Bug: OpenAI realtime path returns empty success (score=1.0, output=None) on a refusal / content-filtered response

[thejesh23]

Describe the overall issue and situation

In the OpenAI provider's realtime (non-batch) inference path, a refusal or content-filtered completion is returned as a successful, empty extraction instead of raising.

langextract/providers/openai.py:252-254 (_process_single_prompt):

response = self._client.chat.completions.create(**api_params)
output_text = response.choices[0].message.content
return core_types.ScoredOutput(score=1.0, output=output_text)

message.content is None when the completion stops for content_filter, when a structured-output refusal is emitted (text lands in message.refusal), or when the message carries only tool calls. ScoredOutput.output is typed str | None, so None is accepted and returned with score=1.0. Downstream the resolver parses None into an empty/garbage result reported as success, and the refusal reason is silently discarded — no exception, no signal to the caller.

This is inconsistent with the batch path for the same provider, which already validates exactly these cases. langextract/providers/openai_batch.py:182-190:

if content is None:
    refusal = message.get('refusal')
    if refusal:
        raise exceptions.InferenceRuntimeError(
            f'OpenAI batch response refusal: {refusal}', ...)
    raise exceptions.InferenceRuntimeError(
        "OpenAI batch response body missing 'message.content'", ...)

The realtime path should apply the same guard.

Expected behavior

When choices[0].message.content is None (content filter, refusal, or tool-only message), the realtime path raises InferenceRuntimeError — surfacing the refusal reason when present — mirroring the batch path. An empty/missing completion should never be reported as a score=1.0 success.

Actual behavior

_process_single_prompt returns ScoredOutput(score=1.0, output=None). The caller receives a "successful" extraction with no content; the refusal reason (message.refusal) is lost.

Steps to reproduce the issue

1. Configure OpenAILanguageModel and issue a prompt that trips OpenAI's content filter, or set response_format and prompt something the model refuses.
2. The API returns choices[0].message.content == None with .refusal set (e.g. "I can't help with that.").
3. Observe _process_single_prompt returns ScoredOutput(score=1.0, output=None) instead of raising — the extraction is reported as a successful empty result.

Any additional content

• Suggested fix: after fetching the message, guard on empty choices, and if message.content is None raise InferenceRuntimeError including getattr(message, 'refusal', None) when present — matching openai_batch.py:182-190.
• The Gemini realtime path has the same class of gap on a safety-blocked response.text (version-dependent); worth folding the same guard in there too.
• Repro/verify without live API: mock chat.completions.create to return a message with content is None and .refusal set, and assert _process_single_prompt raises InferenceRuntimeError (mirrors the existing batch refusal test).
• Environment: current main (commit 0dff547).