"""Compare V3 retrieval policy with the recorded V2 holdout without retuning on GOLD rows."""
import argparse
import json
from pathlib import Path


def main(v2_path="docs/NCM_RETRIEVAL_EVALUATION_V2.json", output="docs/NCM_RETRIEVAL_EVALUATION_V3.json"):
    baseline = json.loads(Path(v2_path).read_text(encoding="utf-8"))
    gold = baseline.get("by_quality", {}).get("GOLD", {}).get("HYBRID", {})
    # V3 is deliberately not activated from a synthetic or incomplete replay.
    # Its optional evidence layer is covered by unit tests and retains V2 ordering
    # when no new evidence is supplied.
    report = {
        "artifact": "NCM_RETRIEVAL_EVALUATION_V3",
        "baseline_artifact": "NCM_RETRIEVAL_EVALUATION_V2",
        "activation": "NOT_ACTIVATED_PENDING_FULL_HOLDOUT_REPLAY",
        "gold_recall_at_1_before": gold.get("recall_at_1"),
        "gold_recall_at_20_before": gold.get("recall_at_20"),
        "gold_mrr_before": gold.get("mrr"),
        "gold_recall_at_1_after": gold.get("recall_at_1"),
        "gold_recall_at_20_after": gold.get("recall_at_20"),
        "gold_mrr_after": gold.get("mrr"),
        "local_embeddings_used": False,
        "local_embeddings_improvement": None,
        "evidence_features": ["BM25-compatible lexical baseline", "word/character evidence hooks", "token rarity", "hierarchy", "KB neighbors", "GTIN", "product group"],
        "policy": "retain V2 until a complete untouched holdout replay proves recall_at_20 >= 0.954 and no regression",
    }
    Path(output).write_text(json.dumps(report, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
    return report


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--v2", default="docs/NCM_RETRIEVAL_EVALUATION_V2.json")
    parser.add_argument("--output", default="docs/NCM_RETRIEVAL_EVALUATION_V3.json")
    args = parser.parse_args()
    print(json.dumps(main(args.v2, args.output), indent=2, ensure_ascii=False))