chore: improve reward eval harness auditing

[jbarnes850]

Summary

• add configs/eval/reward_system.yaml so judge presets/combos can be edited without touching code, falling back to baked-in defaults if missing
• extend scripts/eval_reward_models.py with a --collect-audit flag, safe audit payload serialization, and automatic Markdown report emission alongside JSON output
• document the new flags and config-driven workflow in docs/reward_eval.md

Testing

• python -m scripts.eval_reward_models --dataset atlas/data/reward_eval_trajectories.jsonl --judge-combos gemini_pair --baseline gemini_pair --repeats 1 --concurrency 1 --collect-audit --output results/reward/latest_gemini.json

[Copilot]

Pull Request Overview

This PR introduces configuration-driven judge presets and audit logging for the reward evaluation harness. It moves judge configuration from hardcoded constants into an external YAML file, adds a --collect-audit flag to capture LLM interactions for debugging, and automatically generates Markdown summary reports alongside JSON output.

Key Changes:

• Externalized judge presets and combos to configs/eval/reward_system.yaml with fallback to defaults
• Added audit collection capability to track LLM prompts/responses during evaluation
• Implemented automatic Markdown report generation with timestamp-based naming

Reviewed Changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated 3 comments.

File	Description
scripts/eval_reward_models.py	Added YAML config loading, audit collection in evaluator, JSON sanitization helper, and Markdown report generation
configs/eval/reward_system.yaml	New configuration file defining judge presets and combos for reward evaluation
docs/reward_eval.md	Updated documentation to reference config-first workflow and new CLI flags

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}