone", fmt(24_000, 1)),
+            ("odyssey", fmt(48_000, 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)",
+            "a narrowing silent sink must not beat clean real hardware"
+        );
+        assert!(
+            w.loopback_narrowing.is_none(),
+            "the chosen endpoint is intact"
+        );
+        // The mic assignment is untouched by any of this.
+        assert_eq!(
+            w.mic_render.unwrap().0,
+            "CABLE Input (VB-Audio Virtual Cable)"
+        );
+    }
+
+    /// …but a silent sink that carries the mix intact keeps its historical preference: the
+    /// client-only routing default is not being abandoned, only made conditional on quality.
+    #[test]
+    fn intact_silent_sink_still_wins() {
+        let renders = [
+            ep("Speakers (Realtek HD Audio)"),
+            ep("CABLE Input (VB-Audio Virtual Cable)"),
+            ep("Speakers (Steam Streaming Microphone)"),
+        ];
+        let p = probe(vec![
+            ("steam streaming microphone", fmt(48_000, 2)),
+            ("realtek", fmt(48_000, 2)),
+        ]);
+        let w = plan_with_formats(&renders, &[], None, false, &p, 2);
+        assert_eq!(
+            w.loopback_render.unwrap().0,
+            "Speakers (Steam Streaming Microphone)"
+        );
+    }
+
+    /// Narrow audio still beats NO audio: with nothing else available the narrowing sink is taken
+    /// and flagged, not refused.
+    #[test]
+    fn narrowing_sink_is_taken_when_it_is_all_there_is() {
+        let renders = [
+            ep("CABLE Input (VB-Audio Virtual Cable)"),
+            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);
+        assert_eq!(
+            w.loopback_render.as_ref().unwrap().0,
+            "Speakers (Steam Streaming Microphone)"
+        );
+        let why = w.loopback_narrowing.expect("must be flagged");
+        assert!(why.contains("16000"), "{why}");
+    }
+
+    /// Real hardware can narrow too — a headset in its hands-free profile is 16 kHz mono — and
+    /// must be flagged just the same. The flag is about the CHOSEN endpoint, not about which tier
+    /// it came from.
+    #[test]
+    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);
+        assert_eq!(
+            w.loopback_render.as_ref().unwrap().0,
+            "Headset (Hands-Free AG Audio)"
+        );
+        assert!(w.loopback_narrowing.is_some());
+    }
+
+    /// An unknown format must never make the plan WORSE than it was before formats existed: a
+    /// probe that answers nothing has to reproduce `plan` exactly.
+    #[test]
+    fn unknown_formats_reproduce_the_formatless_plan() {
+        let renders = [
+            ep("Speakers (Apple Audio Device)"),
+            ep("CABLE Input (VB-Audio Virtual Cable)"),
+            ep("Speakers (Steam Streaming Speakers)"),
+            ep("Speakers (Steam Streaming Microphone)"),
+        ];
+        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);
+            assert_eq!(a, b, "host_audio={host_audio}");
+            assert!(a.loopback_narrowing.is_none());
+        }
+    }
+
+    /// `host_audio` still prefers real hardware, and a narrowing silent sink stays last in that
+    /// mode too.
+    #[test]
+    fn host_audio_ordering_survives_formats() {
+        let renders = [
+            ep("Speakers (Realtek HD Audio)"),
+            ep("Speakers (Steam Streaming Microphone)"),
+        ];
+        let p = probe(vec![
+            ("steam streaming microphone", fmt(24_000, 1)),
+            ("realtek", fmt(48_000, 2)),
+        ]);
+        let w = plan_with_formats(&renders, &[], None, true, &p, 2);
+        assert_eq!(w.loopback_render.unwrap().0, "Speakers (Realtek HD Audio)");
+    }
+
+    /// The narrowing test is channel-count aware: an endpoint that is fine for stereo narrows a
+    /// 5.1 session.
+    #[test]
+    fn narrowing_depends_on_the_session_channel_count() {
+        let stereo_only = fmt(48_000, 2);
+        assert_eq!(stereo_only.narrowing(2), None);
+        assert!(stereo_only.narrowing(6).is_some());
+        // Rate is judged independently of channels.
+        assert!(fmt(44_100, 8).narrowing(2).is_some());
+        // And an endpoint wider than the session is never "narrowing".
+        assert_eq!(fmt(48_000, 8).narrowing(2), None);
+        // Both wrong: the message must name both problems.
+        let both = fmt(16_000, 1).narrowing(6).unwrap();
+        assert!(both.contains("16000") && both.contains("channel"), "{both}");
+    }
+
     /// Operator override beats the candidate order.
     #[test]
     fn env_override_wins() {

crates/punktfunk-host/src/native.rs

@@ -26,9 +26,7 @@
 #![deny(clippy::undocumented_unsafe_blocks)]
 
 use anyhow::{anyhow, Context, Result};
-use punktfunk_core::config::{
-    mtu1500_shard_payload_for, CompositorPref, FecConfig, FecScheme, GamepadPref, Role,
-};
+use punktfunk_core::config::{CompositorPref, FecConfig, FecScheme, GamepadPref, Role};
 use punktfunk_core::input::{InputEvent, InputKind};
 use punktfunk_core::packet::{FLAG_PIC, FLAG_PROBE, FLAG_SOF};
 use punktfunk_core::quic::{
@@ -72,6 +70,9 @@ use input::{input_thread, ClientInput};
 /// The Hello→Welcome→Start negotiation (plan §W1); `serve_session` calls `handshake::negotiate`
 /// after the pairing gate.
 mod handshake;
+/// MTU resilience for the video data plane: `PUNKTFUNK_WIRE_MTU` override, the per-session
+/// path-MTU watch on the control connection, and the per-peer learned shard-payload clamp.
+mod wire_mtu;
 
