perf(runtime): reduce idle mailbox polling

[boadij]

Closes #319

Problem

Managed-Agent request pumps and owner-side result/ask watchers repeatedly polled durable mailbox files every 250 ms, including while idle. This PR replaces those long-lived polling loops with native directory notifications and a one-second reconciliation fallback.

Changes

• Adds one shared observeMailbox() helper using fs.watch() on the mailbox directory, coalesced reconciliation, periodic fallback, watcher reattachment after failure, and explicit disposal.
• Uses the observer for managed-Agent request detection and owner-side result and ask observation; removes the old watchFile() machinery and its retry maps.
• Continues to use the existing durable mailbox readers and request/result/ask identity checks. Notifications are only hints; durable records remain authoritative. Existing acknowledgement, Pi semantic input, result/ask delivery, retry, cancellation, recovery, and cleanup paths remain in place.
• Clears a pending-start status entry after result delivery only when its request ID matches, preventing a fast completion from leaving a stale starting projection without clearing a replacement assignment.

Validation

• npm run check — PASS: 886 passed, 1 skipped, 0 failed ((rerun before commit)).
• npm run build — PASS.
• npm run package:audit — PASS.
• git diff --check — PASS.
• npm run smoke -- core — PASS, cleanup passed.
• npm run smoke -- continuation — PASS, cleanup passed.
• Idle observer measurement: over 10,001.31 ms with no filesystem events, the candidate invoked reconciliation 10 times (59.99 callbacks/minute), with 0 error callbacks. The main baseline of 250 ms schedules 40 checks over the same interval, or 240/minute; this is source-derived, not measured by running the baseline, and excludes the immediate startup pump. CPU usage was not measured. This is a callback-count comparison, not a CPU or filesystem-I/O benchmark.

Remaining validation / risks

The initial GitHub CI run exposed a test-harness issue: the native-watcher test had a pending promise when the event loop became idle on Ubuntu, macOS, and Windows (no assertion failures were reported). A follow-up test-only commit holds a referenced timer while the test waits for notification or fallback reconciliation. The subsequent full CI run passed on Ubuntu, macOS, and Windows, as well as package, container, and scope jobs: run 37935048347. CPU usage and filesystem I/O were not measured. Live smoke was run in the local isolated harness against the current worktree's built extension.

The atomic-publication test verifies eventual reconciliation, which may be triggered either by a native filesystem notification or the one-second fallback; it does not guarantee native notification delivery. Mocked emitMailboxChange() tests separately exercise the notification path.

The mismatched-ID regression confirms an unrelated result is not delivered, but does not directly reach the pending-start cleanup guard: result observation is keyed to the active request ID, and deliverResult() independently rejects a result whose ID differs from the runtime's active/completed request ID. A valid public-flow result therefore cannot pair with a different pending-start ID; directly testing that internal impossible state would require a production-only hook or weakening identity checks. The matching fast-completion regression exercises cleanup during the initial result check.

[boadij]

Addressed External Review #2 in dafe61f:\n\n- Delegation now returns the immutable request ID returned by submit(), even if fast result delivery clears runtime.activeRequestId.\n- Ask observation is registered only while the request remains active after result-observer reconciliation.\n- Extended the fast-completion regression to assert the response request ID and that no mailbox watcher remains after delivery.\n\nValidation: focused regression, npm test -- extension/controller-lifecycle.test.ts (60/60), npm run validate (886 passed, 1 skipped; build and package audit passed), and post-format npm run check (886 passed, 1 skipped). git diff --check passed.