zebra-report/docs/tickets/0001-fedora-chrome-cannot-hear-speakers.md

0001: fedora chrome cannot hear speakers

Status: fixed Priority: P0 (silent regression — every invariant green except audible output) Surface: fedora chrome — observed under both listener and speaker roles First seen: 2026-06-04 (or earlier — "for days") Reporter: fox (blanka-chrome identity) Fixed: 2026-06-07, commits dad697d (source) + 1f27b89 (deploy)

Root cause

Chromium will not run its WebRTC audio decoder for a remote MediaStreamTrack whose only consumer is a MediaStreamAudioSourceNode. The decoder requires an HTMLMediaElement (an <audio> or <video> element) bound to the track to actually decode. Without that anchor, the source node produces silence even though packets arrive at full rate, the worklet's started event fires, and audioCtx.state === 'running'. Firefox does not have this restriction.

PulseAudio Playback during repro: chrome stream visible (chrome WAS producing an output stream), level meter pinned at 0 (the stream's content was silence). RTP telemetry: aud.recv pkt=N → N+50/sec bytes=growing level=0.000 jbuf=?. Worklet telemetry: jitter-buffer started uuid=… target=0.7s (queue filled — but with zeros).

Fix

In attachAudioStreamViaWorklet, alongside the existing createMediaStreamSource(stream) → gain → destination chain, create a hidden muted <audio> element with srcObject = stream, append it to the DOM, and call play(). The element makes no audible output (muted = true); its sole purpose is to keep chromium's WebRTC decoder running so the worklet's MediaStreamSource sees real samples. setWorkletStream swaps the anchor's srcObject alongside the source node; detachListenerStream tears the anchor down.

Pinned by tests in test/listener-audio-attach.test.js:

• anchor created on attach
• anchor follows the stream on setWorkletStream
• anchor torn down on detach

Cross-references

• ../../CLAUDE.md "Audio pipeline — load-bearing invariants" rules 3
  • 4 (anchor + always-SFU) are sourced from this ticket.
• feedback-chromium-decoder-anchor memory carries the apply-everywhere version of rule 3 so a future session can't silently undo it.
• test/listener-audio-attach.test.js pins the anchor contract.

Symptom

On fedora chrome, no audio from any speaker in the room is audible. Affects every role chrome takes: as listener, no mic audio plays. As speaker, no peer mic audio plays through the mesh+SFU path. Same chrome can publish mic + camera + screen fine — only the receive-side audio is silent.

Other browsers on the same machine (firefox) hear the room correctly once the same code is loaded — so the codebase is not uniformly broken.

Telemetry / Repro

Signal-server log (/var/log/zebra-spaces-signal.log, CLIENT_LOG actor_handle="blanka-chrome") on 2026-06-06 ~23:24 UTC:

· role=speaker sub=connected/connected pub=connected/connected
  mesh=2 sListen=0 streamMode=0 muted=1 xcr=off
  aud.recv pkt=308 lost=0 bytes=197428 jitter=0.0080 level=0.000 jbuf=? lp=0.2s
  aud.recv pkt=22075 lost=3 bytes=44147 jitter=0.3210 level=0.000 jbuf=? lp=0.9s
  aud.recv pkt=44242 lost=4 bytes=751664 jitter=0.0260 level=0.000 jbuf=? lp=0.3s

• RTP arriving at ~50 pps on the active speaker (real Opus voice traffic)
• level=0.000 on EVERY receiver
• mesh=2 (both mesh peers connected)
• sub=connected and pub=connected (both SFU PCs healthy)

Earlier in the session:

audio attach uuid=2d4a target=0.5s (current pool=3)
audio via AudioContext 2d4a target=0.5s ctxState=running
jitter-buffer installed uuid=2d4a target=0.5s
jitter-buffer started uuid=2d4a target=0.5s
mesh stream swapped into worklet for 2d4a — buffer cushion now applies to mesh too

Every attach invariant green. Worklet started event fired. ctxState=running. Yet the user hears nothing.

User report: chrome tab indicator shows NO audio playing.

