Verify 02: opening group (clean-room)

ETS27 prototype, round 2 · 2026-10-13 · scope: prototype/index.html, prototype/shared.css and the fork's hand-off into prototype/mind.html

Stance: clean-room. This pass used the finding list, the fixer's claims and skipped items, the brief and the design direction. It did not open the fixer's report (studio/fix-02-opening.md) or the critique files. Every verdict below comes from fresh pixels and measurements taken in this pass or from the code, never from the fixer's own screenshots (r2-*).

Verdict

Score: 78 / 100 for the opening group as shipped now.

Owner decisions 1, 2 and 3 are in place and correct: exact fork copy, accessible names, keyboard contract, dated lockup and a single raspberry accent with no retired ember left in any page. The realm-loading and fallback engineering is solid. Self-hosted three.js matches the pinned hashes, the DOM never waits for the realm, and stalls, failed imports, context loss and moderately slow GPUs all land on the CSS realm with the phase clock carried over. Draws are capped at 30 fps after the intro settles.

Two things keep this from a high score:

1. Decision 5 (the head) is partial. The skull defects are gone: no dark sockets, streaks, drips or teeth. The minimum luminance on the head rose from 117 to 164, and no pixel is under 150. But the eye hollows are oversized, and the 0.60 shading floor flattens their interior into a uniform grey plateau. They read as flat grey patches more than soft mannequin dishes.
2. The slow-device probe fails open on the slowest devices. Those are the phones this probe exists for.

Smaller issues:

• The 10-second nudge overflows common laptop viewports.
• Small phones (375x667 and 360x640) overflow.
• Fast typing on the fork keeps only the first character.
• Muted text sits right at the AA line over drifting clouds.

Method

• Headless Chrome for Testing (chromium-1234) through the Playwright library, with the GPU on (ANGLE Metal, --ignore-gpu-blocklist) and deviceScaleFactor 1. No headed browser was used.
• Pages were served by a static python3 -I -m http.server 8704 --bind 127.0.0.1 from prototype/. The server was stopped by exact PID at the end, and port 8704 was confirmed free.
• Zero live Claude calls. server.py (8777) was never contacted. mind.html was exercised only in offline demo, through ?demo=1 or the failed /api/health probe on the static server. The network log shows zero POST requests. No visitor text was logged anywhere.
• Instruments used:
• DOM and computed-style reads.
• Pixel probes on the saved PNGs: luminance histograms, dark-run scans under the cut, border and text contrast against the darkest adjacent background.
• Real keyboard sequences.
• Route interception to stall or abort three.js.
• requestAnimationFrame throttling to emulate slow GPUs.
• Forced SwiftShader and --disable-gpu browsers.
• WEBGL_lose_context.
• A drawArrays counter for the 30 fps cap.
• Every screenshot listed at the end was read as an image before it was used as evidence.

Owner decisions

Fixer claims

Skipped items: the socket glow, lockup hierarchy, poster still, SVG wordmark, return-visitor intro and the --ets-ember deletion were checked against the code and are absent as the fixer stated. Two notes:

• The --ets-ember alias now has no users in any page, so it can go (N8).
• The fixer said the ink focus-ring alternative is recorded as a deviation in its report. That was not verified: opening the fixer's report would break the clean room.

Blocker and major findings

Minor and nit findings, status in brief

• Palette cream, heavy hairlines, 800 headings, forced caps: fixed. The sky is white and the hairline is #e2e2df. Interactive outlines use --line-ui on purpose so they pass 3:1.
• Notation and microcopy drift: fixed on the opening. Corner tags and the hint are lowercase; [ skip ] and [ shortcuts off ] use the bracket notation. The mind and standard parts belong to the other group.
• Phone composition (two items): fixed at 390x844 and wider. Partial on short phones (N4).
• Lockup hierarchy: skipped as an orchestrator call. The block mark stays.
• Skip affordance and wordmark: skip fixed. The wordmark is still unavailable.
• No-JS blank sky: fixed.
• Single-key shortcuts scope and off switch: fixed.
• Interactive outlines under 3:1: fixed, measured at 3.45-3.53:1.
• Focus ring colour: raspberry, per the direction. The ring is 2 px with a 4 px offset and has no white halo; it remained visible in every capture.
• Leftover tokens: blue deleted and ember renamed. A dead --ets-sun: #e8a06a and two unused aliases remain (N8).
• Frame governor cannot see slow devices: partial (N1).
• No Subresource Integrity on the CDN import: fixed by self-hosting, with hashes verified.
• Retired ember in marks and hovers: fixed.
• Render loop housekeeping: fixed.
• Cached prefix tokens and production hygiene: out of this group and not verified here.

New findings from this pass

N1 (major). The frame-time probe fails open on the slowest devices.

The probe only falls back when it has at least 4 samples (index.html 937). Its window starts on the first loop call (919), and that call includes shader compilation. So a device whose first frame takes more than 1.5 s, or whose frames take more than about 375 ms, ends the window with 0-3 samples and passes.

Evidence:

Only real devices whose browsers report a major performance caveat are caught, as --disable-gpu was.

Fix:

• Treat fewer than 4 samples in the window as slow, or extend the window until 4 samples arrive.
• Start the window after the first rendered frame.
• Add a wall-clock guard: if the realm clock lags wall time by more than about 1 s after 3 s, fall back.

N3 (minor). The 10 s nudge overflows common laptop viewports.

.nudge is absolutely positioned below the hint (index.html 162). Its bottom edge lands at:

In each case the line is clipped and the page becomes scrollable after 10 s (v2-opening-vp-1280x720-nudge.png, v2-opening-vp-1024x680-nudge.png). It fits at 1440x900, but only with 1 px to spare.

