fix: prevent combine_scores from re-consuming aggregation keys on repeated calls

[Copilot]

combine_scores iterated over all top-level keys in scores without guarding against known aggregation keys (label_scores, overall_*, total_evals, etc.). Passing an already-combined dict back in would re-consume those entries as crop data, corrupting TP/FP/FN accumulators and skewing all PQ outputs.

Changes

• aggregation.py – Defines _AGGREGATION_KEYS frozenset of all known non-crop top-level keys; adds if crop_name in _AGGREGATION_KEYS: continue guard at the top of the crop-iteration loop in combine_scores
• tests/test_evaluate_metrics.py – Adds test_combine_scores_idempotent_on_combined_dict regression test verifying scores are stable across two sequential calls

first  = combine_scores(scores, ...)
second = combine_scores(first, ...)   # previously corrupted results; now identical
assert np.isclose(first["overall_score"], second["overall_score"])  # passes
assert first["label_scores"]["instance"]["tp"] == second["label_scores"]["instance"]["tp"]  # passes

---

⌨️ Start Copilot coding agent tasks without leaving your editor — available in VS Code, Visual Studio, JetBrains IDEs and Eclipse.