"""
convergence.py — Turn an infer.Pipeline result into a single ForagerResult.

Ported from forager_ml with the Hailo coupling removed: instead of a
RawPrediction it takes the plain dict returned by infer.Pipeline.identify().
The safety-first philosophy is unchanged — abstain by default, flag DEADLY
prominently, never present a below-threshold guess as an identification.
"""

from dataclasses import dataclass

from .metadata import SPECIES_METADATA, UNKNOWN_META

LOW_CONFIDENCE_THRESHOLD    = 0.50  # below this -> flagged low_confidence
EXPERT_CONFIDENCE_THRESHOLD = 0.60  # a committed (non-deadly) ID must clear this.
                                    # These experts are accurate but underconfident
                                    # on SAFE classes (avg ~0.5–0.6); 0.60 balances
                                    # decisiveness (75% safe-correct) against residual
                                    # deadly-as-safe. The UX never treats SAFE as a
                                    # green light — see app.py — so the gate is a
                                    # usability dial, not the safety mechanism.
DEADLY_VETO_FLOOR           = 0.40  # a DEADLY call at/above this overrides any
                                    # higher-confidence SAFE/CAUTION call from
                                    # another expert (safety-biased arbitration)


@dataclass
class ForagerResult:
    domain:          str
    species:         str    # class key, or "unknown"
    scientific_name: str
    confidence:      float
    safety:          str    # SAFE | CAUTION | DEADLY | UNKNOWN
    lookalike:       str
    key_diff:        str
    low_confidence:  bool
    expert_model:    str
    abstained:       bool
    reason:          str     # why we abstained (or "" when committed)

    @property
    def is_deadly(self) -> bool:
        return self.safety == "DEADLY" and not self.low_confidence

    @property
    def is_unknown(self) -> bool:
        return self.species == "unknown"


_ABSTAIN_REASON = {
    "uncertain_domain": "Couldn't confidently place this in a known domain.",
    "off_domain":       "This doesn't look like anything in the trained domains.",
    "low_confidence":   "Leaning toward an answer, but not confident enough to commit.",
}


def _abstain(domain: str, reason: str, conf: float = 0.0, expert: str = "none") -> ForagerResult:
    return ForagerResult(
        domain=domain, species="unknown", scientific_name=UNKNOWN_META["scientific"],
        confidence=conf, safety="UNKNOWN", lookalike=UNKNOWN_META["lookalike"],
        key_diff=_ABSTAIN_REASON.get(reason, UNKNOWN_META["key_diff"]),
        low_confidence=True, expert_model=expert, abstained=True, reason=reason,
    )


def _commit(domain: str, c: dict) -> ForagerResult:
    species = c["species"]
    meta = SPECIES_METADATA.get(species, UNKNOWN_META)
    conf = float(c["confidence"])
    return ForagerResult(
        domain=domain, species=species, scientific_name=meta["scientific"], confidence=conf,
        safety=meta["safety"], lookalike=meta["lookalike"], key_diff=meta["key_diff"],
        low_confidence=conf < LOW_CONFIDENCE_THRESHOLD, expert_model=c["expert"],
        abstained=False, reason="",
    )


def build_result(call: dict) -> ForagerResult:
    """
    `call` is the dict from infer.Pipeline.identify().

    Safety-biased arbitration across the domain's experts:
      1. If ANY expert flags a DEADLY species at >= DEADLY_VETO_FLOOR, surface
         that — a deadly vote vetoes a more-confident SAFE/CAUTION call from an
         off-domain expert (prevents e.g. hemlock -> "ramps" because highvalue
         is more confident than medicinals).
      2. Otherwise take the highest-confidence call, abstaining if it can't
         clear EXPERT_CONFIDENCE_THRESHOLD.
    """
    domain = call.get("domain", "unknown")

    if call.get("abstain") and "calls" not in call:
        return _abstain(domain, call.get("reason", "low_confidence"),
                        float(call.get("confidence", 0.0)), call.get("expert", "none"))

    calls = call.get("calls", [])
    if not calls:
        return _abstain(domain, "low_confidence")

    for c in calls:
        c["safety"] = SPECIES_METADATA.get(c["species"], UNKNOWN_META)["safety"]

    deadly = [c for c in calls if c["safety"] == "DEADLY" and c["confidence"] >= DEADLY_VETO_FLOOR]
    if deadly:
        return _commit(domain, max(deadly, key=lambda c: c["confidence"]))

    best = max(calls, key=lambda c: c["confidence"])
    if best["confidence"] < EXPERT_CONFIDENCE_THRESHOLD:
        return _abstain(domain, "low_confidence", best["confidence"], best["expert"])
    return _commit(domain, best)