After caa0548 (fixes A + B for ticket 0001), blanka-chrome was still
silent. Fresh telemetry pinned a third failure mode:
21:31:31 audio via AudioContext 47e5 sink=c38572ec… ← good
…leave + rejoin…
21:31:33 audio via AudioContext 47e5 sink=default ← drifted
No `audioCtx sink → c38572ec…` log line between the two attaches.
applySinkToAudioCtx was only invoked inside the `if (!audioCtx)` branch
in attachAudioStreamViaWorklet — a persistent audioCtx whose .sinkId
getter returned '' (system default) after a leave/rejoin never got
its sink re-applied, so every subsequent attach emitted to the system
default speaker instead of the device the user picked in the dropdown.
Fix: when audioCtx already exists AND speakerDeviceId is set AND
audioCtx.sinkId !== speakerDeviceId, call applySinkToAudioCtx() to
restore routing. Idempotent (same-sink call is a noop).
Telemetry: every 5s tick now includes `ctx.sink=…` and `want=…` when
the active sink doesn't match the picked one, so drift is visible
without grepping for the rare attach event.
Test: attach re-applies sink when audioCtx persists across leave/rejoin
and the picked sink drifted. Sandbox now exposes setSpeakerDeviceId so
the test can drive the picked-device path through the real
attachAudioStreamViaWorklet branch.
Ticket 0001 updated with the second-pass telemetry and fix (C).
Russell Ballestrini
4cedd85dc6