Provider Errors Returned In A 200 Response Body Are Never Classified

In langchain-openai, the _create_chat_result method raises ValueError(response_dict.get('error')) directly instead of routing the error through _handle_openai_api_error. Streaming paths receive an openai.APIError with no status_code because the SDK only sets it from HTTP status, not from the error body. Consequently, _handle_openai_api_error cannot classify these errors, leaving them without is_retryable or ModelError type.

1. Run the provided fake gateway Python script that returns HTTP 200 with body {'error': {'code': 502, 'message': 'Provider returned error', 'type': 'server_error'}} for both streaming and non-streaming requests.\n2. Instantiate ChatOpenAI with base_url pointing to the fake gateway and max_retries=0.\n3. Call llm.invoke('hi') and observe a ValueError with the error dictionary but no is_retryable attribute.\n4. Call llm.stream('hi') and observe an openai.APIError with no status_code and no is_retryable attribute.

This fix changes error types from ValueError to ModelError subclasses, which may break existing code that catches ValueError for these cases. The existing test test__construct_lc_result_from_responses_api_error_handling expects ValueError and must be updated to expect ModelError. The openai.APIError constructor signature may vary between SDK versions; if the SDK changes, the helper method may need adjustment. In concurrent environments, ensure that the status_code extraction and assignment are thread-safe (the current approach mutates the exception object, which is local to the thread). Rollback recommendation: revert to the original raise ValueError line if unexpected regressions occur, but be aware that error classification will be lost.