[miguel-heygen]

What

Music keeps playing when a preview video swaps to a new copy. The video shows a still of its last frame until the copy can draw. Before, the whole film stopped, music included, for about 200 ms.

Why

Each transport tick stops the film when a clip in view has no data (mediaNotReadyAt, then haltPlayback). A copy swap empties the video on purpose, so the swap itself counted as a stall. The clock and every playing track stopped, then restarted once the copy could draw.

How to reproduce: a film of muted video clips over one continuous audio clip, with a swap during play, such as a change of preview quality. A level meter on the audio element reads silence at the swap.

Related work

#6922 added the still that stands in during a swap; this PR lets the film rely on it. A swap of a video that has no frame to show still waits, as before.

How

• stillStandsIn (in proxySrc.ts): true while a video shows its held still, for up to 2 s. mediaNotReadyAt skips such a video, so the film plays on.
• After 2 s the copy counts as a real stall and the film waits with Buffering, as it does for any clip with no data.
• heldPosters now also keeps when the hold began.

Known limit: the copy lands slightly behind the film (0.16 s in the run below), since the clock kept moving while it loaded. I did not measure how quickly that closes before the next cut.

Test plan

• Unit tests added: init.buffering.test.ts
  • The music plays on while a video shows a still across a copy swap.
  • A copy that never lands stops standing in, and the film waits for it.
• fails without the fix: packages/core/src/runtime/init.buffering.test.ts › the music plays on while a video shows a still across a copy swap
• Removing the stillStandsIn check fails the first test; removing its time limit fails the second.
• packages/core: init.buffering.test.ts and mediaProxy.test.ts passed 3 runs in a row; vitest run src/runtime 2702 passed; tsc --noEmit clean; oxlint and oxfmt on the changed files.
• Manual testing performed: the desktop app on Linux with no GPU, 10 back-to-back muted HEVC clips over one mp3 (49.5 s, made from repo fixtures), the runtime built with and without this change, a quality step-up swapping every video during play, a level meter on the music.
  • Without the change, 2 runs: the music paused once at the swap in each, with 184 and 538 ms of silence.
  • With the change, 2 runs: no pause during play and no silence at the swap. The shown video held its still for about 180 ms, then drew.
  • Separate from this change: in 1 of the 2 runs with it, the music stayed at its first moments for about 2 s after Play and was re-seeked several times while the film ran, 30 s before any swap and with no still held. I have not found the cause.
• Not run on macOS or Windows.
• Comments follow CONTRIBUTING.md "Comments"

[github-actions]

Edit accuracy: accurate 271 (base branch 271), smooth 124 of those

The gate passes.
Smoothness is reported in the artifact, not gated. A case fails only if it fails 2 of 3 runs.

Quarantined, measured but not gated (0)

Unstable (1)

• seqplainkeys-none-pct-r0-root-z100: tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo false, teleport true / tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true / tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true

[heygengenesis]

Verdict: APPROVE · 0 blocker · 1 important · 0 nits · head 50a927b0

The Comments ratchet fails, but it is not a required check. See the inline comment.

Scope

I read all 4 files. The only places heldPosters is set and released are in mediaProxy.ts, and both moved to the new {poster, since} shape. mediaNotReadyAt is the only stall gate, so stillStandsIn is the only bypass. A still is held only when the old copy had a frame, so a video with nothing to show still waits.

The fails without the fix: line holds: with !stillStandsIn(el) removed, the named test fails (1 failed, 5 passed); with it back, all 6 pass.

The ~0.16 s lag the PR lists as a known limit gets closed by rate steering, not a seek. The swap evicts the sync state, and 0.16 s is under the 0.5 s hard-sync threshold. That matters only for a swapped video that has its own audible track.

CI, CodeQL, captures, regression, Windows, Player perf and Safari audio are green. Player Safari perf was still running when I reviewed.

[heygengenesis]

important: Comments fails here. proxySrc.ts goes from 14.3% comments (2/14) to 17.4% (4/23). Cutting this line is enough to pass, because the constant's name already says it.

[somanshreddy]

Approving at 5b91b744 on Miguel's re-stamp ask, after my own pass over the full diff. The only change since Genesis's approval at 50a927b0 is one deleted comment line in proxySrc.ts.

• stillStandsIn: excludes a video showing its held still from mediaNotReadyAt for at most 2 s, measured from heldPosters.since. So a proxy swap no longer stops the clock and every playing track, while a copy that never lands still falls back to Buffering after 2 s. Both directions are pinned in init.buffering.test.ts: stand-in is ~1 s of play with the music unpaused, and expiry with since 60 s back gives a held clock and Buffering.
• Type change: moving heldPosters from string | null to { poster, since } is applied at both readers: unheldPoster and the letGo restore in mediaProxy.ts.
• Nits, non-blocking:
  • A re-swap while a still is already held keeps the first since (holdShownFrame returns early on heldPosters.has). So a panel that settles over more than 2 s of swaps reaches the Buffering fallback sooner than "2 s per swap" suggests.
  • The ~0.16 s lag matters only for a swapped video that plays its own audio, as stated in the PR.

— Somu