chore: sync main with staging

[vrtornisiello]

Summary

Make agent runs survive client disconnects, and persist a meaningful row when a graceful shutdown forces cancellation.

• Decouple agent execution from event streaming: The route handler now spawns the agent run as a background task and returns a StreamingResponse that forwards events from an asyncio.Queue. If the client disconnects, only the consumer is cancelled — the producer keeps running and persists its result.

• Track in-flight runs and drain on shutdown: The lifespan registers each producer task in app.state.running_runs and, on SIGTERM, waits up to SHUTDOWN_DRAIN_TIMEOUT_SECONDS for them to finish naturally, then cancels and awaits the rest so their finally blocks and done-callbacks complete before loguru and the engine are torn down.

• Persist interrupted runs distinctly: On cancellation, the producer catches CancelledError, sets the user-facing INTERRUPTED message and a new MessageStatus.INTERRUPTED (only if no final_answer has been observed — otherwise the real status is preserved), and re-raises. A dedicated status keeps shutdown-driven cancellations separable from real agent errors in queries and metrics.

• Other fixes and improvements along the way: Persistence failure now surfaces via complete.error_details; exception strings no longer leak into client payloads; run_id is consistently typed as str across boundaries; Created model_call_limit event type and MODEL_CALL_LIMIT status to distinguish from a success final answer.

Notable design choices

• Producer-crash hang is by design. If the producer dies before its finally reaches queue.put(complete), the consumer hangs and the frontend eventually sees a timeout error. We deliberately don't synthesize a complete in _cleanup because it would mask the bug and hide real schema-drift errors.
• Producer opens its own DB session rather than reusing the request-scoped one, since it outlives the HTTP request.