[pksgit]

Follow-up to #1465, which listed this path under Follow-ups: a selected track's own EOS reaching the selector while the egress is running.

input-selector forwards the active pad's EOS downstream. When a selected track ends on its own while the egress is still running, its EOS goes through the selector into video_encoder_queue. The track's removal then switches the selector to the filler behind that EOS. Two ways a track can end like this:

• its appsrc stops with not-negotiated: basesrc pushes EOS right after posting the error, and removal follows from the bus message;
• its app writer finishes before the track is removed, and pushes EOS as part of its cleanup.

The EOS travels through the track's video_input_queue on its own thread, so whether it reaches the selector before the removal's switch is a race. When it wins:

• the video encoder has been sent EOS while the egress keeps running;
• the filler renegotiates after the switch and sends an ALLOCATION query. If it reaches video_encoder_queue while that EOS is still draining, gst_queue keeps the query after its sender gives up and frees it, and the filler's STREAM_START later restarts the queue onto the freed pointer: Unexpected item ... dequeued from queue video_encoder_queue (refcounting problem?), followed by a stream error. This queue behaviour is the same in GStreamer 1.24 and 1.28.

Change

Each track's selector pad gets an event probe that calls onTrackEOS. When an EOS arrives while the pipeline is still running and the track is selected, the probe switches the selector to the filler before the EOS reaches the selector, and clears a pending handover to that track. The selector then holds the EOS as it does for any inactive input, until the track's removal releases the pad. This is the switch the track's removal already makes, moved ahead of the EOS.

A track's EOS at shutdown passes unchanged. Shutdown EOS can reach the selector before the video bin's EOS function runs, but always after the pipeline has left the running state, so the probe checks the pipeline state.

Order of events	Before	After
Selected track's EOS reaches the selector before its removal switches	EOS into the encoder; the filler follows behind it; can fail as above	Filler takes over at the EOS; the selector holds the EOS; the output continues on the filler
Removal switches before the track's EOS arrives	EOS held on an inactive pad	Same
Track pending (filler on screen) when its EOS arrives	EOS held; the pending handover is cleared at removal	EOS held; the pending handover is cleared at the EOS
Selected track's EOS during shutdown	Forwarded, ends the video stream	Same

Testing

• New TestSelectorHandover_SelectedTrackEOSWhileRunning runs the sequence against the production gates, a real input-selector and the production video_encoder_queue: the filler is on screen, a track takes the selector, the track's EOS arrives (if forwarded, it is held past the encoder queue by a probe, standing in for the encoder draining it), the track's removal switches to the filler, the filler's ALLOCATION query is sent and freed once answered, then later filler frames arrive. It checks that the track's EOS never reaches the encoder queue, that no query is queued behind it, that pushes keep succeeding and that the filler's frames leave the selector. Without the probe it fails every run on both 1.24 and 1.28 core elements, and some runs log Unexpected item ... dequeued from queue video_encoder_queue (refcounting problem?), after which every push fails. With it, it passes.
• New TestSelectorHandover_SelectedTrackEOSAtShutdown checks that once the pipeline is ending, a selected track's EOS leaves the selector and the track stays selected. It fails if the probe ignores the pipeline state.
• go test ./pkg/... with the deadlock tag and golangci-lint, as run by build/test/Dockerfile, pass.
• Integration, locally on linux/arm64 with an image built from this branch: file-media, stream, segments and edge pass. In multi, the Web and ParticipantComposite subtests fail in this local setup because the test can't find their manifest files; both egresses complete, and the same two subtests fail the same way without this change. CI builds amd64.

🤖 Generated with Claude Code

[devin-ai-integration]

🔍 Devin Review: 1 flag

Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)

[github-actions]

AV-sync stats summary: view in run #39637731391