perf: event-driven msg_wait wake-up via asyncio.Event

[bertheto]

Summary

• Replace the 1s �syncio.sleep polling loop in handle_msg_wait with an �syncio.Event-based notification mechanism
• When msg_post inserts a message, it signals the per-thread event, waking up all waiters immediately
• The 1s timeout fallback is preserved via �syncio.wait_for(event.wait(), timeout=1.0) for correctness in multi-process/stdio mode

Changes

• src/tools/dispatch.py (+33 lines, -1 line):
  • Add _thread_events: dict[str, asyncio.Event] registry at module level
  • Add _get_thread_event() helper for lazy event creation
  • In _poll(): replace �syncio.sleep(1.0) with event.clear() + �syncio.wait_for(event.wait(), timeout=1.0)
  • In handle_msg_post: signal _thread_events[thread_id].set() after successful message insertion

Design decisions

• In-memory dict (not DB-backed): Works because SSE mode runs a single uvicorn process. The 1s fallback guarantees correctness if events are missed (e.g., in multi-process stdio mode).
• **
  eload_excludes**: dispatch.py is excluded from hot-reload (
  eload_excludes=[src/tools/*.py] in main.py), so the dict persists across code changes during development.
• Spurious-wakeup safe: The outer while True loop in _poll() re-checks crud.msg_list after every wake-up, so spurious wakes are harmless.

Performance impact

• Before: ~500ms average latency per msg_wait call (uniform distribution over 0-1s sleep)
• After: ~10ms (event signal + context switch)
• Per session (8 msg_wait calls in a typical 4-round debate): ~4s saved

Test plan

• Verified single-worker uvicorn configuration (no --workers flag)
• Verified
  eload_excludes covers dispatch.py
• Manual test: run a multi-agent session and confirm reduced latency
• Verify no regression in stdio mode (fallback 1s polling)