 /// The mid-stream control task (plan §W1); `serve_session` spawns `control::run` after the
 /// handshake to multiplex renegotiation / speed-test control messages onto the data-plane channels.
@@ -1307,9 +1308,18 @@ async fn serve_session(
         let stop = stop.clone();
         let cap = audio_cap.clone();
         let channels = welcome.audio_channels;
+        // Read the granted bit back off the Welcome (the cursor plane's precedent), so the wire
+        // the client was promised and the wire we actually send cannot disagree — then re-derive
+        // the SAME budget rung from it, so the encode tier and the redundancy decision are one
+        // choice made once rather than two settings that can drift apart.
+        let budget = handshake::audio_budget(
+            welcome.host_caps & punktfunk_core::quic::HOST_CAP_AUDIO_RED != 0,
+            welcome.bitrate_kbps,
+            channels,
+        );
         std::thread::Builder::new()
             .name("punktfunk1-audio".into())
-            .spawn(move || audio_thread(conn, stop, cap, channels))
+            .spawn(move || audio_thread(conn, stop, cap, channels, budget))
             .map_err(|e| tracing::warn!(error = %e, "audio thread spawn failed — session continues without audio"))
             .ok()
     } else {

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

@@ -1,8 +1,13 @@
 //! The native audio plane (plan §W1 — carved out of the [`super`] module): desktop capture → Opus
-//! (48 kHz, 5 ms, CBR — the same tuning as the GameStream path) → `AUDIO_MAGIC` QUIC datagrams, at
-//! the negotiated channel count. The encoder ([`NativeAudioEnc`]) and the capture/encode/send loop
-//! ([`audio_thread`]) are gated to linux/windows (libopus + a real capturer); other targets get the
-//! stub, so a dev build streams video-only rather than failing to compile.
+//! (48 kHz, 5 ms, constrained VBR at the configured [`AudioTier`](punktfunk_core::audio::AudioTier))
+//! → `AUDIO_MAGIC` QUIC datagrams — or `AUDIO_RED_MAGIC` when the session negotiated redundancy —
+//! at the negotiated channel count. The encoder ([`NativeAudioEnc`]) and the capture/encode/send
+//! loop ([`audio_thread`]) are gated to linux/windows (libopus + a real capturer); other targets
+//! get the stub, so a dev build streams video-only rather than failing to compile.
+//!
+//! Two things here deliberately DIVERGE from the GameStream plane, which used to share this
+//! tuning: hard CBR (its audio FEC needs fixed-size packets; this plane has no FEC, so CBR was a
+//! pure quality tax) and the fixed 128 kbps stereo bitrate. See [`NativeAudioEnc::new`].
 
 use super::*;
 
@@ -17,20 +22,36 @@ enum NativeAudioEnc {
 
 #[cfg(any(target_os = "linux", target_os = "windows"))]
 impl NativeAudioEnc {
-    /// Build the encoder for `channels` (2/6/8), hard-CBR + RESTRICTED_LOWDELAY like the
-    /// GameStream path; bitrate from the shared layout table (stereo keeps the validated 128 kbps).
-    fn new(channels: u8) -> Result<NativeAudioEnc, opus::Error> {
+    /// Build the encoder for `channels` (2/6/8) at `tier`, RESTRICTED_LOWDELAY like the GameStream
+    /// path but — unlike it — in CONSTRAINED VBR.
+    ///
+    /// **Why not hard CBR (WP1.2).** The layout table's comment justifies `set_vbr(false)` with
+    /// "constant packet size, which GameStream's audio FEC relies on" — true of the GameStream
+    /// plane, and irrelevant here: the native `punktfunk/1` audio plane has no FEC at all (see
+    /// `punktfunk_core::audio::AudioGapTracker`, which exists precisely because a lost packet has
+    /// nothing to rebuild it from). So this path was paying a pure quality tax for a constraint
+    /// that does not apply to it. Constrained VBR keeps the same average bitrate and the same
+    /// bounded packet size, and spends the bits where the signal needs them.
+    ///
+    /// The GameStream encoder (`crate::gamestream::audio`) is deliberately NOT changed: its FEC
+    /// really does need fixed-size packets.
+    fn new(
+        channels: u8,
+        tier: punktfunk_core::audio::AudioTier,
+    ) -> Result<NativeAudioEnc, opus::Error> {
+        let l = punktfunk_core::audio::layout_for(channels, false);
+        let bitrate = l.bitrate_for(tier);
         if channels == 2 {
             let mut e = opus::Encoder::new(
                 crate::audio::SAMPLE_RATE,
                 opus::Channels::Stereo,
                 opus::Application::LowDelay,
             )?;
-            e.set_bitrate(opus::Bitrate::Bits(128_000)).ok();
-            e.set_vbr(false).ok();
+            e.set_bitrate(opus::Bitrate::Bits(bitrate)).ok();
+            e.set_vbr(true).ok();
+            e.set_vbr_constraint(true).ok();
             Ok(NativeAudioEnc::Stereo(e))
         } else {
-            let l = punktfunk_core::audio::layout_for(channels, false);
             let mut e = opus::MSEncoder::new(
                 crate::audio::SAMPLE_RATE,
                 l.streams,
@@ -38,8 +59,9 @@ impl NativeAudioEnc {
                 l.mapping,
                 opus::Application::LowDelay,
             )?;
-            e.set_bitrate(opus::Bitrate::Bits(l.bitrate)).ok();
-            e.set_vbr(false).ok();
+            e.set_bitrate(opus::Bitrate::Bits(bitrate)).ok();
+            e.set_vbr(true).ok();
+            e.set_vbr_constraint(true).ok();
             Ok(NativeAudioEnc::Surround(e))
         }
     }
@@ -52,8 +74,8 @@ impl NativeAudioEnc {
     }
 }
 
