src/explicit_learning/contamination/match.py

"""Decontamination matching: automatic rejection and a clean allowlist.

``docs/02_DATA_PIPELINE.md`` §5.2/§5.3 define two outcomes for every train
candidate, measured against the frozen evaluation index:

- ``REJECT_CONTAMINATION`` — an automatic, non-reviewable exclusion.
- near-duplicate — automatically excluded and written to an immutable rejection
  report.  There is no human override or re-admission path.

The final allowlist is the train base_ids that are neither exact contamination
nor near-duplicates.  This fail-closed rule preserves the zero-per-example-human
contract and cannot be relaxed to meet a quota.

Thresholds are frozen here, not tuned to results (``docs/02`` §5.3).
"""

from __future__ import annotations

from collections.abc import Mapping, Sequence
from dataclasses import dataclass, field
from typing import Any

from ..atomic_io import atomic_write_jsonl
from .fingerprints import estimated_jaccard, phash_hamming
from .index import LoadedFingerprints, load_all

# Frozen thresholds (docs/02 §5.3).
PHASH_HAMMING_MAX = 4
OCR_JACCARD_MIN = 0.85
QUESTION_JACCARD_MIN = 0.90

REJECT = "REJECT_CONTAMINATION"
NEAR_DUP = "NEAR_DUPLICATE"
CLEAN = "CLEAN"


@dataclass(frozen=True)
class Match:
    """One near-duplicate match between a train item and an eval item."""

    eval_base_id: str
    reason: str
    evidence: dict[str, Any] = field(default_factory=dict)


@dataclass
class ItemDecision:
    """The decontamination verdict for one train base_id."""

    base_id: str
    source: str
    outcome: str  # REJECT | NEAR_DUPLICATE | CLEAN
    reject_reason: str | None = None
    matches: list[Match] = field(default_factory=list)


@dataclass
class DecontaminationResult:
    rejected: list[ItemDecision]
    near_duplicates: list[ItemDecision]
    clean: list[ItemDecision]

    @property
    def allowlist_base_ids(self) -> list[str]:
        return sorted(d.base_id for d in self.clean)


def _classify(
    train: LoadedFingerprints,
    eval_fps: Sequence[LoadedFingerprints],
    *,
    eval_image_shas: set[str],
    source_native_to_base: dict[str, str],
    eval_by_base: dict[str, LoadedFingerprints],
    canonical_q_to_bases: Mapping[str, Sequence[str]],
) -> ItemDecision:
    # --- auto-reject (docs/02 §5.2) ---
    shared_images = [sha for sha in train.image_sha256 if sha in eval_image_shas]
    if shared_images:
        return ItemDecision(
            base_id=train.base_id,
            source=train.source,
            outcome=REJECT,
            reject_reason="train image SHA-256 matches an eval image",
        )
    # canonical question SHA-256 AND >=1 image SHA-256 both match the SAME eval item.
    if train.question_canonical_sha256 in canonical_q_to_bases:
        for eval_base in canonical_q_to_bases[train.question_canonical_sha256]:
            eval_fp = eval_by_base[eval_base]
            if train.image_sha256 and set(train.image_sha256) & set(eval_fp.image_sha256):
                return ItemDecision(
                    base_id=train.base_id,
                    source=train.source,
                    outcome=REJECT,
                    reject_reason=(
                        "canonical question SHA-256 and an image SHA-256 both match "
                        "the same eval item"
                    ),
                )
    if f"{train.source}\0{train.native_id}" in source_native_to_base:
        return ItemDecision(
            base_id=train.base_id,
            source=train.source,
            outcome=REJECT,
            reject_reason="source/native_id explicitly matches the evaluation registry",
        )
    if train.derived_from_eval:
        return ItemDecision(
            base_id=train.base_id,
            source=train.source,
            outcome=REJECT,
            reject_reason="provenance states the row derives from an evaluation-only source",
        )

    # --- near-duplicate queue (docs/02 §5.3) ---
    matches: list[Match] = []
    seen_eval: set[str] = set()
    for eval_fp in eval_fps:
        if eval_fp.base_id in seen_eval:
            continue
        match = _near_dup_match(train, eval_fp)
        if match is not None:
            matches.append(match)
            seen_eval.add(eval_fp.base_id)
    if matches:
        matches.sort(key=lambda m: m.eval_base_id)
        return ItemDecision(
            base_id=train.base_id, source=train.source, outcome=NEAR_DUP, matches=matches
        )
    return ItemDecision(base_id=train.base_id, source=train.source, outcome=CLEAN)


def _near_dup_match(train: LoadedFingerprints, eval_fp: LoadedFingerprints) -> Match | None:
    # pHash Hamming distance <= 4 between any train image and any eval image.
    for t_hash in train.image_phash:
        for e_hash in eval_fp.image_phash:
            if phash_hamming(t_hash, e_hash) <= PHASH_HAMMING_MAX:
                return Match(
                    eval_base_id=eval_fp.base_id,
                    reason="phash_hamming_le_4",
                    evidence={"train_phash": t_hash, "eval_phash": e_hash},
                )
    # OCR MinHash estimated Jaccard >= 0.85 (only when both sides have OCR text).
    if (
        train.ocr_minhash
        and eval_fp.ocr_minhash
        and (estimated_jaccard(train.ocr_minhash, eval_fp.ocr_minhash) >= OCR_JACCARD_MIN)
    ):
        return Match(eval_base_id=eval_fp.base_id, reason="ocr_minhash_jaccard_ge_0.85")
    # question MinHash estimated Jaccard >= 0.90.
    if (
        train.question_minhash
        and eval_fp.question_minhash
        and (
            estimated_jaccard(train.question_minhash, eval_fp.question_minhash)
            >= QUESTION_JACCARD_MIN
        )
    ):
        return Match(eval_base_id=eval_fp.base_id, reason="question_minhash_jaccard_ge_0.90")
    # question + choices nearly the same but image crop differs.
    if (
        train.question_canonical_sha256
        and train.question_canonical_sha256 == eval_fp.question_canonical_sha256
        and train.choices_sha256
        and train.choices_sha256 == eval_fp.choices_sha256
        and set(train.image_sha256).isdisjoint(set(eval_fp.image_sha256))
    ):
        return Match(eval_base_id=eval_fp.base_id, reason="same_question_choices_different_image")
    return None


