"""Evaluation harness for FixSense (Phases P1-P2).

- ``splits``        : deterministic by-clip train/test partition (no step leakage).
- ``retrieval``      : Recall@k / MRR / nDCG@k over the AutoRepair-VideoQA benchmark (P1).
- ``faithfulness``   : LLM-as-judge cause correctness + step grounding + citation
  validity of the full diagnosis system's output (P2).
"""