Fix: show the nudge in place of or inside the hint row, or reserve its line inside the stage's spare bottom row, then re-measure scrollHeight after it appears.

N4 (minor). Small phones overflow.

The crown of the head sits under the corner tags, and the hint and toggle fall below the fold (v2-opening-vp-375x667.png, v2-opening-vp-360x640.png).

Fix: below about 700 px tall, shrink the head (raise lookYTall, increase halfH), tighten the row gaps and let the spare bottom row collapse to zero.

N5 (minor). Fast typing keeps only the first character.

Typing who at 40 ms per key on the fork lands w in the composer. The other keys are lost during navigation and before mind.html focuses #q after the fixture and health awaits. The storyboard (§2) promises "their words".

Fix: buffer keys on the fork for about 300-400 ms (until a pause) before navigating, then append the buffer. Separately, have mind.html focus #q synchronously at the top of its script and append any keydown that arrives before focus.

N6 (minor). Muted text is borderline AA over drifting clouds.

The minimum measured is 4.47:1 (see F13).

Fix: darken --muted slightly (about #5c5b57), or put a faint white veil behind the fork and hint band.

N7 (nit). The no-GPU path logs two three.js console.error lines before the CSS realm takes over.

Fix: probe with a bare canvas.getContext('webgl2', {failIfMajorPerformanceCaveat: true}) before constructing the renderer.

N8 (nit). Dead tokens.

--ets-sun: #e8a06a, --ets-ember and --ets-glow are used by no page. The realm uses its own shader constants and the CSS realm's own rgba halo. Delete them so the retired colour cannot creep back.

N9 (nit). [ shortcuts off ] shows on touch-only devices, where there are no shortcuts (v2-opening-390x844-touch.png).

Fix: hide it under (hover: none) and (pointer: coarse) until a keydown is seen.

N10 (nit, cross-group). Copy mismatch on mind.html. The unavailable-state door card (mind.html 684) uses the em-dash explainer; the fork uses the colon form from the decision. Align them.

N11 (nit). Small direction deviations:

• The door B socket glow on hover (§3.3) is not built; the fixer skipped it.
• The door B top hairline is 2 px rather than 1 px.
• The eyebrow is 12 px rather than mono-m 14 px.

Head: what would finish decision 5

The probe proves the darkness problem is solved; the remaining issue is shape and gradation.

1. Lower the floor (0.60 → about 0.50-0.52) and add an up-facing hemisphere or bounce term, so the lower half of each dish lifts toward 200 and the interior keeps a gradient instead of a plateau.
2. Narrow the hollows. Socket radius x 0.14 → about 0.11, centre x 0.27 → about 0.25, ssub k 0.16 → about 0.10. They should stop well inside the silhouette.
3. Check the turned profile at yaw ±9°: the temple cutter plus socket currently leaves a notch on the turned side.
4. Re-run the same probe. Expect no histogram spike within 2 levels of the floor, and keep zero pixels under 150 and zero dark runs under the cut.

Regressions against round 1

Compared with opening-1440x900.png, opening-390x844.png, opening-mid-intro-1440.png, opening-no-webgl-1440.png, opening-reduced-motion-1440.png and c2v-opening-head-1440.png:

• No regressions in unclaimed areas. Checked: the realm and cloud deck, the theme line, the corner tags, the lockup position against the head, the mid-intro beat and the no-WebGL layout. Corner tags are now visible from frame 0, which matches the storyboard. The mark is slightly smaller (96 px cap) to fit short screens.
• One claimed property breaks over time. Short laptop viewports fit without scroll at settle, then scroll once the 10 s nudge appears (N3).

Console and network

• Opening flows had zero console errors, zero failed requests and zero HTTP errors. These covered settled, mid-intro, focus, hover, ?still=1, reduced motion, 390x844 (with and without touch), 1024-1920 widths and resize.
• Expected warnings only, one per forced fallback (for example [ets27] CSS realm: forced via ?nowebgl=1).
• No-JS runs show two blocked modulepreload requests. That is how Playwright disables scripts, not a page fault.
• On mind.html served by the static server, the /api/health 404 is an environment artifact (no API on port 8704), and the page falls back to demo as designed.
• Forced-failure runs logged what they forced: the aborted request, and three.js context errors under --disable-gpu.

Not verified

• Real phone GPUs, Safari and Firefox, and a real screen reader. The accessible names were read from the DOM.
• The fixer's report contents (clean room).
• Server-side items (prefix token count, telemetry, frame-ancestors).

Captures (prototype/shots/, all new, prefix v2-opening-)

• Core: 1440x900, mid-intro-1440x900, mid-intro-b-1440x900, focus-1440x900, focus-map-1440, hover-mind-1440, hover-map-1440, still-1440x900, nowebgl-1440x900, 390x844, 390x844-touch, nojs-1440x900, nojs-390x844, reduced-motion-early-1440, head-1440, head-still-1440.
• Intro and keys: skip-visible-1440, after-skip-1440, focus-early-1440, shortcuts-off-1440, nudge-1440, handoff-slash-mind-1440, handoff-draft-mind-1440.
• Robustness: stall-1440, slow-device-1440, context-lost-1440, swiftshader-1440.
• Viewports: vp-1024x680, vp-1024x680-nudge, vp-1280x720, vp-1280x720-nudge, vp-1366x768, vp-375x667, vp-360x640.
• Decisions 6 and 7 (mind.html, offline demo): decision6-mind-first-answer-1440x900, decision6-mind-first-answer-390x844, decision7-mind-after5-1440x900, decision7-mind-after5-390x844.