feat: reward ensembles and inter-rater reliability metrics

[RUFFY-369]

PR Type

• RL Environment PR - Complete Environment Snapshot & Zero-Training sections
• Non-Environment PR - Complete Description, Related Issues & Type of Change sections

---

📝 General Information

Description

I’ve added a new EnsembleReward class to atroposlib/envs/reward_fns/ to mitigate reward hacking and high variance in RL training. Instead of relying on a single scoring model, this allows for aggregating multiple scorers using mean, median, min, or majority_vote.

I also integrated Krippendorff's alpha to track inter-rater reliability (IRR) across the ensemble. This is a critical observability tool to catch when scorers are fundamentally disagreeing—a common signal that the agent is finding an unaligned edge case.

Related Issues

• Part of [Enhancement] RL Training Infrastructure Stabilization & Observability #431 (RL Infrastructure Enhancements)

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update
• Code refactor (no functional changes)
• Build/CI/CD related changes
• Other (please describe):

---

✅ Developer & Reviewer Checklist

• Code follows project style (black, isort, flake8 pass with pre-commit)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• New and existing unit tests pass locally with my changes (17/17 verified)
• Docstrings added for all new public classes / functions
• If .env vars required, did you add it to the .env.example in repo root? (N/A)