fix: guard against None content in JudgeRubric judge response

[vominh1919]

Problem

In verifiers/rubrics/judge_rubric.py:98, the judge response is extracted with:

judge_response = str(judge_response.choices[0].message.content)

When the judge model returns tool_calls or a refusal (e.g., content filter), message.content is None. In Python, str(None) produces the literal string "None", which is then parsed as the judge's scoring response. This causes:

1. Silent incorrect scoring — the judge "evaluates" the string "None" instead of the actual response
2. No error signal — the caller has no idea the judge failed; it just gets a wrong score
3. Hard-to-debug training issues — if using this for RL training, the wrong rewards silently corrupt the training signal

Fix

Check for None content and raise a clear RuntimeError with diagnostic info:

content = judge_response.choices[0].message.content
if content is None:
    raise RuntimeError(
        f"Judge model returned None content. "
        f"This usually means the model returned tool_calls or a refusal. "
        f"Model: {self.judge_model}"
    )
judge_response = str(content)

The error is raised inside the existing try/except Exception block (line 126), so it will be caught and logged properly with the model name.

Before vs After

Scenario	Before	After
Normal response	Works correctly	Works correctly
content = None	str(None) = "None" → silent wrong score	RuntimeError → logged + propagated
Tool calls only	str(None) = "None" → silent wrong score	RuntimeError → logged + propagated

Tests

This change only affects the error path when content is None. The happy path (non-None content) is completely unchanged. Existing tests that mock the judge client with valid responses will continue to pass.

---

Note

Medium Risk
Changes judge scoring behavior to raise an exception when the model returns None content (e.g., tool calls/refusals), which may surface new runtime failures for callers that previously received a silent "None" string.

Overview
Prevents silent mis-scoring in JudgeRubric by validating the judge model response before parsing.

When choices[0].message.content is None (commonly due to tool calls or refusals), the code now raises a descriptive RuntimeError instead of converting None to the literal string "None" and caching/returning it.

^{Reviewed by Cursor Bugbot for commit 2e8644b. Bugbot is set up for automated code reviews on this repo. Configure here.}