bug: speaker device alignment silently fails — users can't hear audio (Bluetooth)

[JamesPHoughton]

Summary

Users report "can't hear" during video calls. Auto-diagnostics correctly identify the cause as speakerNotSet, attempt to re-align the speaker device, and report success — but post-fix validation shows the speaker is still null. The setSpeaker() call completes without error but the device never actually activates.

Sentry issues:

• DELIBERATION-EMPIRICA-PE — reportedAVError: 36 events, 22 users (ongoing)
• DELIBERATION-EMPIRICA-PG — avDiagnosticResponse: 33 events, 18 users (correlated roommate diagnostics)
• DELIBERATION-EMPIRICA-QD — "Preferred speaker not found, showing picker": 75 events
• DELIBERATION-EMPIRICA-QC — same: 42 events
• DELIBERATION-EMPIRICA-QN / QP / QS — "Preferred speaker not found, auto-switching to fallback": ~21 events combined

Investigation Findings

What the Sentry data shows

Diagnostic data from PE events consistently shows all three device types with currentId: null — not just the speaker:

{
  "deviceAlignment": {
    "camera": { "currentId": null, "currentLabel": null },
    "microphone": { "currentId": null, "currentLabel": null },
    "speaker": { "currentId": null, "currentLabel": null }
  }
}

Breadcrumb sequence from PE event 3348c86c:

1. 14:14:04 — "Speaker aligned via id match" — alignment logs success
2. 14:14:04 — TypeError: Cannot read properties of undefined (reading 'then') — bug: startRecording() returns undefined during stage transitions, crashing VideoCall #1226 crash
3. 14:14:21 — Diagnosed causes: speakerNotSet — speaker is null despite "successful" alignment
4. 14:14:21 — Re-aligned speaker device — fix attempt
5. 14:14:22 — Validation: stillPresent: speakerNotSet — fix failed

Key observation: diagnostics vs alignment use different APIs

There is a disconnect between how devices are set and how they are read back:

• Alignment uses devices.setSpeaker() and reads devices.currentSpeaker (from Daily's useDevices() React hook)
• Diagnostics read from callObject.getInputDevices() / callObject.getOutputDevices() (synchronous call object methods)

These are two different APIs to the same underlying state. If the call object is in a bad state (e.g., after the #1226 crash/remount), getInputDevices()/getOutputDevices() may return nulls even though useDevices() shows valid devices.

Hypothesis: many PE events are caused by the #1226 crash

The same TypeError: Cannot read properties of undefined (reading 'then') from #1226 appears in the PE breadcrumbs. The sequence is:

1. Stage transition triggers spurious VideoCall mount (the stale data race from bug: startRecording() returns undefined during stage transitions, crashing VideoCall #1226)
2. VideoCall crashes during initialization (startRecording TypeError)
3. ErrorBoundary remounts VideoCall, but the call object's internal state is corrupted
4. getInputDevices()/getOutputDevices() return nulls for all device types
5. User can't hear → reports "can't hear" → diagnostics find speakerNotSet → fix fails

If this hypothesis is correct, deploying the #1226 fix should significantly reduce PE/PG events.

Remaining concerns (independent of #1226)

Even if #1226 accounts for many events, there are likely real speaker issues too:

1. setSpeaker() can silently fail: The Promise resolves without error even when the underlying HTMLAudioElement.setSinkId() doesn't actually change the output device (e.g., stale device ID from a disconnected Bluetooth device).

2. No post-set verification: After setSpeaker() succeeds, neither the alignment code nor the FixAV recovery verifies that currentSpeaker actually reflects the change.

3. Safari gesture requirement: Safari requires a user gesture for setSinkId(). When alignment runs from a React effect (no gesture context), setSinkId throws NotAllowedError. The gesture prompt system handles this, but QD shows 7 consecutive failures in one session.

4. Bluetooth device churn: Bluetooth devices (AirPods, "sairpods", WH-1000XM4) can disconnect between alignment and use, leaving the speaker unset with no recovery.

Plan

1. Deploy bug: startRecording() returns undefined during stage transitions, crashing VideoCall #1226 fix first (PR fix(call): prevent stale-data VideoCall mount and guard startRecording #1230) and observe whether PE/PG events drop
2. If events persist, investigate:
   • Add post-set verification to alignSpeaker() — re-query devices.currentSpeaker after setSpeaker() and retry with fallback if unchanged
   • Align the diagnostics API to use devices.currentSpeaker (the hook) instead of callObject.getOutputDevices() (the snapshot), or verify they return the same data
   • Add Bluetooth disconnect detection and proactive fallback

User Impact

High — 22 distinct users unable to hear audio during calls. The FixAV panel diagnoses the problem but can't fix it, leaving users stuck.

Affected Code

• client/src/call/hooks/useDeviceAlignment.js — alignSpeaker() lacks post-set verification
• client/src/call/FixAV.jsx — collectAVDiagnostics() reads from callObject.getOutputDevices() which may differ from devices.currentSpeaker
• client/src/call/utils/avRecovery.js — soft fix for speakerNotSet reports false success