feat(moq-gst): moqsrc follows a restart on the same pads

[t0ms]

Problem

moqsrc resolves its path once. Now that subscriptions are sticky (#5087), it keeps playing a replaced broadcast until that broadcast's route goes away, and then ends with EOS or a session error. A restarted publisher, a newer instance taking over, or a lite-06 end and start never reaches the pipeline as a switch.

Completes quest/m0/broadcast-epoch/moqsrc.md.

Approach

• Follow announcements. The session follows its path with origin::Consumer::follow from feat(apps): players follow announce Start, Restart, and End #5154. A Start or Restart cancels the old run's pumps, then starts a run on the route serving the path. An Update starts one only when nothing is playing, meaning the path came back after a gap that the announcements folded away. An End holds the pads.

• Pads by rendition. A slot per rendition name holds the pad between pumps, so a restart or an in-run caps or container change hands the same pad to the next pump. The new pump pushes stream-start and its caps; refused caps fail as not negotiated. On its first buffer it pushes a segment whose base is the current running time, so a synced sink does not drop the new run as late. A rendition the catalog retires ends with EOS, and a new one gets a new pad.

• Source loss holds. A track or catalog that ends or loses its source hands its pad back without EOS. This works whether the loss arrives before or after End, and the next Start resumes on the same pads. A catalog that the live source lacks or refuses (NotFound, Unauthorized) still posts the session error, as does a malformed one, and losing the connection still ends the session. Mid-stream only malformed content counts against the source, because a producer that goes away surfaces as almost any error, including NotFound for a group that can never be served.

• Tests. Unit tests cover both orders of source loss and End, a reshape keeping its pad, and a broadcast without a catalog failing the session with NotFound. tests/source.rs runs moqsrc ! fakesink sync=true max-lateness=20ms against an in-process relay. It covers:

  • a newer instance taking over while the old one stays up;
  • an epochless republish;
  • a restarted prefix route;
  • an update that does not switch;
  • new caps on the same pad;
  • an end followed by a start;
  • a closed publisher session.

  Each case checks that the new run renders at its frame rate on one pad with no bus error. Pushing the segment with a base of zero fails all seven, and switching on every Update fails the update case.

Impact

• No new property. An ended broadcast no longer sends EOS on its pads; they hold until the path is announced again.
• A caps or container change within a run keeps its pad instead of replacing it.
• A downstream flow error other than flushing, not-linked, or EOS now posts an element error instead of stopping the pad silently.
• Losing the source is no longer a session error. A missing, refused, or malformed catalog still is.
• Wire: none.

Alternatives

• Fresh pads for each run: rejected in planning, because pad ids come from a process-wide counter, so a pipeline linked by name would go idle.
• A linger timer before EOS: rejected in planning, because announcements already say when the source is back, and a timer cannot tell a slow restart from a dead source.

Follow-ups

• Stacked on fix(net): don't renew a prefix whose winner already restarted #5182 (a restarted prefix restarted twice on lite-06, which a_restarted_prefix_switches found). feat(apps): players follow announce Start, Restart, and End #5154 (Follow) has landed and main is merged in. Main gets merged in again once fix(net): don't renew a prefix whose winner already restarted #5182 lands. If quest: plan moqsink's first catalog and moqsrc A/V alignment #5189 lands first, this PR also drops its two quests' Required links to the moqsrc quest it deletes.
• After a restart, a rendition that the new run's first catalog lists late ends with EOS and gets a new pad. moqsink lists a pad only once its caps arrive. Planned in quest: plan moqsink's first catalog and moqsrc A/V alignment #5189 as quest/m1/gst-first-catalog.md: moqsink reserves each pad's catalog slot when the pad is requested.
• Each pump starts its timestamps and its segment at its own first buffer, so audio and video can be offset by the gap between their first buffers. Per-pump timestamps predate this change. Planned in quest: plan moqsink's first catalog and moqsrc A/V alignment #5189 as quest/m1/gst-src-av-sync.md.

(Written by Claude Opus 5.5)

[t0ms]

Outcome: implementation complete. This PR stays a draft until its dependencies land.

• Local just check passes on this head.
• The relay tests in rs/moq-gst/tests/source.rs catch both regressions they target: all seven fail when the segment base is pushed as zero, and the update test fails when every Update switches. The six exact-path cases passed 8 of 8 repeated nextest runs.
• a_restarted_prefix_switches exposed a moq-net bug. On lite-06, a front still holding a path beneath a prefix renewed it after the prefix's winner had already changed, so the prefix restarted twice. The test failed in 4 of 6 runs. With the fix in fix(net): don't renew a prefix whose winner already restarted #5182 merged in, it passed 10 of 10.
• Decided with the user: stack on feat(apps): players follow announce Start, Restart, and End #5154 and fix(net): don't renew a prefix whose winner already restarted #5182, and fix moq-net in its own PR rather than loosening the test. Main gets merged in once both land.

(Written by Claude Opus 5.5)

[t0ms]

Closing as superseded by #5191, which landed the same moqsrc quest: following the path's announcements, pads kept by rendition across restarts, held pads on source loss, and a missing or refused catalog failing the session. The two follow-ups planned from this PR (moqsink's first catalog and moqsrc A/V alignment) are in #5189. The moq-net fix this PR found, a restarted prefix restarted twice on lite-06, landed separately in #5182.

(Written by Claude Opus 5.5)