-/// The audio thread: desktop capture → Opus (48 kHz, 5 ms, CBR — same tuning as the GameStream
-/// path) → `AUDIO_MAGIC` datagrams, at the negotiated `channels` (2 stereo / 6 = 5.1 / 8 = 7.1,
+/// The audio thread: desktop capture → Opus (48 kHz, 5 ms, constrained VBR at the configured
+/// tier) → `AUDIO_MAGIC` (or `AUDIO_RED_MAGIC`) datagrams, at the negotiated `channels` (2 stereo / 6 = 5.1 / 8 = 7.1,
 /// canonical wire order FL FR FC LFE RL RR SL SR). QUIC already encrypts; no extra layer. The
 /// capturer comes from (and returns to) the persistent slot — see [`AudioCapSlot`].
 #[cfg(any(target_os = "linux", target_os = "windows"))]
@@ -62,11 +84,16 @@ pub(super) fn audio_thread(
     stop: Arc<AtomicBool>,
     audio_cap: AudioCapSlot,
     channels: u8,
+    budget: punktfunk_core::audio::AudioBudget,
 ) {
     use crate::audio::SAMPLE_RATE;
     const FRAME_MS: usize = 5;
     const SAMPLES_PER_FRAME: usize = SAMPLE_RATE as usize * FRAME_MS / 1000; // 240
     let want = punktfunk_core::audio::normalize_channels(channels);
+    // Tier and redundancy are ONE decision, budgeted against the session's video bitrate — see
+    // `handshake::audio_budget`. An unparseable `audio.quality` was already warned about there
+    // and fell back to the default, so nothing here can silently downgrade someone's audio.
+    let (tier, redundancy) = (budget.tier, budget.redundancy);
 
     // Reuse the cached capturer ONLY when its channel count matches this session's; a stereo
     // capturer left by a prior session must not feed a 5.1/7.1 session (the encoder + the client's
@@ -92,7 +119,7 @@ pub(super) fn audio_thread(
             }
         }
     };
-    let mut enc = match NativeAudioEnc::new(want) {
+    let mut enc = match NativeAudioEnc::new(want, tier) {
         Ok(e) => e,
         Err(e) => {
             tracing::warn!(error = %e, "opus encoder init failed — session continues without audio");
@@ -120,9 +147,16 @@ pub(super) fn audio_thread(
     // A stuck Opus encoder would fail on every 5 ms frame (~200/s); power-of-two throttle the
     // warn so it can't flood stderr + the log ring while still surfacing that it's failing.
     let mut opus_encode_errs: u64 = 0;
+    // WP3.1 — the previous frame's Opus bytes, for the redundant `0xD2` plane. Cleared whenever
+    // continuity breaks (a capture reopen), so we never advertise a predecessor the client's
+    // sequence numbering does not agree with.
+    let mut prev_frame: Vec<u8> = Vec::new();
     if capturer.is_some() {
         tracing::info!(
             channels = want,
+            tier = tier.as_str(),
+            kbps = budget.kbps,
+            redundancy,
             "punktfunk/1 audio streaming (Opus 48 kHz, 5 ms datagrams)"
         );
     }
@@ -138,6 +172,10 @@ pub(super) fn audio_thread(
                     capturer = Some(c);
                     last_failed = None;
                     acc.clear(); // drop the partial frame straddling the gap
+                                 // The next frame has no valid predecessor across the gap: sending the
+                                 // pre-gap frame as "the previous one" would hand the client audio from
+                                 // before the discontinuity to splice in.
+                    prev_frame.clear();
                 }
                 Err(e) => {
                     tracing::debug!(error = %format!("{e:#}"), "audio reopen failed — will retry");
@@ -162,11 +200,24 @@ pub(super) fn audio_thread(
             let pts_ns = now_ns();
             match enc.encode_float(&frame, &mut opus_buf) {
                 Ok(n) => {
-                    let d =
-                        punktfunk_core::quic::encode_audio_datagram(seq, pts_ns, &opus_buf[..n]);
+                    let opus = &opus_buf[..n];
+                    let d = if redundancy {
+                        punktfunk_core::quic::encode_audio_red_datagram(
+                            seq,
+                            pts_ns,
+                            opus,
+                            &prev_frame,
+                        )
+                    } else {
+                        punktfunk_core::quic::encode_audio_datagram(seq, pts_ns, opus)
+                    };
                     if conn.send_datagram(d.into()).is_err() {
                         break 'session; // connection gone
                     }
+                    if redundancy {
+                        prev_frame.clear();
+                        prev_frame.extend_from_slice(opus);
+                    }
                     seq = seq.wrapping_add(1);
                 }
                 Err(e) => {
@@ -199,6 +250,7 @@ pub(super) fn audio_thread(
     _stop: Arc<AtomicBool>,
     _audio_cap: AudioCapSlot,
     _channels: u8,
+    _budget: punktfunk_core::audio::AudioBudget,
 ) {
     tracing::warn!("punktfunk/1 audio requires Linux or Windows — session continues without it");
 }

