async

Provider boundaries, latency handling, and failure modes that surfaced while wiring Gemini and Ollama into the same product

The polling bug that made completed backend analysis keep looking like it was still running