feat(ci): the UI audit can now see what nav FORGOT, not just what it wrote

[catomean]

The audit's two detectors both look for something someone put on the page: ink below the contrast floor, rows starting at different x. An entire class of navigation defect is structurally invisible to that shape, because each instance is something someone left out.

A fleet-wide nav audit across 20 repos (~13,300 lines of navigation, nothing shared between them) found three, repeatedly.

The three detectors

1. A current page no link announces. Every repo styles the active item; nine never set aria-current, and four more set it on every surface but one. The highlight exists only for people who can see it.

2. Nav targets below 44px. That's this fleet's own floor, not the WCAG 2.2 AA minimum of 24px — fleetcrown enforces it centrally, kivvi and wild-spirit state it explicitly. So the message says "out of step with the fleet", not "non-compliant".

3. A label that looks like a control but isn't. The orangecat sidebar shipped an <h3> beside a chevron-only button carrying the whole onClick. Visitors aimed at the word and nothing happened.

Rendered, not grepped — and that's the point. This fleet holds Next apps, CSS modules, Tailwind, and one hand-rolled static generator with no framework at all. No source-level lint spans that. The DOM does.

Each detector refuses to fire where the absence is correct

• aria-current is demanded only when a link in that nav actually points at the page being rendered. A footer of outbound links has no current page to mark; flagging it would fire on every site that has one.
• A dead label is claimed only when the heading sits outside any control and a sibling control has no text of its own. That pairing is the bug; a heading beside a labelled button is not.
• Nested controls are measured by their parent, so one small icon inside a large row isn't reported as the row.
• Small targets print one line per distinct size, not per element — a nav of twelve identical 32px links is one decision, and twelve lines of it buries everything else.

Tests

Seven new self-tests, both sides pinned as the existing ones are. 19/19 pass. The failing side of each is a real bug from a real repo; the passing side is the corrected markup.

Lands while the workflow still defaults to --warn-only, so findings surface in the job summary before anything fails on them.

Also: SHARED.md

Records navigation chrome under "What must NOT be centralized" — with evidence rather than opinion. sitekit is the one shared nav renderer, serves 2 of 20 repos, and shipped ~28px targets and no focus style to both consumers, unfixable downstream. Its flat {path,label} model can't express solon's mega-menu or reparaturbonus-zh's drawer, which is why every repo with real nav complexity reinvented its own.

Plus the six-rule contract that is shareable, and where each rule is enforced (rendered audit for public pages; a repo-local check for anything behind a login — disjoint surfaces, neither substitutes for the other).

One reversal recorded honestly: the audit first recommended putting nav duplication on the ratchet. Writing the "do not centralize" section showed the two can't both be right — the ratchet counts concerns that should converge on one implementation, and every repo is supposed to have its own nav config.

Unrelated pre-existing failure

shared-inventory.sh --check currently exits 1 on health-route: 9 → 10. Not from this PR — the inventory scans remote default branches, this diff touches only local docs and audit scripts, and the baseline file is untouched here. A tenth health route landed somewhere independently and needs either adoption or a deliberate --update.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XeELB8b3N4JrT2asYL9WvE