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

@@ -24,6 +24,63 @@ use super::*;
 /// paints on a Mutter virtual stream), and only a can't-blend backend falls back to the
 /// compositor EMBED. THE single predicate: the Welcome's `HOST_CAP_CURSOR` bit is computed
 /// from it, and the session wiring reads that bit back.
+/// THE single audio-plane decision for a session: the encode tier AND whether the redundant
+/// `0xD2` plane is sent. The Welcome's `HOST_CAP_AUDIO_RED` bit is computed from it, and
+/// `serve_session` reads that bit back to configure the audio thread — so the wire the client is
+/// promised and the wire we send cannot disagree.
+///
+/// Capable-and-agreed for redundancy: the client must have advertised `CLIENT_CAP_AUDIO_RED`, so a
+/// session with an older client keeps the plain `0xC9` wire byte-for-byte.
+///
+/// **Both halves are then BUDGETED against the session's video bitrate**
+/// ([`plan_audio_budget`](punktfunk_core::audio::plan_audio_budget)). Tier `High` and redundancy
+/// were introduced separately, each costed as "~1 % of the video budget", and they multiply:
+/// 256 kbps stereo sent twice is 512 kbps — ~10 % of a 5 Mbps session. Audio rides QUIC datagrams,
+/// outside the ABR loop, so ABR can neither see that nor reclaim it. The budget is what stops a
+/// constrained link silently handing a tenth of its bandwidth to audio.
+///
+/// The operator's `audio.quality` / `audio.redundancy` settings are the REQUEST; the budget may
+/// lower them, never raise them.
+///
+/// NB the plan's "only while the link is actually losing packets" gate is deliberately not here:
+/// turning redundancy on and off mid-session changes the wire tag, and the client's decoder would
+/// have to re-derive which plane it is on from every datagram. Deciding once, at handshake, against
+/// a bitrate we already know is both cheaper and more predictable.
+/// `wants_redundancy` is the caller's answer to "is `0xD2` even on the table" — at handshake that
+/// is the client's cap AND the operator's setting; afterwards it is the GRANTED
+/// `HOST_CAP_AUDIO_RED` bit, so the audio thread re-derives the same rung of the same ladder.
+pub(super) fn audio_budget(
+    wants_redundancy: bool,
+    video_kbps: u32,
+    channels: u8,
+) -> punktfunk_core::audio::AudioBudget {
+    let configured = pf_host_config::config().audio_quality.as_deref();
+    let requested = match configured {
+        None => punktfunk_core::audio::AudioTier::default(),
+        Some(s) => punktfunk_core::audio::AudioTier::parse(s).unwrap_or_else(|| {
+            // Once per process: this runs per session, and an operator with a typo in host.env
+            // does not need it on every connect. Never silently downgrade someone's audio.
+            static WARNED: std::sync::Once = std::sync::Once::new();
+            WARNED.call_once(|| {
+                tracing::warn!(
+                    value = %s,
+                    "audio.quality (PUNKTFUNK_AUDIO_QUALITY) is not one of low/standard/high — \
+                     using the default"
+                );
+            });
+            punktfunk_core::audio::AudioTier::default()
+        }),
+    };
+    punktfunk_core::audio::plan_audio_budget(video_kbps, channels, requested, wants_redundancy)
+}
+
+/// The operator's answer to "may this session use redundancy at all", before the budget is
+/// consulted: the client must be able to decode it and the operator must not have forced it off.
+pub(super) fn redundancy_offered(client_caps: u8) -> bool {
+    client_caps & punktfunk_core::quic::CLIENT_CAP_AUDIO_RED != 0
+        && pf_host_config::config().audio_redundancy.unwrap_or(true)
+}
+
 pub(super) fn cursor_forward(
     client_caps: u8,
     compositor: Option<crate::vdisplay::Compositor>,
@@ -491,7 +548,12 @@ pub(super) async fn negotiate(
         // per-datagram loss on Wi-Fi — the "100 Mbps badly fails on the phone" root cause.
         // Negotiated, so the client follows. Jumbo (≈8900) is a future negotiated bump (needs
         // MAX_DATAGRAM_BYTES raised + end-to-end 9000 MTU).
-        shard_payload: mtu1500_shard_payload_for(peer.ip()) as u16,
+        // Resolution order (wire_mtu.rs): `PUNKTFUNK_WIRE_MTU` operator override, then a path
+        // budget learned from a prior session whose QUIC MTU discovery settled below the
+        // video-datagram ceiling (the "VPN on the host blackholes every video packet" field
+        // shape — small flows pass, the stream is an endless black screen), then this family
+        // default. Healthy paths take the default branch and are byte-identical to before.
+        shard_payload: wire_mtu::negotiated_shard_payload(peer.ip()) as u16,
         encrypt: true,
         key,
         salt,
@@ -564,6 +626,20 @@ pub(super) async fn negotiate(
                 punktfunk_core::quic::HOST_CAP_PEN
             } else {
                 0
+            }
+            // Redundant desktop-audio plane (0xD2): the client asked, the operator has not forced
+            // it off, AND it fits the session's audio budget. Capable-and-agreed like the cursor
+            // bit — a client that did not ask keeps the plain 0xC9 wire byte-for-byte.
+            | if audio_budget(
+                redundancy_offered(hello.client_caps),
+                bitrate_kbps,
+                audio_channels,
+            )
+            .redundancy
+            {
+                punktfunk_core::quic::HOST_CAP_AUDIO_RED
+            } 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
@@ -658,6 +734,10 @@ pub(super) async fn negotiate(
     let start =
         Start::decode(&io::read_msg(recv).await?).map_err(|e| anyhow!("Start decode: {e:?}"))?;
     bringup.mark("start");
+    // The session is real: watch this connection's MTU discovery settle and turn it into a
+    // path verdict (WARN + learned clamp for the next session on a constrained path; clears a
+    // stale clamp on a healthy one). Bounded ~10 s task, ends by itself.
+    wire_mtu::spawn_watch(conn.clone(), welcome.shard_payload as usize);
     Ok::<_, anyhow::Error>((
         hello,
         welcome,

