"""Principle-level gold matching (importance layer #153, component 1).

Flags the SPECIFIC principle a chair relied on — not the whole precedent. A chair
cites a precedent for ONE holding; we embed the citation's ``match_context`` (or a
digest's ``headline_holding``) and cosine-match it to the principles extracted from
that precedent. The best match ≥ threshold gets the flag:

  • OUR_CHAIR (דפנה) cited it      → ``gold_chair``   (tier-1, protective in the cull)
  • another committee chair cited it → ``chair_cited`` (tier-2, weight not protection)
  • a digest highlights it          → ``gold_digest``  (tier-1)

Embedding-only — NO LLM, so it is cheap and re-runnable (the periodic refresh job).
Tuned toward recall (HALACHA_GOLD_MATCH_THRESHOLD): missing a chair-relied principle
(buried by the cull) is worse than over-protecting one.
"""
from __future__ import annotations

import logging
from uuid import UUID

from legal_mcp import config
from legal_mcp.services import db, embeddings

logger = logging.getLogger(__name__)


async def _embed_all(texts: list[str]) -> list[list[float]]:
    """Embed in Voyage-sized batches (embed_texts already chunks at 128)."""
    return await embeddings.embed_texts(texts, input_type="query") if texts else []


async def match_all(threshold: float | None = None, apply: bool = False) -> dict:
    """Run gold matching over all chair citations + digests.

    Returns a stats dict; when ``apply`` writes the flags (idempotent — clears
    prior flags first). Dry-run computes everything but writes nothing.
    """
    threshold = config.HALACHA_GOLD_MATCH_THRESHOLD if threshold is None else threshold
    if apply:
        cleared = await db.reset_principle_gold()
        logger.info("principle_gold: cleared %d prior flags before re-match", cleared)

    citations = await db.gold_chair_citations()
    digests = await db.gold_digest_holdings()

    cit_vecs = await _embed_all([c["match_context"] for c in citations])
    dig_vecs = await _embed_all([d["holding_text"] for d in digests])

    stats = {"threshold": threshold, "applied": apply,
             "chair_citations": len(citations), "digests": len(digests),
             "gold_chair": 0, "chair_cited": 0, "gold_digest": 0,
             "chair_no_match": 0, "digest_no_match": 0, "samples": []}

    async def _flag(case_law_id, vec, *, gold_chair=False, gold_digest=False,
                    chair_cited=False, label=""):
        match = await db.nearest_original_principle(case_law_id, list(vec))
        if match is None or match[1] < threshold:
            return None
        hid, sim = match
        if apply:
            await db.set_principle_gold(
                UUID(hid), gold_chair=gold_chair, gold_digest=gold_digest,
                chair_cited=chair_cited, score=sim)
        if len(stats["samples"]) < 12:
            stats["samples"].append({"label": label, "halacha_id": hid, "sim": round(sim, 3)})
        return hid

    for c, vec in zip(citations, cit_vecs):
        cid = c["cited_case_law_id"]
        if c["is_our_chair"]:
            hid = await _flag(cid, vec, gold_chair=True, label="chair:דפנה")
            stats["gold_chair" if hid else "chair_no_match"] += 1
        else:
            hid = await _flag(cid, vec, chair_cited=True, label=f"chair:{c['citing_chair']}")
            stats["chair_cited" if hid else "chair_no_match"] += 1

    for d, vec in zip(digests, dig_vecs):
        hid = await _flag(d["linked_case_law_id"], vec, gold_digest=True, label="digest")
        stats["gold_digest" if hid else "digest_no_match"] += 1

    if apply:
        stats["coverage"] = await db.gold_coverage_stats()
    return stats