B4 — Remote client, hydration, React

Implement B4 of plans/cleanup-state-machines.md ("Phase B — B4"). The plan wins; detailed design in plans/distrbuted-machines.md ("Remote snapshot store and React API"). Prereq: B3 merged.

Scope (still against the fake transport + test-only IPC machine; no production consumer):

1. One RemoteMachineClient per renderer window; remote actor refs (getStatus/getSnapshot/subscribe/dispatch→receipt/resync).
2. Revisioned remote snapshot stores: pre-bootstrap buffering (bounded), monotonic application, gap-triggered resync, newer-actor-instance supersession, disposed-envelope handling (never accept a late snapshot from a disposed actor).
3. Connection lifecycle: reconnect resubscribes and bootstraps; window recreation gets a fresh client; pending dispatch promises settle with a transport-specific result on renderer destruction.
4. React: useDistributedMachine returning {state, projection, connection, dispatch}; explicit connecting/ready/disconnected/incompatible surfaced so UI capabilities can account for it; state never silently fabricated while disconnected (definitions choose the unavailable/initial view); selectors pure and reference-stable via the A1 bindings; remote snapshots never copied into Jotai.
5. Tests: StrictMode replay, provider replacement, unrelated-key no-rerender, dispatch-receipt vs command-completion distinction, all against fake transport and the two-window harness.

Verify: typecheck, full tests, lint, golden suite green. /deep-review. Branch cleanup-b4-remote-client; /pr-push; update plan status. After this PR, Phase B is complete: transport proven, zero product change shipped — C1 is unblocked.