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

@@ -0,0 +1,192 @@
+//! MTU resilience for the video data plane (the "connects fine, black screen forever" field
+//! shape).
+//!
+//! Video datagrams are sealed at a per-session `shard_payload` sized for a clean 1500-byte MTU
+//! (1472-byte UDP payloads). A host whose route to the client runs through a smaller-MTU hop —
+//! a VPN/overlay adapter (Tailscale/WARP/ZeroTier default to 1280) claiming the LAN route, or a
+//! lowered NIC MTU — delivers every SMALL flow (QUIC control, hole punch, input, audio) while
+//! 100 % of video datagrams die by fragmentation or local `WSAEMSGSIZE`: the client sits on a
+//! black screen reporting `loss_ppm=0` (it can't see gaps in packets it never saw any of) and
+//! the host streams into the void with every gauge green. Neither side observes the failure
+//! directly — but the control connection CAN: its MTU discovery probes up to exactly the sealed
+//! video-datagram size ([`video_datagram_udp_ceiling`], set in `quic/endpoint.rs`), so its
+//! settled MTU is a verdict on the path.
+//!
+//! Three legs, none of which changes a session on a healthy path:
+//! - **`PUNKTFUNK_WIRE_MTU=<bytes>`** — operator override; the shard payload is derived from
+//!   the given on-wire IP MTU. Wire-compatible with every deployed client:
+//!   `Welcome::shard_payload` is already negotiated per session (the v4/v6 split ships two
+//!   values today) and clients follow the negotiated value.
+//! - **Watch** — a per-session task samples the control connection's discovered MTU once the
+//!   search has had time to finish. A connection still alive that settled BELOW the ceiling is
+//!   proof the path can't carry full-size video: log an actionable WARN and record the measured
+//!   budget for the peer.
+//! - **Heal** — the next handshake from that peer clamps `shard_payload` to the recorded
+//!   budget, so a reconnect fixes the stream. A later session that reaches the ceiling erases
+//!   the record (the learn/heal loop is self-correcting in both directions).
+
+use std::collections::HashMap;
+use std::net::IpAddr;
+use std::sync::{Mutex, OnceLock};
+
+use punktfunk_core::config::{
+    mtu1500_shard_payload_for, sealed_datagram_bytes, shard_payload_for_udp_budget,
+    shard_payload_for_wire_mtu, video_datagram_udp_ceiling,
+};
+
+/// Measured UDP-payload budget per peer IP, learned from live control connections whose MTU
+/// discovery settled below the video-datagram ceiling. In-memory only: a host restart
+/// re-learns in one session, and entries self-correct (a later ceiling-hit erases, a lower
+/// re-measure overwrites).
+fn learned() -> &'static Mutex<HashMap<IpAddr, u16>> {
+    static LEARNED: OnceLock<Mutex<HashMap<IpAddr, u16>>> = OnceLock::new();
+    LEARNED.get_or_init(|| Mutex::new(HashMap::new()))
+}
+
+/// The shard payload for a new session to `peer`: `PUNKTFUNK_WIRE_MTU` override, else the
+/// peer's learned path budget, else the family default (today's exact behavior). Logs whenever
+/// the result differs from the default.
+pub(super) fn negotiated_shard_payload(peer: IpAddr) -> usize {
+    let env = match std::env::var("PUNKTFUNK_WIRE_MTU") {
+        Ok(v) => match v.trim().parse::<usize>() {
+            Ok(mtu) => Some(mtu),
+            Err(_) => {
+                tracing::warn!(value = %v, "PUNKTFUNK_WIRE_MTU is not a number — ignoring it");
+                None
+            }
+        },
+        Err(_) => None,
+    };
+    let learned_budget = learned().lock().unwrap().get(&peer).copied();
+    resolve(env, learned_budget, peer)
+}
+
+/// Pure resolution (env override > learned budget > family default) — the tested core of
+/// [`negotiated_shard_payload`].
+fn resolve(env_wire_mtu: Option<usize>, learned_udp_budget: Option<u16>, peer: IpAddr) -> usize {
+    let default = mtu1500_shard_payload_for(peer);
+    if let Some(mtu) = env_wire_mtu {
+        let p = shard_payload_for_wire_mtu(mtu, peer);
+        if p != default {
+            tracing::info!(
+                wire_mtu = mtu,
+                shard_payload = p,
+                default,
+                "wire MTU: shard payload set from PUNKTFUNK_WIRE_MTU"
+            );
+        }
+        return p;
+    }
+    if let Some(budget) = learned_udp_budget {
+        let p = shard_payload_for_udp_budget(budget as usize, peer);
+        if p != default {
+            tracing::info!(
+                peer = %peer,
+                udp_budget = budget,
+                shard_payload = p,
+                default,
+                "wire MTU: shard payload clamped to this peer's measured path MTU (learned \
+                 from a prior session's QUIC MTU discovery) — video datagrams now fit the \
+                 constrained hop"
+            );
+            return p;
+        }
+    }
+    default
