Summary

Seven root-caused fixes for the browser-matrix remainder (run 4012 baseline: 28 failures) plus the canary CIWA gap — every finding agent-diagnosed with probe or arithmetic evidence before any code was written:

• Service-worker navigation hijack (the big one — playwright.config.ts + tests/ui-pwa.spec.ts): in production builds public/sw.js registers in every matrix test (pwa-lifecycle.tsx registers whenever NODE_ENV === "production"), clients.claim()s the page seconds in, and serves every subsequent navigation. A Chromium probe proved SW-served reloads bypass page.route interception entirely, and Playwright-Firefox wedges on exactly this path — its only two page.reload() calls in the whole suite are the two 60-second ui-smoke hangs (2206 and the @critical 2932). Fix: serviceWorkers: "block" in the shared config; ui-pwa.spec.ts — the one spec whose subject is the worker — opts back in with test.use({ serviceWorkers: "allow" }), and its offline/CacheStorage privacy journey passes locally under the opt-in.
• WebKit mode-menu dismissal (src/components/clinical-dashboard/master-search-header.tsx): dismissal relied solely on a wrapper focusout; WebKit's sequential focus navigation can move focus nowhere (links are excluded from its Tab order — the previous tabbable is the skip link) or wrap backward into the open menu, so no wrapper-escaping focusout ever fires. Fix: close on Tab in the trigger's keydown handler (APG menu-button pattern; arrow keys remain the entry). Agent-verified no existing test depends on the old forward-Tab behavior; pointer/Escape/blur paths unchanged.
• WebKit forced-colors capability (tests/ui-accessibility.spec.ts): WebKit has never implemented forced-colors; Playwright plumbs the emulation flag but the @media (forced-colors: active) remap under test cannot engage. Fix: webkit test.skip in the repo's established capability-guard style. Chromium/Firefox keep full coverage; the sibling generic forced-colors usability test stays unguarded (it passed on WebKit).
• WebKit phantom axe contrast (tests/ui-formulation.spec.ts): the reported #b5bbc2 at 1.93:1 is byte-exact the 0.4-opacity disabled composite of --text-muted on white — but the Previous button is provably enabled at scan time and axe skips truly disabled controls. Only consistent mechanism: WebKit serving a stale :disabled computed style after React re-enables the button. Fix: pin toBeEnabled() + toHaveCSS("opacity", "1") before the scan — heals the stale style or converts the failure into direct proof (fallback documented in-test). WCAG 1.4.3 exempts disabled controls, so no design change is warranted.
• Firefox tab order (tests/ui-smoke.spec.ts): Firefox includes scrollable containers in the tab order; the sheet body (overflow-y-auto, measured genuinely overflowing at the test viewport) sits between Close and "New chat" in DOM order. Fix: conditional step-over in the test (an app-side tabIndex={-1} on the shared Sheet primitive would remove Firefox's keyboard-scroll affordance — a product decision deliberately not taken here).
• Canary CIWA content gate (scripts/eval-retrieval.ts): canary runs Save Codex local changes #50 and Refactor monolithic front‑end and decouple domain logic #51 rank the CIWA-Ar dosing-table region in the top 5, yet the content gate reports "ciwa OR score OR threshold" missing — because textContainsClinicalTerm is whitespace-delimited and the scale's written name is the hyphenated token "CIWA-Ar", which the bare fixture term can never match. Fix: ciwa: ["ciwa", "ciwa-ar"] in clinicalContentAliases — the eval's own documented drift-absorption mechanism (plan-authorized fixture-alias route). The clinical substance of the expectation is unchanged.
• Shared stub helper (tests/helpers/zero-touch.ts + six specs): extracts test(e2e): report real zero-touch capability to WebKit for command-surface specs #995's six inline maxTouchPoints stubs into one helper — the CodeRabbit review follow-up.

Verification

• Targeted vitest (eval-retrieval, search-command-surface) — 25/25
• npm run test — 3012 passed; sole failure is the long-baselined container-only pdf-extraction-budget artifact
• npm run verify:cheap — all static checks, lint, typecheck green; same single artifact in the test step
• npm run build + client-bundle secret scan — pass
• Chromium ui-pwa under the new SW config — the cold-offline/CacheStorage privacy journey passes with the allow opt-in; the installability test shows only the known container-only in-incognito artifact (hosted-CI-green since feat(pwa): retirement kill-switch, offline-version binding guard, and dev teardown flag #826)

UI verification not run: the local Chromium UI suites carry documented container artifacts (16/20 ui-universal-search failures identical on unmodified main; Playwright browser-build mismatch after the dependabot bump) — hosted ui-critical/ui-advisory on this PR plus the post-merge matrix dispatch are the authoritative UI verdicts. npm run eval:retrieval:quality cannot run locally (live keys exist only in Actions); the alias change is validated by the next canary run — the failing case's evidence chain (#50/#51 logs) is recorded in the ledger row.

Risk and rollout

• Risk: low-moderate, dominated by the config change. serviceWorkers: "block" returns every non-PWA test to the pre-July-17 (pre-sw.js) behavior they were written under; the worker's own coverage is preserved via the ui-pwa opt-in and the untouched Vitest SW policy suites. The mode-menu Tab-close is a one-branch keyboard refinement matching the APG pattern; remaining changes are test-side or the eval fixture alias.
• Rollback: plain revert of the content commit; the ledger commit is documentation-only.
• Provider or production effects: None from merging. The post-merge matrix dispatch is $0; the alias is exercised by the next scheduled canary.

Clinical Governance Preflight

• Source-backed claims still require linked source verification before clinical use
• No patient-identifiable document workflow was introduced or expanded without explicit governance approval
• Supabase target remains Clinical KB Database (sjrfecxgysukkwxsowpy)
• Service-role keys and private document access remain server-only
• Demo/synthetic content remains clearly separated from real clinical sources
• Source metadata, review status, and outdated/unknown-source behavior remain conservative (the eval alias recognizes the CIWA-Ar scale's written name; it does not relax any expectation, threshold, or ranking behavior)
• Deployment classification/TGA SaMD impact was checked when clinical decision-support behavior changed (mode-menu keyboard dismissal is navigation chrome; no clinical output surface changed)

Notes

• Expected next-matrix delta vs the 28-failure baseline: the two Firefox reload hangs, the Firefox tab-order case, the WebKit menu/forced-colors/formulation trio, plus whatever the test(e2e): report real zero-touch capability to WebKit for command-surface specs #995 stub already cleared. Explicitly not claimed: ui-stress:409, ui-tools:2087, and the WebKit answer-flow subset — the post-merge dispatch measures those; anything surviving gets its own diagnosis round.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EXsJcLrbZUXwnBeG91cVo9

Generated by Claude Code

Summary by CodeRabbit

• Bug Fixes

  • Improved desktop keyboard navigation so mode menus close correctly when tabbing away.
  • Improved focus handling in mobile search flows, including clipped layouts and scrollable content.
  • Corrected clinical content matching for CIWA-A terminology.
  • Improved validation of step-navigation controls and accessibility behavior across browsers.

• Tests

  • Increased cross-browser reliability for PWA navigation, touch detection, forced-colors checks, and UI smoke tests.