a11y(desktop/pdf): struct-tree text layer + annotation tool shortcuts (#316)

[agentfleets]

Closes #316

The last two open items from the issue's remaining tail: expose PDF prose to screen readers, and single-key annotation-tool shortcuts.

What

• Struct-tree accessible text layer (PdfCanvas.tsx): pdfjs-dist 4.10's API-layer TextLayer has no struct-tree wiring (upstream that lives in the viewer's StructTreeLayerBuilder, which we don't bundle). This replicates its minimal core:
  • Text is now streamed with includeMarkedContent: true, so each struct element's run is wrapped in an id'd span.markedContent (zero-size anchors; layout, selection, copy, and the search-highlight <mark>s are unchanged — textLayer.textDivs still yields only the text spans).
  • page.getStructTree() is walked into a hidden span tree (faithful port of upstream's #walk: PDF role → ARIA role map, H<n> → heading/aria-level, alt → aria-label, lang, single-content-child collapse, aria-owns onto the marked-content ids) and appended inside the <canvas> as fallback content — never painted, but exposed to assistive tech in reading order.
  • The page wrapper changes role="img" → role="region" with a localized "Page {n}" label (like the upstream viewer): img would mask all descendants as presentational, hiding the prose.
  • Untagged PDFs resolve a null tree → nothing appended (same as upstream). The split-view mirror pane skips its tree to avoid duplicate-id prose exposure.
• Tool shortcuts: window-level keydown (V/Escape = select, H highlight, U underline, N note, A area, I ink), inert in inputs/textareas/contenteditable and with modifiers held, so the find box, page input, comment textarea, and Ctrl/Cmd zoom/undo keep their keys. Each tool button's tooltip now shows its key, localized via a new read.annToolTitle format key (en (H), zh full-width （H）).

Why

• Screen readers previously got only "Page N, image" — no prose. Tagged PDFs now read structurally (headings, lists, tables, figures with alt text).
• Keyboard users couldn't switch annotation tools without the mouse.

Trade-offs / notes

• This is the pragmatic-equivalent path the issue allows: the struct-tree layer is hand-wired against pdfjs-dist 4.10's getStructTree()/includeMarkedContent contract rather than importing the upstream viewer bundle (which would drag ~430 kB of pdf_viewer into the chunk for one class). Upstream's #addImageInTextLayer (positioned role=img spans for figures inside the text layer) is not ported — figure alt text is still exposed via aria-label on the struct node.
• Untagged PDFs remain prose-inaccessible to AT (as in the upstream viewer); the prior role="img" page label is superseded by the role="region" label.

How verified

• npm run build (sync:tokens + tsc --noEmit + vite build) — green.
• bash ../scripts/lint-desktop-tokens.sh — clean (no new off-token/phantom tokens; CSS additions are color-free).
• git diff origin/main --stat — only the 3 intended files.
• Walk logic smoke-tested against a stub DOM (heading levels, role map, single-child collapse, null-tree); the marked-content id ↔ struct-tree aria-owns contract verified against the bundled pdfjs-dist 4.10.38 sources (worker emits ${pageObjId}_mc${mcid} ids on both sides). No Rust changes.

[physercoe]

APPROVE. Struct-tree/markedContent positioning faithfully mirrors pdf.js, overlay-bleed guard preserved, keydown listener uses the ref-latch pattern (no stale closure), and it fixes a pre-existing hardcoded aria-label. Rebased onto main and resolved the i18n + drilled-t→local-useT() merge (aligned PageView with main's new local-useT idiom). Minor to fold in later: includeMarkedContent: !readOnly so the split-view mirror pane does not emit duplicate DOM ids.