#!/usr/bin/env python3
"""Export exact training-row identities from a foundation feature cache."""

from __future__ import annotations

import argparse
import hashlib
import json
import os

import torch


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--feature_cache", required=True)
    parser.add_argument("--out", required=True)
    args = parser.parse_args()

    payload = torch.load(args.feature_cache, map_location="cpu", weights_only=False, mmap=True)
    rows = payload.get("rows") or []
    keys = [[str(row.get("boss")), int(row.get("fight", -1)), int(row.get("index", -1))] for row in rows]
    if not keys:
        raise ValueError(f"feature cache contains no rows: {args.feature_cache}")
    if len({tuple(key) for key in keys}) != len(keys):
        raise ValueError(f"feature cache contains duplicate row keys: {args.feature_cache}")
    digest = hashlib.sha256(
        json.dumps(keys, ensure_ascii=False, separators=(",", ":")).encode("utf-8")
    ).hexdigest()
    value = {
        "schema_revision": "pact-matched-training-rows-v1",
        "source_feature_cache": os.path.realpath(args.feature_cache),
        "source_request_signature": payload.get("request_signature"),
        "n_rows": len(keys),
        "row_keys_sha256": digest,
        "row_keys": keys,
    }
    os.makedirs(os.path.dirname(args.out) or ".", exist_ok=True)
    tmp = f"{args.out}.tmp.{os.getpid()}"
    with open(tmp, "w", encoding="utf-8") as handle:
        json.dump(value, handle, ensure_ascii=False, indent=2, allow_nan=False)
        handle.flush()
        os.fsync(handle.fileno())
    os.replace(tmp, args.out)
    print(json.dumps({key: value[key] for key in value if key != "row_keys"}, indent=2))
    print(f"wrote {args.out}")


if __name__ == "__main__":
    main()