Adds `canUndo: ReadonlySignal<boolean>` and `canRedo: ReadonlySignal<boolean>`
to `HistoryExtension`'s output, kept in sync via `CAN_UNDO_COMMAND` and
`CAN_REDO_COMMAND` listeners. Signals reset to `false` when the extension is
disabled.

https://claude.ai/code/session_01XeSGvEsnDrdbyLJBGUNYT8

- HistoryExtensionOutput: put canRedo before canUndo (sort-keys-fix rule)
- Flow type: same key order fix
- build(): expand namedSignals({...}) onto multiple lines (was 95 chars)
- Tests: put createInitialHistoryState before delay in configExtension
  objects (sort-keys-fix rule)

https://claude.ai/code/session_01XeSGvEsnDrdbyLJBGUNYT8

…a callback

registerHistory now accepts an optional callback that is invoked once on
registration and again after every mutation of historyState (push, pop, clear).
HistoryExtension uses this callback to keep canUndo/canRedo signals in sync
with the current HistoryState, so the signals are computed directly from the
authoritative source with no dependency on commands.

Removes the previous CAN_UNDO_COMMAND/CAN_REDO_COMMAND listeners. The callback
fires on initialization, which also covers the SharedHistoryExtension case
where historyState may be pre-populated when the child editor is created.

Tests:
- Two updates are required to put an entry on undoStack (first sets current,
  second pushes). makeEditorWithOneUndoEntry helper does both updates.
- The "new edit clears redoStack after undo" test wraps the UNDO dispatch in
  editor.update() so the HISTORIC_TAG from undo's setEditorState does not
  leak into the subsequent edit (which would otherwise discard it).

Verified locally with pnpm run test-unit and pnpm run ci-check (tsc,
tsc-extension, tsc-website, flow, prettier, lint) -- all pass.

https://claude.ai/code/session_01XeSGvEsnDrdbyLJBGUNYT8

…g\` in tests

- Move onHistoryStateChange invocation out of the UNDO/REDO/CLEAR_EDITOR/
  CLEAR_HISTORY command handlers and into the undo / redo / clearHistory
  helper functions themselves, alongside the mutation. The command handlers
  are now thin pass-throughs containing no signal-related code.
- Wrap the canUndo / canRedo writes in HistoryExtension's callback (and the
  disabled-reset path) in batch() so subscribers to both signals are only
  notified once per change.
- Switch all editor variables in the new tests to \`using\` declarations so
  editors are eagerly disposed at end of scope, matching the pattern used by
  other lexical-extension / lexical-react unit tests.

Verified locally with pnpm run test-unit (3045 passed) and pnpm run ci-check
(tsc, tsc-extension, tsc-website, flow, prettier, lint -- all pass).

https://claude.ai/code/session_01XeSGvEsnDrdbyLJBGUNYT8

… effect

These reads are outside any reactive context, so peek() is the appropriate
accessor — it returns the current value without establishing a subscription.

The single remaining .value occurrence is the writable assignment
\`dep.output.historyState.value = populated\` which must stay as a setter.

Verified locally with pnpm run test-unit and pnpm run ci-check.

https://claude.ai/code/session_01XeSGvEsnDrdbyLJBGUNYT8

- Stop wrapping canUndo/canRedo in computed() inside HistoryExtension.build.
  Signal<T> already extends ReadonlySignal<T>, so the explicit
  HistoryExtensionOutput return-type annotation is enough to expose them as
  readonly to consumers — the computed() layer added nothing.
- Drop the now-unused \`computed\` import from @lexical/extension.
- Hoist a single syncFromHistoryState callback in HistoryExtension.register
  that accepts \`HistoryState | null\`. Passing null resets both signals
  (used in the disabled branch); passing a HistoryState derives them from
  its stacks (used by registerHistory on init and after every mutation).
  Eliminates the duplicated batch/reset block.

Verified locally with pnpm run test-unit and pnpm run ci-check.

https://claude.ai/code/session_01XeSGvEsnDrdbyLJBGUNYT8