feat(a11y): engage the audit enforcement ratchet (#21)

The point of the Phase 0 harness finally lands: named findings in
textClipped, dynamicType, hitRegion, elementDetection,
sufficientElementDescription, and trait now FAIL the build on the
empty-state suite. Regressions in five phases of accessibility work are
gated, not narrated.

Three carve-outs, each earned by evidence rather than convenience:

- contrast stays report-only. The two long-standing Settings findings are
  rows scrolled under the translucent tab bar, and their attribution
  flips between a row name and nil run-to-run — no suppression is narrow
  enough to keep CI stable. The centralised palette is the real guard.
- The seeded dense-row tests run reportOnly. Bisection showed the audit
  degrades on children-ignored content — the correct VoiceOver treatment
  for dense rows — emitting unattributed contrast/dynamicType failures on
  rows that measure 6-7:1 and render correctly.
- Characterised noise is suppressed narrowly and always logged as
  [noise: reason]: disabled controls (WCAG 1.4.3 exempt), "nearly passed"
  near-misses, system field placeholders (flagged at any length — proven
  by shortening them to no effect), and unattributed clipped/dynamic-type
  artifacts. noiseReason(for:) records each rule's provenance inline.

Validated in both directions. Positive: the full 11-test suite passes
with enforcement live, dark and light, on an erased simulator. Negative:
re-injecting the Phase 3 icon-exposure regression produced two named
[FAIL] findings and a failed suite — on both screens sharing the
component — then went green again on revert.

Docs and the CI workflow comment updated to describe the engaged state.

1 files changed, 32 insertions, 21 deletions

diff --git a/Docs/ACCESSIBILITY.md b/Docs/ACCESSIBILITY.md
index b35f02c..3f8b3ba 100644
--- a/Docs/ACCESSIBILITY.md
+++ b/Docs/ACCESSIBILITY.md
@@ -110,27 +110,38 @@ Users override it under Settings → Display.
 Dark mode reports 18 findings and light mode 21; the three extra are the section
 headers above. Everything the app actually controls passes in both schemes.
 
-## Findings are reported, not enforced
-
-The audit surfaces violations that exist today, so failing on all of them would
-block every unrelated change until the whole pass lands. Instead, findings are
-logged and attached to the result bundle tagged `[report]` or `[FAIL]`.
-
-Enforcement is the committed constant
-`AccessibilityAuditHarness.enforcedAuditTypes`. Widen it as each phase clears a
-category:
-
-| After phase | Enforce |
-| --- | --- |
-| 2 — semantic colors + light mode | `.contrast` |
-| 3 — Dynamic Type + reflow | `.textClipped`, `.dynamicType`, `.hitRegion` |
-| 4 — VoiceOver | `.elementDetection`, `.sufficientElementDescription`, `.trait` |
-
-A constant rather than a CI setting, for two reasons. Environment variables do
-not work: neither a plain `xcodebuild` env var nor a `TEST_RUNNER_`-prefixed
-build setting reaches the UI test process, so the toggle silently did nothing.
-And a committed value makes "when did contrast become enforced?" answerable with
-`git blame` instead of CI tribal knowledge.
+## Enforcement — the ratchet is engaged
+
+With phases 1–5 landed, `AccessibilityAuditHarness.enforcedAuditTypes` enforces
+**`.textClipped`, `.dynamicType`, `.hitRegion`, `.elementDetection`,
+`.sufficientElementDescription`, `.trait`** on the empty-state test suite. A
+named finding in any of these fails CI — regressions in five phases of work are
+now gated, not merely reported.
+
+Three deliberate carve-outs, each with its evidence:
+
+1. **`.contrast` stays report-only.** The two long-standing Settings findings
+   are rows scrolled under the translucent tab bar; their attribution flips
+   between a row name and nil run-to-run, so no suppression is narrow enough to
+   keep CI stable. The centralised palette in `Shared/Colors.xcassets` is the
+   actual guard against contrast regressions.
+2. **The seeded tests run `reportOnly`.** Bisecting the row/badge accessibility
+   modifiers showed the audit degrades on `children: .ignore` content — the
+   *correct* VoiceOver treatment for dense rows — emitting unattributed
+   contrast/dynamicType failures on rows that measure 6–7:1 and render
+   correctly. Their burndown still prints; it just doesn't gate.
+3. **Characterised noise is suppressed narrowly and always logged** with a
+   `[noise: reason]` marker — disabled controls (WCAG 1.4.3 exempt), "nearly
+   passed" near-misses, system field placeholders (clipped at any length —
+   proven by shortening them to no effect), and unattributed
+   clipped/dynamic-type artifacts. Nothing disappears silently; see
+   `noiseReason(for:)` for each rule's provenance.
+
+Enforcement is a committed constant rather than a CI setting, for two reasons.
+Environment variables do not work: neither a plain `xcodebuild` env var nor a
+`TEST_RUNNER_`-prefixed build setting reaches the UI test process, so the toggle
+silently did nothing. And a committed value makes "when did clipping become
+enforced?" answerable with `git blame` instead of CI tribal knowledge.
 
 ## Why coverage is split between local and CI