Handle OpenRouter API errors and free-model rate limits correctly

Documented steps

1. All errors share one shape: {"error": {"code": <http status>, "message": "...", "metadata": {...}}} — branch on code, then on metadata for detail (e.g. provider_code, moderation error_type).
2. Meanings: 400 invalid params; 401 missing/invalid/disabled key or expired OAuth; 402 insufficient credits — top up account; 403 permissions OR moderation-flagged input (inspect metadata); 408 timeout; 429 rate limited — back off; 502 provider unavailable/invalid response; 503 no provider meets your routing requirements.
3. Free model variants (ids ending in :free) are rate limited: 20 requests/min, and 50 requests/day if the account has never purchased at least $10 in credits — 1000 requests/day once it has.
4. On 429/502, retry with exponential backoff; on 503, loosen provider requirements or choose another model rather than retrying unchanged.
5. During streaming, errors after the 200 OK arrive as SSE events with finish_reason "error" — check for that in stream chunks, not just HTTP status.
6. Docs: https://openrouter.ai/docs/api_reference/limits and https://openrouter.ai/docs/api_reference/errors-and-debugging