### What happened?

When a streaming Responses API request uses the MCP gateway (`server_url: litellm_proxy/...`) and the **initial LLM call fails** (e.g. an invalid/expired `previous_response_id` producing an OpenAI 400), the proxy still returns **HTTP 200** and the SSE stream emits the pre-generated `mcp_list_tools` discovery events with **no `response.created` before them**, then ends with `[DONE]`. No `error` event, no `response.failed`.

This violates the Responses API streaming contract (streams begin with `response.created` and end with a terminal event) and hard-crashes SDK stream accumulators. openai-node's `responses.stream()`:

``` Error: When snapshot hasn't been set yet, expected 'response.created' event, got response.mcp_list_tools.in_progress at ResponseStream._ResponseStream_accumulateResponse (.../openai/src/lib/responses/ResponseStream.ts:207:15) ```

The underlying provider error is only visible in proxy stderr (`mcp_streaming_iterator.py: Error creating initial response iterator: ...`); the client has no way to know what failed.

Expected: the request fails with a real 4xx carrying the provider error body (nothing has been streamed yet at that point), or at …