def decontaminate(
    train_fps: Sequence[LoadedFingerprints],
    eval_fps: Sequence[LoadedFingerprints],
) -> DecontaminationResult:
    """Classify every train item against the eval index."""
    eval_by_base = {fp.base_id: fp for fp in eval_fps}
    eval_image_shas: set[str] = set()
    source_native_to_base: dict[str, str] = {}
    canonical_q_to_bases: dict[str, list[str]] = {}
    for fp in eval_fps:
        eval_image_shas.update(fp.image_sha256)
        source_native_to_base[f"{fp.source}\0{fp.native_id}"] = fp.base_id
        if fp.question_canonical_sha256:
            canonical_q_to_bases.setdefault(fp.question_canonical_sha256, []).append(fp.base_id)

    rejected: list[ItemDecision] = []
    near_duplicates: list[ItemDecision] = []
    clean: list[ItemDecision] = []
    for train in sorted(train_fps, key=lambda f: f.base_id):
        decision = _classify(
            train,
            eval_fps,
            eval_image_shas=eval_image_shas,
            source_native_to_base=source_native_to_base,
            eval_by_base=eval_by_base,
            canonical_q_to_bases=canonical_q_to_bases,
        )
        if decision.outcome == REJECT:
            rejected.append(decision)
        elif decision.outcome == NEAR_DUP:
            near_duplicates.append(decision)
        else:
            clean.append(decision)
    return DecontaminationResult(rejected=rejected, near_duplicates=near_duplicates, clean=clean)


def _allowlist_rows(base_ids: Sequence[str], source_of: Mapping[str, str]) -> list[dict[str, Any]]:
    return [
        {"base_id": bid, "source": source_of.get(bid, ""), "policy": "train_candidate"}
        for bid in sorted(base_ids)
    ]


def _near_duplicate_rejection_rows(
    near_duplicates: Sequence[ItemDecision],
) -> list[dict[str, Any]]:
    rows: list[dict[str, Any]] = []
    for decision in sorted(near_duplicates, key=lambda d: d.base_id):
        for match in decision.matches:
            rows.append(
                {
                    "base_id": decision.base_id,
                    "source": decision.source,
                    "eval_base_id": match.eval_base_id,
                    "reason": match.reason,
                    "evidence": match.evidence,
                    "status": "auto_rejected",
                }
            )
    return rows


def write_allowlist_and_review(
    result: DecontaminationResult,
    eval_base_ids: Sequence[str],
    *,
    allowlist_path: str | None = None,
    review_path: str | None = None,
    decisions: Mapping[str, str] | None = None,
) -> dict[str, Any]:
    """Write the allowlist and automatic near-duplicate rejects as JSONL.

    ``decisions`` is retained only to fail loudly for callers using the removed
    human-override API.  No evaluation base_id or near-duplicate may ever enter
    the allowlist.
    """
    if decisions is not None:
        raise ValueError(
            "per-example decontamination decisions are forbidden; "
            "near-duplicates are always auto-rejected"
        )
    train_source = {d.base_id: d.source for d in result.clean + result.near_duplicates}
    allowlist_ids = list(result.allowlist_base_ids)

    eval_id_set = set(eval_base_ids)
    summary: dict[str, Any] = {
        "allowlist_count": len(allowlist_ids),
        "rejected_count": len(result.rejected),
        "near_duplicate_count": len(result.near_duplicates),
        "near_duplicates_auto_rejected": len(result.near_duplicates),
        "undecided_near_duplicates": 0,
        "eval_ids_in_allowlist": sum(1 for bid in allowlist_ids if bid in eval_id_set),
    }

    if allowlist_path is not None:
        atomic_write_jsonl(allowlist_path, _allowlist_rows(allowlist_ids, train_source))
    if review_path is not None:
        atomic_write_jsonl(review_path, _near_duplicate_rejection_rows(result.near_duplicates))
    return summary


def decontaminate_dbs(
    train_db: str,
    eval_db: str,
    *,
    allowlist_path: str | None = None,
    review_path: str | None = None,
    decisions: Mapping[str, str] | None = None,
) -> dict[str, Any]:
    """Load both indexes, decontaminate, and write outputs. Returns the summary."""
    eval_fps = load_all(eval_db)
    result = decontaminate(load_all(train_db), eval_fps)
    summary = write_allowlist_and_review(
        result,
        [fp.base_id for fp in eval_fps],
        allowlist_path=allowlist_path,
        review_path=review_path,
        decisions=decisions,
    )
    summary["rejected"] = [
        {"base_id": d.base_id, "reason": d.reject_reason} for d in result.rejected
    ]
    return summary