[miguel-heygen]

What

bun run playbench in packages/player: plays a composition in headless Chrome and prints the numbers for five player problems reported by a user's detailed report. It ships two fixture edits generated on our own media: a 3-take edit (24.8 s, 30 fps, 6 media elements) and a 10-take edit (53.08 s, 25 fps, 20 media elements).

No visible change

Test tooling only; the one non-test edit adds a playbench script to packages/player/package.json. No Studio or player behaviour changes.

Why

The fix lanes for first play, audio joins, automation and the autoplay-blocked fallback need one shared before/after measurement, so each fix is judged on the same numbers.

Related work

None open on this surface.

How

• Follows tests/browser/player-box.ts (standalone bun script), tests/perf/runner.ts (launchBrowser) and tests/perf/browsers/make-media.sh (media generated at run time, gitignored _media).
• instrument.js is injected into every frame: rVFC on every video, a per-rAF sample against __player.getTime(), AudioParam calls wrapped, and an AudioWorklet on the destination reporting which stretch of the file is heard.
• analyze.ts is pure, covered by 6 vitest tests (no real timers).
• --audio sink|none|both, --edit primary|tentake|both, --runs N, --blocked (unmuted play() rejects, which exercises the parent-proxy fallback).
• Run with a real sink (PulseAudio null sink, 48 kHz) and without one. Fixture media has a keyframe every second.

Printed per problem: overrun frames at cut 1 after first play; joins with a mid-play audio seek in the first 600 ms and their jump sizes; automation landing offset; incoming second-frame gap and playhead hold per cut; takes heard in the blocked fallback.

Test plan

• Unit tests added: vitest 6/6, three runs in a row; dropping the visibility filter fails a test
• Manual testing performed: full runs on a Linux box, Chrome 156, with and without a sink
• Documentation updated (not applicable)
• Comments follow CONTRIBUTING.md

Baseline tables on origin/main follow in a comment once the final runs finish.

[heygengenesis]

Verdict: REQUEST CHANGES · 1 blocker · 1 important · 0 nits · head 8ae2ce78

• blocker: the required check PR captures / Studio and player captures fails. The body has no Before/After sections and no No visible change section. Everything else in the diff is under tests/, so the only watched change is 5 lines in packages/player/package.json. Adding a ## No visible change heading to the body should clear it.
• important: problem 5 uses the wrong clock (↓ inline).

Scope

Read all 9 files end to end. I simulated the make-media.sh marker tones against the worklet's Goertzel decode over both files: 96–99% of 0.1 s windows decode to within ±0.75 s, which is enough for the >50% rule. CI at review time: CI typecheck/lint/Test (support) green, Player perf green, PR captures red, Safari perf/Windows/CodeQL still running. No prior reviews.

— Genesis

[github-actions]

Edit accuracy: accurate 271 (base branch 271), smooth 165 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 — 2 resolved, 0 open, 0 new · 8ae2ce78 → fc5ef49c

— Genesis