+}
+
+/// Sample the control connection's discovered MTU after the search has settled and turn it
+/// into a verdict. Spawned once per negotiated session; the task ends by itself after the
+/// final sample (bounded ~10 s lifetime, holding only a cheap `Connection` handle).
+pub(super) fn spawn_watch(conn: quinn::Connection, session_shard_payload: usize) {
+    tokio::spawn(async move {
+        let peer = conn.remote_address().ip();
+        let ceiling = video_datagram_udp_ceiling() as u16;
+        // Discovery finishes in a handful of RTTs on a LAN (well under the first sample) but
+        // needs a loss timeout per failed probe on a constrained path — the second sample
+        // covers that with margin. Max, because discovery only ever raises `current_mtu`.
+        let mut settled = 0u16;
+        for wait_s in [3u64, 7] {
+            tokio::time::sleep(std::time::Duration::from_secs(wait_s)).await;
+            settled = settled.max(conn.stats().path.current_mtu);
+            if settled >= ceiling {
+                break;
+            }
+        }
+        if settled >= ceiling {
+            // The path carries full-size video datagrams — erase any stale learned clamp so
+            // the next session returns to the default wire.
+            if learned().lock().unwrap().remove(&peer).is_some() {
+                tracing::info!(peer = %peer,
+                    "wire MTU: path re-measured at full size — learned clamp cleared");
+            }
+            return;
+        }
+        // A closed connection stops discovering, so a session that ended before the final
+        // sample proves nothing (a healthy high-RTT path could still be mid-search): learn
+        // only from a connection that stayed alive through the whole window.
+        if conn.close_reason().is_some() {
+            return;
+        }
+        learned().lock().unwrap().insert(peer, settled);
+        if sealed_datagram_bytes(session_shard_payload) <= settled as usize {
+            // This session was already clamped small enough — the path is still constrained
+            // (keep the record fresh) but video fits, so no alarm.
+            tracing::info!(peer = %peer, discovered_udp_mtu = settled,
+                "wire MTU: constrained path re-measured; this session's video is sized to fit");
+        } else {
+            tracing::warn!(
+                peer = %peer,
+                discovered_udp_mtu = settled,
+                needed_udp_mtu = ceiling,
+                "wire MTU: this path CANNOT carry full-size video datagrams — the control \
+                 plane works but every video packet is oversized for a hop, which streams as \
+                 an endless black screen with zero reported loss. Typical cause: a VPN/overlay \
+                 adapter (Tailscale / Cloudflare WARP / ZeroTier) claiming the LAN route, or a \
+                 lowered NIC MTU — compare `ping <client> -f -l 1450` vs `-l 1200` and check \
+                 `netsh interface ipv4 show subinterfaces` (Windows) / `ip link` (Linux). The \
+                 measured budget is recorded: the NEXT session from this client sizes video to \
+                 fit automatically. To pin it for all sessions set PUNKTFUNK_WIRE_MTU."
+            );
+        }
+    });
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::net::{IpAddr, Ipv4Addr, Ipv6Addr};
+
+    const V4: IpAddr = IpAddr::V4(Ipv4Addr::new(192, 168, 1, 2));
+    const V6: IpAddr = IpAddr::V6(Ipv6Addr::new(0x2001, 0xdb8, 0, 0, 0, 0, 0, 1));
+
+    #[test]
+    fn default_when_nothing_known() {
+        assert_eq!(resolve(None, None, V4), mtu1500_shard_payload_for(V4));
+        assert_eq!(resolve(None, None, V6), mtu1500_shard_payload_for(V6));
+    }
+
+    #[test]
+    fn env_override_beats_learned() {
+        // 1280 wire − 28 IP/UDP − 64 header/crypto = 1188.
+        assert_eq!(resolve(Some(1280), Some(1472), V4), 1188);
+    }
+
+    #[test]
+    fn learned_budget_clamps() {
+        // A WARP-shaped path: 1280-byte UDP budget → 1280 − 64 = 1216.
+        assert_eq!(resolve(None, Some(1280), V4), 1216);
+    }
+
+    #[test]
+    fn learned_at_or_above_ceiling_is_the_default_wire() {
+        assert_eq!(resolve(None, Some(1472), V4), mtu1500_shard_payload_for(V4));
+        assert_eq!(resolve(None, Some(2000), V4), mtu1500_shard_payload_for(V4));
+    }
+
+    #[test]
+    fn env_full_mtu_is_the_default_wire_both_families() {
+        assert_eq!(resolve(Some(1500), None, V4), mtu1500_shard_payload_for(V4));
+        assert_eq!(resolve(Some(1500), None, V6), mtu1500_shard_payload_for(V6));
+    }
+}

docs-site/content/docs/troubleshooting.md

@@ -302,6 +302,73 @@ told your client so. [When the client and the host
 disagree](/docs/client-settings#when-the-client-and-the-host-disagree) lists what it does with each
 one.
 
