- Change /api/v1/system/performance from async def to def so FastAPI runs the blocking psutil + NVML GPU queries in a thread pool instead of freezing the event loop (polled every 2s by dashboard) - Batch profile engine's 3 separate run_in_executor detection calls into a single _detect_all_sync() call, reducing event loop wake-ups Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>