#453 LLM tool asyncio.run() causes worker zombie state via anyio corruption

In llm.py lines 573-577, when no event loop is running: ```python try: asyncio.get_running_loop() except RuntimeError: return asyncio.run(coro) # No timeout! ``` When anyio (used by httpx) has corrupted internal state (e.g., after sandbox fallback), asyncio.run() can hang indefinitely because: 1. anyio callbacks fail with 'no running event loop' 2. HTTP request never completes 3. asyncio.run() waits forever **Evidence from logs:** - 10:01:07 asyncio error: RuntimeError: no running event loop -> InvalidStateError - Workers stopped claiming tasks entirely after this - LISTEN thread still ran (daemon), making workers appear 'alive' - 590 pending + 7000+ completed tasks, zero being processed **Root cause:** Mixing sync sandbox fallback -> direct execution -> LLM async call corrupts anyio state **Fix options:** 1. Add timeout to asyncio.run() path (wrap in thread with timeout) 2. Ensure httpx client lifecycle is properly managed 3. Add watchdog to detect zombie main loop