[miguel-heygen]

What

hyperframes check had no audio section, so a film whose sound the render drops, refuses or plays as silence passed with 0 errors (#2775). It now reads every clip the render mixes, from the same media list the render builds, and reports:

Code	Severity	When	What render does (checked on a real render)
audio_source_rejected	error	the clip's file has no audio track	stops at compile: "composition asset(s) do not match their authored media element type (expected: audio)"
audio_source_unreadable	error	the file does not decode as audio	the audio mix fails and the render stops: "Audio mix failed; output would be video-only"
audio_starts_past_source_end	warning	data-media-start is at or past the end of the file	the clip plays nothing
audio_clip_silent	warning	the whole part of the file the clip plays is silent	the clip plays silence
audio_none	info	no clip in the film plays audio (an unmuted video with an audio track counts, as in render)	a film with no sound

Findings carry the clip's #id, the composition file it was written in, and its time on the film. A 3-scene film with one defect per scene, on main and on this branch:

main:   ◇  Check passed   (exit 0)

branch:
Audio
  ⚠ audio_clip_silent: audio/vo-0.wav is silent for the whole part this clip plays (0-3.5 s of the file).
    compositions/s0.html #s0-vo t=0.5s
  ✗ audio_source_rejected: audio/vo-1.mp4 has no audio track, so render stops before it starts.
    compositions/s1.html #s1-vo t=5.5s
  ⚠ audio_starts_past_source_end: audio/vo-2.wav is 3.5 s long but the clip starts 9 s into it, so it plays nothing.
    compositions/s2.html #s2-vo t=10.5s
  1 error(s), 2 warning(s), 0 info(s)
◇  Check failed   (exit 1)

Design

• Options: (a) metadata checks only by default, the silence decode behind a flag; (b) decode every clip in full by default; (c) by default, sample a long clip at three 1 s spots and stop at the first one with sound, reading the whole played part only when every spot is quiet, while the browser pass runs.
• Picked (c): a silent clip is the defect Audio pipeline: reused narration silently dropped + voice-clip seam collisions — both pass every check #2775 describes and a flag nobody passes would not catch it; a full decode of a 6-minute bed cost about 0.4 s on a CPU-only box, and sampling brings a clip with sound down to one short decode. A clip is only called silent after its whole played part is read.
• Not added: sample-rate or channel checks (render resamples every source), loudness targets (taste, and hyperframes audio reports them), overlap across tracks (music under a voice is normal; same-track overlap is already lint's duplicate_audio_track).

One owner per decision: the bundle and render media list are built once and shared by the HDR inspection and the audio pass (the HDR inspection now takes that list instead of bundling its own); the media-type rule is render's own mediaProfileMatchesElementType, now exported from the producer; the probe is the engine's cached extractAudioMetadata that render uses; the decode and silence measure are the ones hyperframes audio uses.

Timing

check --json wall time, built CLI, CPU-only Linux VM (shared, so each main/branch pair runs back to back after the load drops), 3-composition 15 s film and a 6-scene 6-minute film, each with a music bed and one voice per scene.

Fixture	main median	branch median	delta	median paired delta	budget (max of +10%, +150 ms)
3 compositions, 15 s	1948 ms	1983 ms	+35 ms (+1.8%)	+38 ms	195 ms
6 scenes, 6 min	3755 ms	3808 ms	+53 ms (+1.4%)	+26 ms	376 ms

9 pairs per fixture, every run exit 0, branch reports carry an audio section with no findings (the fixtures' clips all sound). A first version that decoded every clip in full measured +695 ms (+11.5%) on the 6-minute film, over budget; that is what the spot sampling fixes.

Tests

• checkAudio.test.ts (10): real WAV and PNG files through the real probe, render's media-type rule, decoder and silence measure: a sounding clip passes; rejected source; unreadable source; start past the end; silent played part (only that clip, only the part it plays, playback rate respected); a long clip is only called silent after a full read; composition file named; video sound counted.
• check.test.ts (+3): an audio error fails the check and lands in its own section; audio is still checked when lint errors skip the browser; an audio pass that throws becomes an info finding, not a failed check.
• Red on main: the 3 new pipeline tests fail and checkAudio.test.ts cannot import. One deliberate break per check turns exactly its test red. 3 green runs in a row.

Fixes #2775

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
hyperframes	🟢 Ready	View Preview	Oct 26, 2026, 4:14 PM

💡 Tip: Enable Automations to automatically generate PRs for you.

[github-actions]

Edit accuracy: accurate 271 (base branch 271), smooth 166 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-px-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 · 0 important · 2 nits · head faf489da

• nit — Comments (not required) fails: the comment share in checkFindings.ts rose from 2/42 to 3/61 because findingAtRoot moved there with its doc line.
• nit — checkPipeline.ts inspectProjectMedia: if the browser pass throws, the outer catch returns without awaiting audio. The probes and decodes keep running after the failure report until they finish. That's harmless (the promise has a .catch), but it can hold the process open for a moment.

Scope

Read all 13 files and checked the claims against render. compileStage preflights compiled.audios, video -audio tracks included, with mediaProfileMatchesElementType("audio"), so audio_source_rejected matches render. An audio_processing_failed warning fails the render (renderOrchestrator.ts:838), which matches audio_source_unreadable. Locally: cli check, checkAudio, hdrPromotion and checkPipeline tests pass 200/200. I ran a built-CLI check on a fixture: a silent clip and a start past the file's end were flagged; a clip with sound only near its end, and a sounding clip, were not. CI is green apart from Comments.

[heygengenesis]

Verdict: APPROVE — 0 resolved, 0 open, 0 new · faf489d → 28a0e54. This head only drops one comment line, and that fixes Comments. All CI is green on this head.