+## Streamed audio sounds worse than the host does
+
+The host does not capture "the sound card" — it captures a **render endpoint**, and by default it
+picks one that is *silent on the host* so the audio plays on your client only. On a PC with Steam
+installed that silent endpoint is Steam's **Streaming Microphone**, which exists to carry remote
+*voice*. If Windows has it configured as a narrow device — mono, or below 48 kHz — then the whole
+desktop mix is squeezed through that before it is ever encoded, and no amount of bitrate will bring
+it back.
+
+Since 0.25 the host checks for this: it reads each candidate endpoint's real format, prefers a real
+output device over a narrow virtual one, and says so in the log —
+
+```
+WARN  the desktop-audio loopback endpoint mixes at 24000 Hz, so the stream is band-limited …
+INFO  audio loopback capturing device="…" engine_hz=48000 engine_ch=2 engine_bits=32
+```
+
+That `engine_*` line is the endpoint's **own** format, so it tells you directly whether the source
+was ever full quality. To choose the routing yourself, set in `host.env`:
+
+```ini
+# client_only     — default; audio plays on the client only (a silent endpoint)
+# host_and_client — capture a real output device; audio plays on BOTH ends
+# follow_default  — capture whatever YOUR default playback device is, and never change it
+PUNKTFUNK_AUDIO_OUTPUT_MODE=host_and_client
+```
+
+`host_and_client` is also the quickest way to A/B the problem: if the stream sounds right that way
+and wrong on the default, the endpoint was the cause.
+
+Two related knobs:
+
+```ini
+PUNKTFUNK_AUDIO_QUALITY=high    # low | standard | high (default high — stereo 256 kbps)
+PUNKTFUNK_AUDIO_REDUNDANCY=1    # force the loss-resilient audio plane on (default: automatic)
+```
+
+Both are a **request**, not a guarantee: the host budgets audio against the session's video
+bitrate and steps it down on a narrow link, because audio is not managed by adaptive bitrate — so
+whatever it takes is taken off the top. On a roomy link you get 256 kbps plus loss redundancy; as
+the link narrows the host drops redundancy first, then the tier, and never goes below ~96 kbps. The
+session log line says what it settled on:
+
+```
+INFO  punktfunk/1 audio streaming … tier=high kbps=512 redundancy=true
+```
+
+`standard` reproduces the pre-0.25 encoder exactly if you want to A/B it.
+
+## Audio lags behind the picture
+
+The client buffers a little audio to absorb network jitter. Since 0.25 that buffer **corrects
+itself**: if it drifts deeper — a Wi-Fi burst, a stall, or just the two devices' clocks running at
+fractionally different speeds — it trims itself back a few milliseconds at a time, inaudibly.
+Before, it could only grow, so a single hiccup left audio permanently behind the video and the only
+cure was reconnecting.
+
+If audio is still noticeably late:
+
+- **Reconnect once.** It confirms whether the delay was accumulated (gone after a reconnect) or
+  constant (something else).
+- **Check for underruns** rather than guessing. The client logs its buffer depth periodically; a
+  rising `underruns` count means the buffer is being starved, which is a network or CPU problem, not
+  a buffering one.
+- **Wired or 5 GHz Wi-Fi.** Arrival jitter is what the buffer exists to absorb; less jitter lets it
+  run shallower.
+
 ## Windows: the host or the web console won't start
 
 The **`PunktfunkHost` service** runs both halves of the Windows host: the streaming host itself and

include/punktfunk_core.h

@@ -312,6 +312,13 @@
 // `PunktfunkStatus` code).
 #define PUNKTFUNK_CLIP_ERROR 6
 
+// The protocol's audio frame, in milliseconds — every host datagram carries exactly one
+// ([`crate::quic::encode_audio_datagram`]), so it is also the smallest useful shed unit.
+#define PUNKTFUNK_AUDIO_FRAME_MS 5
+
+// Sample rate of every audio plane in the protocol.
+#define PUNKTFUNK_AUDIO_SAMPLE_RATE_HZ 48000
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // The uniform no-TTL-host staleness bound: a legacy host refreshes state every 500 ms, so two
 // missed refreshes = quiet host → silence. Replaces the per-platform zoo (1.6 s / 60 s / 1.5 s /
@@ -333,6 +340,12 @@
 #define INBOUND_REQ_FLAG 2147483648
 #endif
 
+// Floor for a negotiated `shard_payload` (even, well under every real path). A path whose UDP
+// budget lands below this can't carry the QUIC control plane either (QUIC's own minimum is a
+// 1200-byte UDP payload), so shrinking video shards further buys nothing — the clamp helpers
+// bottom out here instead of producing degenerate confetti-sized shards.
+#define MIN_SHARD_PAYLOAD 512
+
 // 16-byte AEAD authentication tag appended by either session cipher.
 #define TAG_LEN 16
 
