refactor(viewer): consolidate view state into a store hook

[denolfe]

Consolidates App's nine scattered view-state fields into a single useViewState store hook, and makes createCommands stable by reading state through a latest-state ref. This is a pure structural refactor with no behavior change.

Each view-state field was declared in five places: a useState in App, a read field and a setter in CommandDeps, and an entry in two useMemo dependency arrays. Adding a field was a five-site edit, and a missed dependency array was a silent staleness bug. Commands also rebuilt on every keystroke and scroll because they closed over each field.

Key Changes

• View-state store hook
  • useViewState({ seedVisible }) holds the nine fields in one useState<ViewState> with a stable useCallback writer per field, grouped into an actions object.
  • covering and status stay as local useState. They are App-local UI flags, not part of the shared view-state read model.
• Stable commands via a latest-state ref
  • useLatest exposes the current ViewState as a ref updated on every render. createCommands reads stateRef.current.* and writes via actions.*.
  • The commands memo no longer lists per-field state. It rebuilds only on doc/nav swap and viewport resize, not per keystroke or scroll.
• Narrowed CommandDeps
  • The read/set bag is replaced by stateRef, actions, and a scalar historyDepth (nav-derived, so kept out of the store).
• Docs and tests
  • commands.test.ts builds deps via makeDeps({ state }) and asserts on actions.* mocks. The architecture doc and CLAUDE.md describe the new seam.

Design Decisions

A custom store hook was chosen over useReducer. It keeps named writer methods and the smallest call-site diff, and avoids introducing a second dispatch/action vocabulary alongside the existing keyboard dispatcher.

Command state reads flow through a latest-state ref rather than closures. Within a single command, all reads see one committed snapshot, matching the previous fresh-closure behavior. This frees createCommands from per-field dependencies, so it stops rebuilding on every state change.

[github-actions]

Startup benchmark (--render test/exhaustive.md, linux-x64)

build	mean	ratio	verdict
baseline (main@8cbde71)	740.0ms ± 14.5ms	—
PR	726.7ms ± 18.1ms	0.98×	✅ ok

Thresholds: warn ≥ 1.1×, fail ≥ 1.25×. Baseline built from main.