fix: guard against empty choices and message=None in make_api_call

[qizwiz]

What

Add explicit guards in make_api_call() before accessing response.choices[0].message.content in both the is_final_answer=True and is_final_answer=False branches.

Why

client.chat.completions.create() (Groq's OpenAI-compatible API) can return two empty-response shapes:

1. choices = [] — on content-policy rejections, rate-limit errors, or provider failures
2. choices[0].message = None — e.g. when content is filtered, some OpenAI-compatible providers return HTTP 200 with the message field absent

Both crash with IndexError or AttributeError inside the retry loop. The surrounding except Exception as e: catches these crashes but produces a confusing "Failed to generate step after 3 attempts. Error: list index out of range" message, masking the actual cause.

With the explicit guards, the ValueError is raised before the attribute access, caught by the existing retry logic, and reported accurately.

Change

# Before
return response.choices[0].message.content

# After
if not response.choices or response.choices[0].message is None:
    raise ValueError("LLM returned empty or filtered response")
return response.choices[0].message.content

Applied to both the final-answer branch and the step-reasoning branch.

Corpus context

Detected by pact (llm_response_unguarded mode), a Z3-verified static analyzer for LLM crash vectors. This pattern was found across 13.7k violations in 786 repos.