fix(eval): block self-judging rows for llm_judge

[ulises-jeremias]

Description

Prevent self-judging in offline evaluation when llm_judge is active.

This change adds a guard that rejects any evaluated row where the trajectory model matches --judge-model (with litellm_proxy/ prefix normalization). This avoids candidate=self-judge comparisons in the default evaluation path.

Type of Change

• Bug fix

Impact on Benchmarking

• No change to baselines: This change enforces evaluator safety checks and does not modify scoring logic itself.

Related Issues

• Fixes: Default agent and judge models are identical — out-of-the-box self-evaluation #336

Verification Steps

1. uvx ruff check src/evaluation/evaluator.py src/evaluation/runner.py src/evaluation/cli.py src/evaluation/tests/test_evaluator.py
2. uv run pytest src/evaluation/tests

Checklist

• I have added tests that prove my fix is effective.
• My code follows the project's Ruff formatting and linting rules.
• I have signed off my commits (DCO).

Notes

• Wired judge_model into Evaluator and evaluation.runner.evaluate(...) so the same guard applies to CLI and programmatic evaluation.
• Added focused evaluator tests for exact match, normalized model-ID match, and non-LLM scorer bypass.
• Updated evaluation docs and instructions to document the self-judging guard behavior.
• Updated evaluation examples so default claude-agent candidate runs are no longer paired with the same judge model in docs.

Additional Quick Notes

• Scope is intentionally narrow and isolated to evaluation-path safety checks; no runtime agent behavior or scoring rubric logic was changed.
• The guard only runs when the active scorer is llm_judge and a judge_model is provided.
• Comparison normalizes litellm_proxy/ to avoid false negatives with equivalent model IDs.