webgl-gate-audit
Procedures for auditing apps/landing TERMINAL VELOCITY WebGL phase gates - driving the browser to exact playhead positions across the 9 scenes, screenshot discipline, scrub + VAT determinism (scrub down THEN rewind to the same playhead -> identical frame), FPS sampling at the risk scenes, draw-call + bundle-size probes vs budget, the blackout->dawn luminance-delta/strobe check at max scrub velocity, copy-parity vs landing/index.html, and reduced-motion + no-GL fallback verification. Load when running /gate or reviewing rendered GL output.
saeedkolivand · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add saeedkolivand/ai-job-hunter-app --skill webgl-gate-audit技能说明
name: webgl-gate-audit description: Procedures for auditing apps/landing TERMINAL VELOCITY WebGL phase gates - driving the browser to exact playhead positions across the 9 scenes, screenshot discipline, scrub + VAT determinism (scrub down THEN rewind to the same playhead -> identical frame), FPS sampling at the risk scenes, draw-call + bundle-size probes vs budget, the blackout->dawn luminance-delta/strobe check at max scrub velocity, copy-parity vs landing/index.html, and reduced-motion + no-GL fallback verification. Load when running /gate or reviewing rendered GL output.
apps/landing WebGL gate-audit procedures (TERMINAL VELOCITY)
The accessibility tree is blind to a canvas, so GL/canvas output is judged only via
screenshots, console, and performance traces - never edit code from this skill, it audits rendered
output only. Semantic-layer gates are the exception: copy parity, real crawlable anchors,
no-GL fallback rendering, and canvas absence are DOM facts, not pixels, and REQUIRE DOM inspection
(accessibility snapshot / eval querying the DOM / DevTools Elements panel), not screenshots -
see Copy parity and Reduced-motion + no-GL fallback render below. Contract + numbers live in
.claude/dormant/skills/webgl-standards/SKILL.md; experience decisions in
docs/knowledge/decision-records/0016-terminal-velocity-scroll-film-landing.md.
Driving the page
Dev server: pnpm --filter @ajh/landing dev (http://localhost:3000). The film is canvas-only once
GL mounts; DOM/accessibility snapshots see nothing but the hidden Semantic layer + a11y overlay.
Scroll = the playhead t in [0,1] over ~3,000 svh (a frozen pixel scroll-track computed once
at mount, never a live vh unit - see webgl-standards). Drive it to a target T:
window.scrollTo(0, (document.documentElement.scrollHeight - innerHeight) * T)
then wait ~2 s for Lenis + scrub damping to settle before screenshotting. The 9 scenes use
half-open playhead ranges [lo, hi) so adjacent scenes never overlap at a shared boundary
(scene 8 alone is closed at both ends, [0.95, 1.00]) - matching webgl-standards' scroll map:
| # | Scene | t range | sample mid-scene T |
|---|---|---|---|
| 0 | Cold open | [0.00, 0.05) | 0.025 |
| 1 | The canyon | [0.05, 0.30) | 0.175 |
| 2 | The surface | [0.30, 0.38) | 0.340 (splash/VAT) |
| 3 | The deep | [0.38, 0.52) | 0.450 |
| 4 | Blackout | [0.52, 0.58) | 0.550 |
| 5 | The catch | [0.58, 0.64) | 0.610 |
| 6 | The ascent | [0.64, 0.85) | 0.745 (ascent fold) |
| 7 | Dawn | [0.85, 0.95) | 0.900 |
| 8 | Finale/credits | [0.95, 1.00] | 0.975 |
Sample mid-scene, not on a range boundary (transition gutters). Confirm the playhead never
hijacks: no wheel preventDefault, no snap, native scroll maps 1:1.
Tools
Preferred: Chrome DevTools MCP (mcp__chrome-devtools) - performance traces, CPU/network
throttling, console with stack traces, screenshots. Fallback if the DevTools MCP is unavailable:
the agent-browser CLI (agent-browser open <url>, eval "<js>", screenshot <path>, console)
plus rAF-counter FPS sampling; mark those checks self-reported. Screenshots stay in the audit
context - report only pass/fail + evidence lines, never raw images.
Scrub determinism (the core gate)
Everything is a pure f(t), so a given playhead must produce ONE frame regardless of direction.
Pick a T inside a scene (not a boundary). Approach it from below (scroll to T-0.02, settle,
then T) and from above (scroll to T+0.02, settle, then T); screenshot both. The frames
MUST match - identical camera pose, character pose, light/grade state, post state. Run it on a
motion-heavy scene (the canyon fall) AND a light-transition scene (the deep). Any drift = a
time-accumulated state or a crossFadeTo leaked in = HIGH failure.
Down-AND-rewind determinism
Scroll forward through a scene into the next, then back to the earlier T; after settle the
frame must match the first pass exactly - no leftover pose, no half-played VAT, no stuck light
state, no doubled particles. The film is reversible end to end; a scene that stays "played" on
scroll-back is a HIGH failure. Exercise it across scene 2->3 (surface into deep) and scene 6->7
(ascent into dawn - the axis inverts here, highest risk).
VAT scrub determinism (scene 2 splash)
At the splash (T ~= 0.34), the crown is a VAT played by sampling at time t. Approach T from
below and from above (as in scrub determinism) AND forward-then-rewind: the crown geometry MUST be
frame-identical each time (VAT sampling is deterministic; inter-frame interpolation must be
symmetric). A crown that differs by approach direction, or that "sticks" at its peak on rewind,
means the VAT time is not pure f(t) = HIGH failure.
FPS sampling (risk scenes first)
Sample the two heaviest scenes explicitly: the splash/VAT (scene 2, T~=0.34) and the ascent
fold (scene 6, T~=0.745, paper-planes-in-formation); then spot-check the canyon storm (scene 1).
DevTools MCP: record a trace while scrubbing the segment, read the FPS track. Fallback rAF sampler
(~2 s, scrub during it):
new Promise(res => { let f=0; const t0=performance.now(); const loop=()=>{f++; performance.now()-t0<2000 ? requestAnimationFrame(loop) : res(Math.round(f/2));}; requestAnimationFrame(loop); })
The governor should hold the tier steady above 45 fps (its downgrade floor); a scene that sits below
45 fps without the governor stepping down a rung is a failure. LOW tier: CPU-throttle 4x (DevTools)
and re-measure; also verify on real budget-Android where possible (fp16 banding / PBR cost).
Draw-call probe vs budget
At each sampled scene read renderer.info.render.calls (expose it, or read via the R3F store in an
`eval