Hypothesis

Primary (deployed 2026-06-06, awaiting retest): the speaker-output dropdown in the page calls setSinkId on <audio> elements only. The worklet path (every listener AND every speaker monitor since c3ff58c routed mesh through the worklet) runs through audioCtx.destination, which always emits to the system default sink. If the user picks a non-default sink in the dropdown, the worklet keeps emitting to the wrong device — silence on the device they're monitoring.

Fox 2026-06-06 confirmed picking a non-default sink on fedora chrome. Firefox defaults to system default which happens to match the right speaker for him, which is why firefox music started working without needing this fix.

Secondary (not yet ruled out):

• AudioContext.setSinkId() may not be honored on this chrome version / flag set → silent fallthrough to default
• PulseAudio per-app default routing for chrome may override the picked sink at the OS layer
• The worklet's started=true may flip back to false on a network blip and never re-arm if the queue can't refill to target (rearm guard fires)

Fix

Landed: 467146a 2026-06-06 — applySinkToAudioCtx() calls audioCtx.setSinkId(speakerDeviceId) on creation and in the dropdown's fan-out. Diagnostic log line now includes sink=... so the next telemetry session shows which sink the worklet is routed to.

Deployed: 0f36e02 to www.unturf.com — blanka-chrome needs page reload (it's a code-shell change, not SDP-level).

Pinned by tests:

• audioCtx.setSinkId is called on creation so the worklet routes to the user-picked sink (not just system default)
• audioCtx without setSinkId support (older browser) does not throw — silent fallback to default

2026-06-07 telemetry — sink hypothesis ruled out

Fresh /var/log/zebra-spaces-signal.log capture (00:20:01 — fxhp host publishing real music, blanka-chrome speaker):

audio via AudioContext 085b target=0.5s ctxState=running sink=default
mesh stream swapped into worklet for 085b — buffer cushion now applies to mesh too
...
aud.recv pkt=14622 lost=1 bytes=9117575 jitter=0.0100 level=0.000 jbuf=? lp=0.1s

Comparison — fxhp-phone in the same room, same window, same publisher:

aud.recv pkt=11161 lost=1 bytes=7154201 jitter=0.0090 level=0.001 jbuf=0.49s lp=0.0s

• sink IS routed (deployed fix 467146a confirmed live — later capture shows sink=c38572ec… = the user-picked deviceId, NOT default).
• 50 pps RTP, 9.1 MB of stereo Opus arriving on blanka-chrome's SFU sub PC.
• jbuf=? = jitterBufferEmittedCount === 0 = receiver decodes nothing. There is no consumer wired to the SFU audio receiver.
• fxhp-phone on the same SFU stream has jbuf=0.49s + level=0.001 (native decoder running) — so the SFU is fine, the problem is per- client wiring.

So: sink routing was not the bug, or was only part of it. With the sink fix applied, fedora chrome still hears silence.

Real root cause

Two coupled defects, both visible in handleRemoteSfuTrack (line 4700) and the mesh pc.ontrack (line 6693):

(A) tracks=0 race at mic ontrack. Chrome can fire ontrack where ev.streams[0] exists but has 0 tracks at handler-time (the track is added a microtask later, or MSID-supplant merges new + dead tracks). Screen / camera / game already work around this by constructing a fresh new MediaStream([ev.track]) and passing that. Mic does not. It caches ev.streams[0] directly into sfuStreamsByPubHex and hands the same reference to attachSfuTrack. When that fires before the track shows up on the stream, attachAudioStreamViaWorklet rejects with liveAudio.length === 0 → falls through to the silent <audio> path → logs sfu attach b1a7 fresh=1 tracks=0 + meter for b1a7: MediaStream has no audio track (both verbatim in today's capture).

(B) Mesh swap orphans the SFU receiver. Mesh pc.ontrack calls setWorkletStream(uuid, meshStream) unconditionally — it does NOT verify that mesh is actually emitting audio. The worklet's MediaStreamSource is now bound to the mesh track. The SFU receiver has no consumer, so its decoder doesn't run (matches jbuf=?, level=0). If the mesh track is muted or DTX-silent, the listener hears nothing — and there is no automatic SFU restore until mesh PC actually transitions to failed (line 6739). Mesh that stays connected but silent traps the listener forever.

