"""The experience seam (ExpeL-style): store, structural retrieval, fold (§3 Step 1).

Prior verdicts reach the hypothesis prompt ONLY through this seam: bundle seeding →
in-memory store → structural retrieval → fold-before-generation. Ranking is
structural, never textual — surface text must not contribute to similarity. The
rationale is the carrier of the learning signal: the fold is what lets an expert's
realization-rate correction reach the next hypothesis.
"""

from __future__ import annotations

import hashlib
import json
from dataclasses import dataclass

from pathlib import Path

from .ir import SavingsProposal, load_validator_input
from .okf import navigate_bundle

_VERDICT_TYPE = "verdict"
_DEFAULT_SEED_DECISION = "approved"
# §3 Step 1: frozen similarity weights and magnitude bucket edges.
_JACCARD_WEIGHT = 0.60
_MEASURE_TYPE_WEIGHT = 0.25
_MAGNITUDE_WEIGHT = 0.15
_MAGNITUDE_BUCKET_EDGES = (1e5, 5e5, 1e6)


@dataclass(frozen=True)
class CandidateFeatures:
    """The structural features retrieval ranks over (§4.2 ``proposal_features``)."""

    affected_codes: frozenset[str]
    measure_type: str
    claimed_saving_nok: float

    @classmethod
    def from_proposal(cls, proposal: SavingsProposal) -> CandidateFeatures:
        # The IR projection carries no separate measure-type field; its ``measure``
        # string is the candidate's measure type at this level (§3 Step 1: the
        # query key is read from the IR projection, before any proposal exists).
        return cls(
            affected_codes=frozenset(item.code for item in proposal.affected_items),
            measure_type=proposal.measure,
            claimed_saving_nok=proposal.claimed_saving_nok,
        )


@dataclass(frozen=True)
class VerdictRecord:
    """One store entry: id (the learning-loop key), decision, rationale, features."""

    verdict_id: str
    decision: str
    rationale: str
    features: CandidateFeatures


def mint_verdict_id(features: CandidateFeatures) -> str:
    """First 16 hex chars of SHA-256 over the canonical feature JSON (§4.2).

    Raw JSON number formatting participates in the hash (30000 vs 30000.0 differ),
    which is why a LOADED verdict's id is kept verbatim — never re-minted.
    """
    canonical = json.dumps(
        {
            "affected_codes": sorted(features.affected_codes),
            "claimed_saving_nok": features.claimed_saving_nok,
            "measure_type": features.measure_type,
        },
        sort_keys=True,
        separators=(",", ":"),
    )
    return hashlib.sha256(canonical.encode("utf-8")).hexdigest()[:16]


def _magnitude_bucket(claimed_saving_nok: float) -> int:
    # Buckets [0, 1e5), [1e5, 5e5), [5e5, 1e6), [1e6, ∞) over the claimed saving.
    return sum(1 for edge in _MAGNITUDE_BUCKET_EDGES if claimed_saving_nok >= edge)


def similarity(a: CandidateFeatures, b: CandidateFeatures) -> float:
    """Structural similarity — surface text never contributes (§3 Step 1)."""
    if not a.affected_codes and not b.affected_codes:
        jaccard = 1.0
    else:
        union = a.affected_codes | b.affected_codes
        jaccard = len(a.affected_codes & b.affected_codes) / len(union)
    return (
        _JACCARD_WEIGHT * jaccard
        + _MEASURE_TYPE_WEIGHT * (a.measure_type == b.measure_type)
        + _MAGNITUDE_WEIGHT
        * (_magnitude_bucket(a.claimed_saving_nok) == _magnitude_bucket(b.claimed_saving_nok))
    )


class VerdictStore:
    """In-memory verdict store — FIRST-write-wins per id (idempotent merges, §4.2)."""

    def __init__(self) -> None:
        self._records: dict[str, VerdictRecord] = {}

    def __len__(self) -> int:
        return len(self._records)

    def add(self, record: VerdictRecord) -> None:
        self._records.setdefault(record.verdict_id, record)

    def retrieve(self, features: CandidateFeatures, k: int) -> list[VerdictRecord]:
        """Top-k by structural similarity, ties broken by verdict id ascending."""
        if k <= 0:
            raise ValueError(f"retrieval k must be positive, got {k}")
        ranked = sorted(
            self._records.values(),
            key=lambda record: (-similarity(record.features, features), record.verdict_id),
        )
        return ranked[:k]


def seed_store_from_bundle(store: VerdictStore, bundle_dir: Path) -> int:
    """Seed the store from the bundle's navigable ``type: verdict`` files (§3 Step 1).

    Entries are keyed on the bundle's candidate features, read from the IR
    projection (fail-fast, required input). The rationale is built from the
    ``description`` frontmatter plus, when present, the structured learning fields.
    A ``verdict_id`` in the frontmatter (promoted files, §6) is read VERBATIM
    (§4.2) — re-minting from the bundle features would collide distinct
    promoted candidates into one first-write-wins store slot. Returns the
    number of verdict files seeded.
    """
    features = CandidateFeatures.from_proposal(load_validator_input(bundle_dir))
    seeded = 0
    for concept in navigate_bundle(bundle_dir):
        if concept.type != _VERDICT_TYPE:
            continue
        rationale = concept.frontmatter.get("description", "")
        realization_rate = concept.frontmatter.get("realization_rate")
        expected_actual = concept.frontmatter.get("expected_actual_saving_nok")
        if realization_rate is not None and expected_actual is not None:
            learning = (
                f"[realiseringsgrad={realization_rate}; forventet_faktisk_NOK={expected_actual}]"
            )
            rationale = f"{rationale} {learning}".strip()
        store.add(
            VerdictRecord(
                verdict_id=concept.frontmatter.get("verdict_id") or mint_verdict_id(features),
                decision=concept.frontmatter.get("decision", _DEFAULT_SEED_DECISION),
                rationale=rationale,
                features=features,
            )
        )
        seeded += 1
    return seeded


def fold_experience(
    store: VerdictStore, features: CandidateFeatures, base_context: str, k: int
) -> str:
    """Prepend the retrieved prior verdicts to the generation context (§3 Step 1).

    One line per verdict — id, decision, rationale. An empty retrieval returns the
    base context unchanged (the empty-store control, §11).
    """
    retrieved = store.retrieve(features, k)
    if not retrieved:
        return base_context
    lines = "\n".join(
        f"- {record.verdict_id} [{record.decision}]: {record.rationale}" for record in retrieved
    )
    return f"Prior expert verdicts (most similar first):\n{lines}\n\n{base_context}"