fix(transcoder): snap A/V to a common origin at startup for AV-path sources

[ntt0601zcoder]

What

The transcoder maps audio and video onto its output timeline with a single shared offset, assuming both ride one source clock — true for MPEG-TS (shared PCR) but not always for AV-path/RTMP sources, where audio can flow for seconds before the first video keyframe and the two tracks may carry independently-based timestamps. The first track to emit seeded the offset, so a late-arriving second track inherited it and baked a large, permanent A/V offset into the output.

Observed on an RTMP-pull source: output A/V offset up to ~22 s (audio ahead / video lagging). HLS preserved it (audible lip-sync); DASH distorted it differently. Source probes showed the upstream A/V synced to a few ms, and ingest traces showed the raw-ingest (post-normaliser) V/A coherent — so the skew was introduced in the transcoder, not at ingest.

Fix

A per-track cross-track origin snap mirroring the ingest timeline normaliser's CrossTrackSnapMs: when the second track's mapped start diverges from the already-emitting track's output by more than crossTrackSnapMs (1 s), the source clocks are treated as independently based and the late track is snapped onto the early track's output line so both start in lockstep. Skews below the threshold are preserved so a genuine intrinsic A/V offset survives.

The shared offset, releaseAudio gate, and regress re-anchor are unchanged; the common steady-state path is byte-identical (the snap is zero unless triggered).

Validation

• Unit tests: TestCrossTrackSnap, TestRebase_RTMPStartupAudioBeforeVideo_SnapsVideo, TestRebase_CoherentStartup_NoSnap.
• Live: an RTMP-pull stream went from a ~22 s output A/V offset to within a few ms, stable across a 1.5 h soak with no drift or regression; UDP/MPEG-TS streams unaffected.
• make check green (native package built + raced with local libav; lint 0 issues).