feat(accessibility): add terminal screen reader mode

[LarryHu0217]

Summary

• add an opt-in Terminal screen reader mode setting, disabled by default
• persist and strictly validate the setting across app restarts
• initialize new xterm instances with screenReaderMode and update open terminals when the setting changes

Validation

• npm run check
• npm run check:static
• npm run test:security-rules
• npm test (1,633 passed, 23 skipped)
• npm run build:frontend

Implements the terminal screen-reader mode portion of #211.

[johannesjo]

Important — include the new preference in the autosave snapshot

I double-checked the current head and reproduced this against persistedSnapshot(): once the initial startup autosave has completed, changing only terminalScreenReaderMode leaves the snapshot unchanged, so setupAutosave() does not schedule its 1-second saveState().

Normal window close explicitly awaits saveState(), so clean restarts are covered. However, a crash or hard termination before another tracked/direct save can revert an enable or resurrect a previously enabled value.

saveState() already serializes the field at src/store/persistence.ts:186. Please add terminalScreenReaderMode: store.terminalScreenReaderMode beside terminalFont in src/store/autosave.ts, plus a regression test in src/store/autosave.test.ts asserting that toggling it changes persistedSnapshot().

The xterm constructor/live-update wiring and strict load validation otherwise check out, and CI/full tests pass.

[LarryHu0217]

Addressed the autosave gap in 9c27792. persistedSnapshot() now tracks terminalScreenReaderMode beside terminalFont, and the regression test verifies that toggling the preference changes the snapshot.

Validation: 96 focused store tests and 1,634 full-suite tests passed; compile, typecheck, ESLint, Prettier, Knip, and dependency-cruiser passed.

[johannesjo]

Reviewed at 9c27792, rebased on current main (208a933). LGTM — merge as-is; everything below is a nit.

Verified

• npm run check clean (compile + typecheck + eslint + prettier); full vitest suite: 1630 passed, 25 skipped, 0 failed.
• Touchpoint parity. The new field hits exactly the same 11 sites as terminalFont — both type decls, core.ts default, saveState, LegacyPersistedState, loadState validation, persistedSnapshot, ui.ts setter, store.ts re-export (needed — SettingsDialog imports from ../store/store), TerminalView construction + effect, settings row. There's only one loadState assignment path and no settings-reset/export site, so nothing is missed.
• The runtime toggle genuinely works. @xterm/xterm@6.1.0-beta.195 registers optionsService.onSpecificOptionChange('screenReaderMode', …) inside open(), and _handleScreenReaderModeOptionChange creates/disposes the AccessibilityManager, guarded on !this._accessibilityManager.value — so re-setting the same value is a no-op and open terminals update live.
• No selection/click regression. .xterm-accessibility is an inset:0; z-index:10 overlay, but ships pointer-events: none, and terminal selection here goes through term.getSelection() / getSelectionPosition() rather than window.getSelection(). Bookmarks, copy and the selection-anchored markers are unaffected.
• Effect ordering is sound. It reads store.terminalScreenReaderMode before the if (!term) return, so the dependency registers on the first run, and onMount (l.397) is registered ahead of the effects (l.1097+) with no early return before new Terminal, so term always exists by then.
• Persistence shape matches the coordinatorControlHintDismissed precedent (|| undefined on save, strict typeof === 'boolean' on load), and the test volume matches the house pattern for a persisted boolean.

Nits

1. src/store/autosave.test.ts:33 — the new case sits inside describe('autosave snapshot includes new-task-default fields'), but screen-reader mode isn't a new-task default. Rename the block (e.g. "autosave snapshot tracks persisted settings") or move the case out. The test itself earns its keep: it guards precisely the regression 9c27792 fixed, and persistence.test.ts can't catch it because it calls saveState() directly rather than going through the dirty check.

2. src/store/ui.test.ts:202 — this asserts a one-line pass-through to setStore with no logic in between; the other blocks in that file all cover setters that actually do something (focus clearing, size clamping), and this one also forces a field onto MockStore just to exist. Dropping it loses no coverage (−13 lines). Optional.

3. src/components/TerminalView.tsx:1111 — the effect deliberately omits markDirty() unlike its font/theme neighbours, which reads like an oversight at a glance. Worth a one-line why. It is correct as written: AccessibilityManager's constructor ends with _refreshRowsDimensions(); _refreshRows(), so toggling the setting on an idle terminal populates the current viewport immediately with no repaint needed.

4. src/remote/AgentDetail.tsx:208 — the phone/remote client builds its own Terminal with no access to the desktop store, so that surface stays silent to assistive tech. Separate app, fair to defer, but worth noting Reduced-motion guards, opt-in terminal screen-reader mode, tooltip upgrade #211's problem statement isn't fully closed by this PR.

Behavioural caveat (not a code issue)

The accessibility tree mirrors only the viewport of the active buffer, and the live region cuts a batch off at 21 lines with a "too much output" message. Claude Code runs on the alternate buffer and redraws continuously, so announcements will be chatty during streaming. Arrow-key navigation still reaches scrollback via xterm's boundary-focus scrolling. That's inherent to the xterm approach the issue asked for — just calibrating expectations, nothing to change here.