@@ -627,6 +640,19 @@
 #define CLIENT_CAP_PHASE_LOCK 2
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// `Hello.client_caps` bit: this client can decode the redundant desktop-audio plane
+// ([`AUDIO_RED_MAGIC`](super::datagram::AUDIO_RED_MAGIC), `0xD2`), where every datagram also
+// carries a copy of the previous frame so a single lost packet is reconstructed instead of
+// papered over with packet-loss concealment.
+//
+// Active only when the host answers with [`HOST_CAP_AUDIO_RED`] (capable-and-agreed, the
+// cursor/clipboard precedent). Toward an older host, or a host that declines because the link is
+// clean, the client keeps receiving the plain `0xC9` plane — so a client may always set this bit.
+// `0x04` — `0x01`/`0x02` are cursor / phase-lock.
+#define CLIENT_CAP_AUDIO_RED 4
+#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,
@@ -652,6 +678,20 @@
 #define HOST_CAP_PEN 16
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`Welcome::host_caps`] bit: the host is sending the REDUNDANT desktop-audio plane
+// ([`AUDIO_RED_MAGIC`](super::datagram::AUDIO_RED_MAGIC), `0xD2`) instead of plain `0xC9` — each
+// datagram carries its own frame plus a copy of the previous one.
+//
+// Set only when the client asked via [`CLIENT_CAP_AUDIO_RED`]. It is a statement about the WIRE,
+// not a negotiation the client can decline: with the bit set the client must decode `0xD2`, and
+// without it `0xC9`. The host may also drop back to `0xC9` mid-session (the redundancy is
+// loss-gated — a clean LAN shouldn't pay for it), which is why clients decode BOTH tags
+// unconditionally and treat this bit as "expect redundancy", not "only redundancy".
+// `0x20` — `0x10` is [`HOST_CAP_PEN`], `0x08` is [`HOST_CAP_CURSOR`].
+#define HOST_CAP_AUDIO_RED 32
+#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
@@ -967,6 +1007,41 @@
 #define HIDOUT_MAGIC 205
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Redundant audio datagram, host → client: the [`AUDIO_MAGIC`] plane plus a copy of the PREVIOUS
+// frame, so a single lost datagram is *reconstructed* rather than concealed.
+//
+// `[0xD2][u32 seq LE][u64 pts_ns LE][u16 primary_len LE][primary opus][previous opus]`
+//
+// **Why this and not Opus in-band FEC.** LBRR is a SILK-layer feature: the desktop-audio encoder
+// runs `RESTRICTED_LOWDELAY` (CELT-only) at 5 ms frames, which is below SILK's 10 ms minimum, so
+// `set_inband_fec(true)` on that encoder is a no-op. Nothing in libopus can protect this plane —
+// the redundancy has to be at the application layer. (The mic uplink is a different encoder, VoIP
+// mode at 10 ms, and *does* use real in-band FEC.)
+//
+// **Why it costs no latency.** The copy rides the SUCCESSOR of the frame it protects, and the
+// client is already holding 15–90 ms of de-jitter buffer — far more than the 5 ms the successor
+// takes to arrive. So the recovery happens inside slack that already exists.
+//
+// The previous frame's sequence is implicitly `seq - 1`; a host with nothing to duplicate yet
+// (the first frame of a session, or straight after a capture reopen) simply sends an empty tail,
+// which decodes to `None`.
+//
+// Sent ONLY when the client advertised [`CLIENT_CAP_AUDIO_RED`](super::caps::CLIENT_CAP_AUDIO_RED)
+// and the host answered [`HOST_CAP_AUDIO_RED`](super::caps::HOST_CAP_AUDIO_RED) — the
+// capable-and-agreed handshake the cursor and 4:4:4 planes already use. Every other session keeps
+// the plain [`AUDIO_MAGIC`] wire byte-for-byte.
+//
+// NB `0xD1` is deliberately skipped: the DualSense pad-audio program has reserved it for the
+// per-pad audio plane.
+#define PUNKTFUNK_AUDIO_RED_MAGIC 210
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Fixed header length of an [`AUDIO_RED_MAGIC`] datagram (tag + seq + pts + primary length).
+#define PUNKTFUNK_AUDIO_RED_HEADER (((1 + 4) + 8) + 2)
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // Wire length of a v1 (legacy, level) rumble datagram.
 #define RUMBLE_V1_LEN 7
@@ -1397,6 +1472,14 @@ typedef uint8_t PunktfunkInputKind;
 typedef struct ColorInfo ColorInfo;
 #endif
 
+// Tuning for [`JitterPolicy`], in MILLISECONDS.
+//
+// Denominating the depth in time rather than in device quanta is the point. Every client used to
+// compute its target as `3 × quantum`, which is a sane 15 ms at a 5 ms quantum and a silent 64 ms
+// at a 20 ms one — the same source line meaning two very different latencies depending on what
+// else happened to be using the audio graph that day.
+typedef struct JitterTuning JitterTuning;
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // Opaque handle to a live `punktfunk/1` connection (QUIC control plane + UDP data plane, all
 // pumped on internal threads).
@@ -1788,6 +1871,14 @@ typedef struct {
 
 
 
+
+
+
+
+
+
+
+
 // The multipliers a picker offers. `1.0` (Native) is the default; the rest are the round stops
 // users reason about. Shared so every client's list stays identical.
 #define PRESETS { 0.5, 0.67, 0.75, 1.0, 1.25, 1.5, 2.0, 3.0, 4.0, }