Follow-ups: dart-run replay CLI via pure-Dart i18n split + page-level a11y audit

[albertzhzhou-droid]

Summary

Closes the two concrete follow-ups documented in #89.

1. dart run Local-AI replay CLI — finally unblocked

The CLI had been documented as future work in three consecutive PRs because the
recommendation orchestrator transitively imported Flutter via app_i18n.dart.
The actual Flutter surface there turned out to be just three members, so this
PR splits them out instead of refactoring i18n broadly:

• lib/core/i18n/app_i18n_context.dart (new) — the BuildContext.appI18n
  extension (AppState provider lookup) and appI18nLocaleFor (tag → Locale),
  re-exporting app_i18n.dart. The 13 UI files + app.dart now import this;
  app_i18n.dart itself is pure Dart. No behaviour change (same logic,
  same listen: false semantics).
• tool/run_recommendation_scenario_replay.dart + .mjs wrapper +
  npm run recommend:replay — replays the five archetypes with the same
  offline scripted Local AI stand-in as the test harness, writes
  build/recommendation_scenario_replay/latest.{json,md}, and exits non-zero
  iff the candidate-set invariant is violated.
• Verified byte-identical artifacts between the CLI and
  flutter test test/local_ai_replay_report_test.dart (md5-compared), so the
  two entry points cannot drift from each other.
• Docs updated: the "Why there is no dart run CLI" section is replaced with
  CLI notes; the reviewer order in PUBLIC_VERIFICATION.md now starts with
  npm run recommend:replay.

2. Page-level accessibility audit (was blocked on asset loading)

test/accessibility_semantics_test.dart gains a real CatalogPage widget
test: a stub CachingAssetBundle serves a valid 1×1 PNG plus a
StandardMessageCodec-encoded AssetManifest.bin, so the brand images render
under flutter_test; a local-mode AppState is constructed in a guarded zone
(sqlite init is unavailable on the test VM and irrelevant); a phone-sized test
viewport avoids layout overflow. The test asserts the accessible "view details"
labels are exposed by the real page's semantics tree — upgrading the previous
construct-level checks to a page-level audit.

Validation (all run, all green)

• dart format --set-exit-if-changed . clean; flutter analyze no issues
• flutter test --concurrency=1 → 745 passed
• npm run recommend:replay → 5 scenarios, invariant held, exit 0
• CLI vs test artifact: byte-identical (json + md)
• npm run public:preflight 0 BLOCKER; npm run privacy:preflight 0 blocker
• git diff --check clean

Remaining follow-ups (intentionally untouched)

• Language-tagged alias model (LocalizedAlias) — only if per-locale alias
  filtering becomes a requirement; flat lists still suffice.
• Kana below the catalog engine's CJK threshold — widening _hasCjk would
  shift existing confidence assignments; deliberately left alone.

Educational prototype only; synthetic/demo data only; no medical advice; not
calibrated for real care.

🤖 Generated with Claude Code