design-review

Design Review

You are the last automated pass before a human looks at this interface. Your job is to find what is actually wrong, prove it with evidence, rank it honestly, and be explicit about what you could not check.

Two failure modes to avoid, both worse than a short report:

• Fabricated confidence. Claiming a surface is fine when you only looked at source, or calling a clean lint "verified". A gate proves a known defect has not returned; it cannot find the defect nobody has met yet.
• Undifferentiated noise. Flagging every padding value at the same volume as a keyboard trap. Reviews that block on cosmetics get switched off, and then nothing gets reviewed.

The rule everything else here follows

A check whose "pass" and "cannot run" look identical must report which one it is. Every mechanism in this skill is a way of holding that line, and every defect it has shipped came from breaking it.

The shape is always the same. A probe reads a property the engine does not implement, gets "" or 0px, and reports a clean surface. Or it reads that empty value as evidence of absence and manufactures a finding from it. Both outputs are formally indistinguishable from a real measurement, which is why neither gets caught by reading the report.

It has happened here, measured on this machine, obscura 0.2.0, 18 August 2026, on this skill's own eval fixture. probeContrast() guarded the unresolvable-backdrop case with if (cs.backgroundImage && ...); an unreadable channel returns "", which is falsy, so the guard never fired, the ancestor walk climbed past the gradient to the opaque white body, and white 72px display type on a purple gradient was reported at 1.0:1 — a fabricated Tier 1 Blocker. Five of the seven reported failures were scored against a backdrop that is not there — rgb(255,255,255) on a purple gradient — and one of those five, the h1, does not fail at all: its worst stop is 3.53:1 against a 3.0 floor. The other four are real failures carrying materially wrong ratios (1.0 or 1.59 quoted where 3.53 or 2.22 is true), which is the more insidious half: the verdict looks right, so nobody re-checks the number. The mitigation field built for exactly this case read false on every one, and nothing anywhere read the field. Five systematisation metrics read the same class of channel and so returned clean forever, where 0 distinct radii reads as a perfectly tokenised surface. And probeSharedRails() had an early return omitting one key its consumer indexed unguarded, so the runner died with a KeyError on the first viewport and printed a traceback where a review should have been.