Hard refresh + leave/enter fixes both: fresh PC negotiation → fresh ontrack with the live track in the stream → attach succeeds on SFU before mesh comes up → mesh swap eventually replaces source but by then SFU receiver is fine and… actually no, this still ends up on mesh. The reason hard-refresh works is that mesh negotiation is fresh and the new mesh track arrives in an unmuted state because audio is already flowing. The mesh swap then lands on a real audio source.

Fix (in progress)

(A) In handleRemoteSfuTrack mic branch, mirror the screen/camera pattern: const s = new MediaStream([ev.track]); sfuStreamsByPubHex.set(pubHex, s); attachSfuTrack(uuid, s). Guarantees the stream handed to attachAudioStreamViaWorklet contains exactly the live track, every time.

(B) In mesh pc.ontrack, gate the setWorkletStream swap on !ev.track.muted (or addEventListener('unmute', swap) if it starts muted). If the mesh track never unmutes, keep SFU as the worklet source. Add a watchdog: if a swapped-in mesh track returns to muted for >5s, swap the worklet back to the cached SFU stream — same pattern the connectionState === 'failed' path already uses at line 6755.

2026-06-07 second telemetry pass — sink DRIFT identified

After deploying (A) + (B), blanka-chrome was still silent. Fresh telemetry pinned a different per-attach failure mode:

21:31:30  audio attach uuid=47e5 target=0.5s (current pool=1)
21:31:31  audio via AudioContext 47e5 target=0.5s ctxState=running sink=c38572ec…
21:31:31  jitter-buffer started uuid=47e5 target=0.5s
…leave/rejoin cycle…
21:31:33  audio attach uuid=47e5 target=0.5s (current pool=1)
21:31:33  audio via AudioContext 47e5 target=0.5s ctxState=running sink=default   ← drifted

No audioCtx sink → c38572ec… log line ever appeared. So applySinkToAudioCtx() either never ran or hit the typeof audioCtx.setSinkId !== 'function' early-return at the moment of execution. The earlier sink fix only invoked applySinkToAudioCtx() inside the if (!audioCtx) branch — a persistent audioCtx whose .sinkId getter returned '' (default) was never re-corrected.

Fix (C) — re-apply sink on every attach when it drifted

attachAudioStreamViaWorklet now: if audioCtx already exists AND speakerDeviceId is set AND audioCtx.sinkId !== speakerDeviceId, call applySinkToAudioCtx(). Idempotent — same-sink call is a no-op.

Also added per-tick ctx.sink= to the telemetry stream so sink drift is visible on every 5s line instead of needing to find the rare attach log. When the active sink doesn't match the picked one, the line includes want=… so the mismatch jumps out.

Status notes

• 2026-06-07 (fixed): chrome decoder anchor landed in dad697d, deployed 1f27b89. fox confirmed audio working on chrome.
• 2026-06-07 (root cause pinned): pavucontrol showed chrome stream present + level=0 with worklet started fired — the chromium-only MediaStreamSource-needs-anchor signature.
• 2026-06-07 (en-route fixes kept): always-SFU receive (1984bb1), SPEAKER_PLAYOUT_DELAY_SEC 0.5s → 0.7s, session fingerprint telemetry (1ed5263), per-peer mesh.recv stats (b9e2541), audioCtx.setSinkId re-apply on rejoin (467146a).
• 2026-06-07 (in-progress, fix C deployed): sink-drift on audioCtx-persistent path identified + patched. Need a third telemetry pass to confirm the per-tick ctx.sink= line now stays on the picked device across leave/rejoin.
• 2026-06-07 (fixes A + B deployed): sink hypothesis from 467146a ruled out by fresh telemetry; (A) tracks=0 race + (B) mesh-swap orphan landed in caa0548. blanka-chrome was still silent, leading to fix (C) above.
• 2026-06-06: the Jun 5 cascade + flushSfuStreams prefix-match fix (0ce1339) helped firefox but not chrome. Telemetry above ruled out the FSM / attach defects.
• 2026-06-04 → 2026-06-05: earlier cascade commits (2e74b92, f479878, 144dd15, 61b21a6) tightened the audio attach path but did not address this defect (they targeted double-audio on fxhp-android-firefox).