fix(rubrics): read JudgeRubric questions from Pydantic messages

[crazywriter1]

Summary

• JudgeRubric.judge() only extracted question from dict prompts (isinstance(..., dict))
• Rollouts normalize prompts to Pydantic Message objects, so the Question block was silently empty
• Read content via parser helpers (_message_field / _content_to_text) for dict and Message prompts

Test plan

• Dict prompt still yields the last user content as {question}
• Pydantic-like message with .content no longer produces an empty Question
• Multimodal text parts are joined into the question string (via _content_to_text)

---

Note

Low Risk
Small, localized change to prompt parsing in judge rubric scoring with no auth or data-path changes.

Overview
Fixes judge scoring when rollouts pass Pydantic Message prompts instead of raw dicts.

JudgeRubric.judge() previously built the {question} field only when the last list prompt entry was a dict with "content", so normalized rollout prompts left the Question block empty and weakened judge decisions.

The last message’s content is now read through Parser._message_field and Parser._content_to_text, so dict and object messages (including multimodal text parts) populate {question} consistently.

^{Reviewed by Cursor Bugbot for commit 3cd51ce. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix JudgeRubric.judge to extract question text from Pydantic messages

Previously, when prompt was a list, the method assumed raw dicts and fell back to empty strings for Pydantic Message objects. Now uses self.parser._message_field and self.parser._content_to_text in judge_rubric.py to correctly extract the question string from both dict and Pydantic message formats.

^{Macroscope summarized 3cd51ce.}