fix: guard against empty choices and message=None in LLM responses (3 sites)

[qizwiz]

What

Guards response.choices[0] in the FLAT and REBEL agent helpers against two crash vectors — empty choices list and choices[0].message = None:

src/llm_vm/agents/FLAT/agent_helper/requests/call_open_ai.py (2 sites):

# OPENAI_CHAT path
+ if not response.choices or response.choices[0].message is None:
+     raise ValueError("LLM returned empty or filtered response")
  response_text = response.choices[0].message.content

# OPENAI_COMPLETION path
+ if not response.choices:
+     raise ValueError("LLM returned empty choices list")
  response_text = response.choices[0].text

src/llm_vm/agents/REBEL/utils.py (1 site):

+ if not ans.choices or ans.choices[0].message is None:
+     raise ValueError("LLM returned empty or filtered response")
  response_text = ans.choices[0].message.content

Why

The OpenAI API can return choices=[] (HTTP 200, empty list) or choices[0].message=None when content is filtered. Without the guard, the agent crashes with an unhandled IndexError or AttributeError — there's no outer exception handler in these paths that would catch these specific errors gracefully.

Evidence

• Live production crash — confirmed in the wild from the same pattern
• Gemini 2.5 Flash returns choices[0].message = None on PROHIBITED_CONTENT with HTTP 200
• Found by pact across 13.5k violations in 759 real repositories

Scope

2 files, 6 lines added.