[callumalpass]

Summary

• Refresh Reader development/SDK-refresh documentation and browser-audit selectors without changing dependency pins or the lockfile.
• Delete 165 superseded vendor files (397,165,184 bytes) and the unused historical intake helper. Keep the currently referenced production/LAB/staging SDK archives, metadata/signatures, protocol archive and still-consumed legacy SDK archive. SDK-repin automation remains supported.
• Fix both previously reported P2 product defects: the tab close affordance now calls Reader's guarded dock.close, and shared disclosure menus now escape scrolling ancestors through the same portal/viewport-positioning mechanism as library row menus.
• Preserve native details/summary controls, deferred content, outside-pointer/item dismissal, Escape focus restoration, Tab/Shift+Tab traversal, start/end alignment and existing radio/context-menu arrow navigation. Resize/scroll and asynchronously resized content use the shared positioning hook. Remove the absolute/edge-specific positioning rules rather than adding table-specific or CSS-anchor positioning.
• Correct a stale docking audit: the lone navigator tab is intentionally hidden by existing CSS. Use visible pane actions to expose it before real mouse/native-touch drags; assert drag sources are visible. No forced clicks, weaker assertions, retries or increased timeouts.

Commits

• de2b46b9 docs: explain current Reader development and SDK refreshes
• 376c90be chore: retain only current Reader SDK originals and pins
• 086ebee8 test: align Reader browser audits with current library and editor flows
• baa666ab fix(reader): route the tab close button through the dock's close guard (preserved unchanged)
• a03b73b4 fix(reader): share viewport-positioned portals for disclosure menus
• aac1edf7 test(reader): expose navigator tabs before native docking gestures
• 3fbcfd57 fix(reader): preserve focus reset when a row menu target changes

Validation

All commands below pass on final head 3fbcfd57789fc855b70bd5cdba4780ac0fb20908. Finite code/build gates use rpnpm; browser fixtures run locally in isolated Chromium. Browser runs set TMPDIR="$PWD/.scratch/menu-fix" and PLAYWRIGHT_BROWSERS_PATH="$PWD/.scratch/reader-upkeep/playwright-browsers".

• rpnpm check — lint, Prettier format check, architecture checks/tests, typecheck, 175 publisher tests, workspace tests (Reader: 204 files / 1,125 tests) and pnpm i18n:check all pass.
• rpnpm --pull apps/reader/dist build — full workspace build passes, including i18n and offline-shell verification.
• rpnpm --pull apps/reader/dist --filter @mdbase-reader/app build:feedback-test — configured feedback build passes.
• pnpm --filter @mdbase-reader/app test:browser — all 40 full-audit checks pass with an external 600-second upper bound. A diagnostic completing run also used DEBUG=pw:api; assertions/timings were not relaxed.
• READER_AUDIT_ANNOTATIONS_ONLY=1 pnpm --filter @mdbase-reader/app test:browser — passes, including actual Hide column clicks, start/end alignment, a short scrolling table at the viewport's right/bottom edges, Tab/Shift+Tab, Escape focus restoration and outside-click dismissal.
• READER_AUDIT_READING_ONLY=1 pnpm --filter @mdbase-reader/app test:browser — passes.
• pnpm --filter @mdbase-reader/app test:html-svg — passes.
• pnpm --filter @mdbase-reader/app test:pdf-touch — passes both native-handle and workspace/persistence checks.
• READER_FEEDBACK_TEST_ORIGIN=http://127.0.0.1:8891 pnpm --filter @mdbase-reader/app test:feedback — desktop/mobile acceptance passes; every delivery is intercepted.
• git diff --check origin/main..HEAD — passes; worktree is clean.

At 390px the existing responsive design intentionally hides table column headers. The browser regression checks the reachable shared annotation-view menu at that width, including real action activation and Escape focus restoration. Desktop regressions exercise both column alignments and the table at the actual right/bottom viewport edges. Screenshots were inspected in both cases.

Stall diagnosis and limits

The reported renderer-residency stall did not recur: the full audit visited twenty HTML documents, measured residency rising from one to four and remaining at four, and verified evicted-location recovery. A separate repeatable stall occurred later in mouse.move while the old docking audit dragged a navigator tab hidden by reader-shell.css; the test setup above fixes that unreachable gesture while retaining real mouse and native-touch docking assertions. The final full run passes both.

Early portal-scoped selector/fixture failures were corrected. One tool-runtime reset stopped the owned local server and interrupted remote clients; owned stale leases were cleaned up and interrupted checks rerun. These attempts are not counted as passes.

User-visible behavior: dirty-tab close confirmation works, and shared menus stay available outside scrolling containers and flip/clamp within the viewport instead of using separate hard-coded placements. No UI text, dependency pins, persisted formats, authorization policy or SDK surface changed. No SDK gap or remaining product defect was found in this lane. Existing nonblocking bundle-size/source-map, architecture-size and test-act warnings remain out of scope.

Qualification is Chromium with local disposable fixtures and emulated phone/touch input, not WebKit, native hardware, production or a real collection. No deployment, SDK refresh or real feedback delivery was performed. No merge or review request is included.