### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

Since v1.90.0, a non-streaming request to a `chatgpt/*` model is silently upgraded to a streaming request, so the client never gets the single JSON body it asked for. This is specific to the `chatgpt` provider; `openai/*`, `azure/*` and the other Responses providers are unaffected.

`llm_http_handler.py` decides whether the caller wanted streaming from the client's own `stream` field. v1.90.0 (commit cfcdf871, PR #30202) added a second term to that decision in both the sync and async Responses handlers:

```python stream = response_api_optional_request_params.get("stream", False) # what the client asked for ... data = responses_api_provider_config.transform_responses_api_request(...) ... stream = bool(stream or data.get("stream")) # added in v1.90.0 ... is_stream_request = bool(stream) ```

`data` at that point is the provider payload, not the client request. `ChatGPTResponsesAPIConfig.transform_responses_api_request` sets `request["stream"] = True` unconditionally because the Codex backend only serves SSE, so `data.get("stream")` is alway…