fix(responses): remove reasoning_item from the boundary set, and name the training-variant error

[ananthsub]

Part of #2452

Two corrections to the Responses item plumbing. Neither changes behaviour at the current openai pin.

1. _RESPONSE_OUTPUT_BOUNDARY_TYPES listed "reasoning_item", which is not a Responses output item type at any openai version.

The class is ResponseReasoningItem. Its type tag is "reasoning": https://github.com/openai/openai-python/blob/650be393dedf2a4550092817c2b82c1d04d6e9dc/src/openai/types/responses/response_reasoning_item.py#L34

which is what the responses converter emits:

Gym/nemo_gym/responses_converter.py

Lines 412 to 418 in abc21eb

	reasoning_item = NeMoGymResponseReasoningItem(
	id=f"rs_{uuid4().hex}",
	type="reasoning",
	summary=[
	NeMoGymSummary(text=reasoning_text, type="summary_text") for reasoning_text in reasoning_matches
	],
	status="completed",

2. RESPONSES_TO_TRAIN[item_cls] directly looks up the dict in the responses converter. It was a bare dict access, so an unregistered class raised KeyError and reached the client as a 500 with no explanation. training_variant_of() raises NotImplementedError naming the class and what to add.

Nothing reaches that path today. The lookup is passed response_output[-1] from postprocess_assistant_message_dict, whose local list can only hold NeMoGymResponseReasoningItem, NeMoGymResponseOutputMessage or NeMoGymResponseFunctionToolCall, and all three are registered. The point is that the next converter to emit a new item type gets a message telling it what to add rather than a stack trace.

3. responses_to_chat_completion_create_params matches on each item's type and has a case for four of them. Everything else falls to case _, which raised Unsupported message type: {the whole item}:

   Gym/nemo_gym/responses_converter.py

   Lines 174 to 184 in a62c300

   	match m["type"]:
   	case "message":
   	self._format_message(m, state)
   	case "reasoning":
   	self._format_reasoning(m, state)
   	case "function_call":
   	self._format_function_call(m, state)
   	case "function_call_output":
   	self._format_function_call_output(m, state)
   	case _: # pragma: no cover
   	raise NotImplementedError(f"Unsupported message type: {m}")

That message reads as a gap in the converter. It usually is not. Most types that reach it have no Chat Completions representation at all, so the transcript needs a model server that passes Responses through rather than a new case here. openai_model does; vllm_model and inference_provider downconvert and cannot.

The message now names the type and says which of those two situations it is. It no longer prints the item, because the payload can be large or opaque, and the type tag is the part that identifies the problem.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.