Close the three fault-injection gaps: 8 of 8

[pftg]

Follow-up to the audit (#572) and the gate fixes (#574). The three defects the suite could not see are now caught, each verified by injecting the defect, watching it fail, and reverting — per the rule #574 added.

D — frozen tenure (was blind until 2027-01-01)

The rendered assertion cannot tell derived from frozen while both read "18+". Added the missing half: the stub must use derived: tenure, not a literal. The hardcoded number IS the defect class — it is how the founding year ended up wrong in eight places on 2026-08-14, each with its own copy.

Injection → the tenure stat must be \derived: tenure` so layouts/next/landing.html computes it from site.Params.foundingYear`. A second injection froze a different stat, proving the second assertion is live rather than shadowed by the first.

G — the PurgeCSS trap (the audit called it unguardable)

It is unguardable by the visual gate, which builds production and therefore purges exactly as production does. A unit test catches it: classes styled in the register's source CSS ∩ present on a rendered element − present in the fingerprinted bundle that page links. Non-empty = a silently-purged rule. Keyed on the mechanism, not on <th>.

Injection → flags rr-th-theirs. The control is what makes it believable: the same class moved onto a <td> — an element writeStats does record — goes green. It flags the purge, not merely a new class.

Costs three file reads; the unit suite already builds production.

B + the contrast hole — computed styles, not screenshots

Deliberately not per-band screenshots: those inherit the tolerance problem (font/SVG-heavy bands need a measured floor) and still return a pixel delta rather than a ratio.

1. The "theirs" column must compute the register's own --*-muted token. Injection → computes rgb(27,28,28), not --ed-muted ... the comparison stopped comparing (#564).
2. A WCAG walk over every visible text run on the whole page — viewport-irrelevant, so it closes gap B — compositing translucent layers, 4.5:1 text / 3:1 large / 3:1 control fills, reporting unresolvable cases instead of guessing white, and asserting it measured >50 pairs so an empty walk cannot pass.

Injections caught: the audit's above-fold defect A at 2.31:1; a below-fold defect at y=1090 that the screenshot gate reports as 0 difference; and a 1.12:1 button fill under SC 1.4.11.

This also closes the hole the audit flagged separately: defect A was previously caught only as an anonymous pixel change, indistinguishable from an intentional recolour. Nothing measured contrast until now — and this session shipped a real 3.33:1 button that only human review caught.

Worth reading in the report

The first version of the muted-column assertion ("differs from the sibling cell") passed the injection — the injected ink was a third value, so it found a difference and approved it. Two green runs on a wrong test, exposed only by breaking it. That is the case for the rule.

Gates

unit 289 runs / 6187 assertions / 0 failures in 1.42s (was 287/6143/1.33s) · critical 38 runs, 55 screenshots compared, no failures, fixtures untouched · hugo-build clean · standardrb clean.

Honest limits

The contrast walk runs at 1920×1080 only (a mobile-breakpoint-only colour is uncovered) and covers the three register pilots, not the whole site. Both are loops, not redesigns.

🤖 Generated with Claude Code

https://claude.ai/code/session_011SP5gaqXEgUie8pdFrmbeJ

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 3cf6bb0e-fed8-4d77-a4e5-22bdfde02b76

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}