[tuanzirwar]

Refs #1362 (provider failure handling, not its additional retry/model-switch suggestions).

Provider creation and iteration errors currently bypass the normal event_generator() completion path. An offline model or a broken response stream can omit the terminal events and leave the disconnect-monitor task running.

Protect both response creation and iteration, emit a user-visible failure through the existing message protocol, and finish with END_LLM_RESPONSE/END_RESPONSE. Close the provider iterator and stop the monitor in finally, including consumer shutdown/cancellation. Preserve partial visible text, but do not save failed output as a successfully completed conversation. No retry or automatic model switching is introduced.

Validation: 18 scoped API-generator tests pass across HTTP/WebSocket, stream=True/False, failure before the first chunk, mid-response failure, creation failure, normal completion, early close and cancellation. Tests check terminal event ordering, monitor completion and save behavior. Original-code failures are recorded separately. A real local TCP HTTP flow through the actual API generator and OpenAI SDK exercises a refused provider connection, truncated chunked SSE after partial text, and a complete successful response. Failures terminate visibly without a successful save; the normal response saves once. Ruff 0.12.3 lint/format and diff checks pass.

AI-assisted implementation and automated checks. Database adapters are isolated; PostgreSQL, a real Ollama server and the complete production application were not deployed.