fix(rubric): enforce structured output for judge via tool_choice

[AUTHENSOR]

Summary

The rubric judge in evaluate_rubric (docent_core/docent/ai_tools/rubric/rubric.py) parses the LLM judge output via json.loads(output.first_text) at two sites (the streaming callback at line 199 and the return path at line 286). The LLM call does not enforce structured output — no response_format, no tools, no tool_choice. The agent transcript (ar.to_text_new()) is interpolated into the judge prompt via .format().

This means an agent that embeds a valid JSON verdict matching the output schema in its transcript can cause the judge to output that JSON, passing both json.loads and jsonschema.validate.

Fix

Three changes (+55/-3 lines):

1. _schema_to_tool_info: converts the rubric output_schema into a ToolInfo for structured-output enforcement.
2. _extract_output: extracts the verdict from tool_calls[0].arguments (preferred) with a json.loads fallback for providers that do not support tools.
3. evaluate_rubric: passes tools=[judge_tool], tool_choice='required' to get_llm_completions_async, so the provider enforces the schema natively rather than relying on text parsing.

Verification

• python3 -m py_compile: compiles OK
• black --check --line-length=100: clean
• Regression PoC (json.loads injection on raw text) now fails after the fix
• Backward-compat grep: no external callers of changed functions
• The json.loads fallback in _extract_output preserves compatibility for providers without structured output support