Retire the MoQ audio track on interruption

[kixelated]

Summary

MOQOutputTransport retires its audio track when an interruption catches audio still to be played, publishing a uniquely generation-named successor (bot-audio, bot-audio-1, ...) in its place. This replaces the pacing-clock re-anchor.

The bot writes TTS up to audio_out_max_buffer_ms ahead of real-time, so at barge-in seconds of the abandoned utterance have already left the process. Nothing on the bot side could retract it, so discarding it relied on an out-of-band signal: the bot published RTVI user-started-speaking on the transcript track and the browser flushed its buffer. The two tracks have no ordering relative to each other, and the transcript is subscribed at priority 0 against audio's 80, so the flush signal is scheduled behind exactly the data it exists to flush.

Retiring the track makes the discard structural instead. AudioProducer.finish() drops the producer's catalog rendition, and the successor publishes its own, so the catalog always advertises exactly one audio track and subscribers follow the swap through normal rendition selection.

When the swap happens

Interruptions are broadcast on every user turn, not only on barge-in, so an unconditional swap would cost every subscriber a catalog update and a resubscribe per turn. A swap happens only when the pacing clock's lead over wall-clock exceeds AUDIO_IN_FLIGHT_FLOOR_S (0.3s).

Timestamps

PTS is not per-track. Every track stamps against one broadcast-wide monotonic origin, matching the browser side's performance.now() convention, so tracks stay mutually synchronizable. A successor re-anchors into that timeline at the position it starts playing from rather than continuing from the write-ahead frontier or resetting to zero.

A track takes its epoch at its first write, not when it opens. Only a track's first stamp reaches the wire as given: moq-audio's encode producer takes it as that track's epoch and derives every later PTS from the running sample count. A new producer resets only its group sequence to zero.

Ordering

• process_frame calls super() first so the base class drains the audio queue before the track is retired.
• publish_audio captures its producer up front and drops the chunk if the track was retired while it sat in the pacing sleep.

Test plan

TestMOQAudioTrackRestart and TestMOQAudioTrackRestartLive in tests/test_moq_transport.py cover:

• catalog replacement (finish + uniquely named successor; live catalog advertises exactly one audio track)
• continuous PTS on the broadcast timeline (not zero, not the write-ahead frontier)
• stale paced writes dropped
• group zero on every replacement track (live AudioProducer)

uv run pytest tests/test_moq_transport.py passes 63 tests with the moq extra installed. ruff check and ruff format --check clean.

This supersedes #5443, rebased onto current main with live catalog and group-zero coverage.

(written by grok-4.6)

[codecov]

Codecov Report

❌ Patch coverage is 0% with 37 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/pipecat/transports/moq/transport.py	0.00%	37 Missing ⚠️

Files with missing lines	Coverage Δ
src/pipecat/transports/moq/transport.py	0.00% <0.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.