Every keystroke triggered full JSON serialization of all composer drafts
(including base64 image attachments) and a synchronous localStorage
write. At normal typing speed this caused 5+ writes/sec, blocking the
main thread and creating noticeable input lag.

Wrap the Zustand persist storage with a 300ms debounce. In-memory state
updates remain immediate; only the serialization and storage write are
deferred. A beforeunload handler flushes pending writes to prevent data
loss. The removeItem method cancels any pending setItem to avoid
resurrecting cleared drafts.

Adds unit tests for the DebouncedStorage utility covering debounce
timing, rapid writes, removeItem cancellation, flush, and edge cases.

The useStore subscriber called persistState on every state mutation,
triggering JSON.stringify + localStorage.setItem synchronously. It also
ran 8 localStorage.removeItem calls for legacy keys on every fire.

Wrap the subscriber with a 500ms debounce so rapid state changes batch
into a single write. Move legacy key cleanup behind a one-time flag so
it runs only once per page load. Add a beforeunload handler to flush
the final state.

During active sessions, every domain event triggered a full
syncSnapshot (IPC fetch + state rebuild + React re-render cascade) and
sometimes a provider query invalidation. Events fire in rapid bursts
during AI turns.

Replace per-event processing with a throttle-first pattern: schedule a
flush on the first event, absorb subsequent events within a 100ms
window, then sync once. Provider query invalidation is batched via a
flag. Since syncSnapshot fetches the complete snapshot, no events are
lost by skipping intermediate syncs.

The ProviderHealth layer blocked server startup with two sequential CLI
spawns (codex --version + codex login status), each with a 4-second
timeout, delaying startup by up to 8 seconds.

Run health checks in the background via Effect.runPromise so the layer
resolves immediately with a placeholder status. Add an onReady callback
to ProviderHealthShape so wsServer can push the resolved statuses to
connected clients once checks complete, preventing early-connecting
clients from showing "Checking..." indefinitely.