The audit suite

Runnable UI audits under scripts/audit/. They drive the real app in a real browser against a
scratch studio seeded through the real adapters, and they enumerate the live app at runtime:
the room list comes from the dock, deck and view lists from the Library's own chip rows, lens
lists from the editor's select. Add an app, a deck view, or a lens and the next run walks it with
no script edits. The one hand-grown file is walks/promises.ts, because behavior expectations
are human knowledge: every new feature adds a row there (act, then the observable effect).

Runs

Command	What it walks	What it checks
bun run audit	everything below, in order	everything below
bun run audit:rooms	every dock + catalog app, both themes	style audit per room
bun run audit:library	every deck x every view, staging bar, delete confirm, context menu	style audit per state
bun run audit:editors	one piece of every kind in its editor; every write-for lens; binder views; export + book-rules dialogs; the folder inspector	style audit per state
bun run audit:setup	the first-run wizard, every step, live theme flip, custom accent, final	style audit per step
bun run audit:promises	the promise table: act, then observable effect	import opens a picker, drops import and raise the count, delete deletes, staging stages, lenses change the page, search filters, updates answer, theme flips, create mints and opens, duplicate mints a sibling, rename sticks
bun run audit:clicks	EVERY visible enabled clickable in every room, clicked	flags clicks with no observable effect as dead-click suspects (an eyeball list, not auto-fail; Quit/Restart and external links are blocklisted)
bun run audit:legibility:dark / :light	rooms + library + editors pinned to one theme	style audit only
bun run audit:watch	reruns the quick set (rooms + promises) whenever src/ui changes

The style audit (every visited state)

Three exhaustive checks over the full DOM, from page-audit.ts (deliberately DOM-bound: the whole
function is serialized into the page and runs there):

• tiny-text: any text node under the 10px floor
• low-contrast: any text under WCAG AA against its effective background (4.5:1, or 3:1 for
  large/bold text)
• naked-control: any button or select wearing browser-default fonts instead of the house
  tokens (the unstyled-component class of regression)

Page errors thrown during a walk are ledgered too.

Artifacts

Each run writes audit-out/<run>/: one screenshot per visited state (<theme>-<stop>.png) and
ledger.json holding the visited-state list and every finding. The visited list is the
coverage claim: anything not on it was not audited. A run exits non-zero when it found anything,
so runs can gate CI.

Honest boundaries

• A state never rendered is never audited; coverage grows with the walkers, and the ledger names
  exactly what was walked.
• Dead-click detection is heuristic (some no-ops are legitimate); its output is a suspect list
  for review.
• Behavior beyond the promise table is carried by the unit suite, not these walks.