backend/tests/test_classify_accuracy.py

"""Accuracy audit for taxonomy routing + classify gates (score fixtures)."""

from __future__ import annotations

from pathlib import Path

import pytest
from PIL import Image

from app.api import _classify_from_scores
from app.taxonomy import choose_best_destination, reload_taxonomy


SELECTED_ALL = {
    "fertilization",
    "NTR",
    "incest",
    "nakadashi",
    "fellatio",
    "paizuri",
    "footjob",
    "tentacles",
    "sex",
    "loli",
    "shota",
    "monster_girl",
    "android",
    "furry",
    "Pokemon",
    "bestiality",
    "Voyeur",
}

# Fallback homes are opt-in; they yield to every folder in SELECTED_ALL.
SELECTED_FALLBACKS = {"SFW", "comic", "scenery"}


@pytest.fixture(autouse=True)
def _reload_default_taxonomy() -> None:
    reload_taxonomy()


@pytest.mark.parametrize(
    "scores,expected",
    [
        ({"loli": 0.92, "flat_chest": 0.99, "lolita_fashion": 0.95}, "loli"),
        ({"shota": 0.88, "1boy": 0.99, "child": 0.9}, "shota"),
        ({"netorare": 0.81, "caught": 0.99}, "NTR"),
        ({"incest": 0.77, "siblings": 0.99}, "incest"),
        ({"internal_cumshot": 0.9, "cum_in_pussy": 0.95}, "nakadashi"),
        ({"fertilization": 0.8, "cum_in_pussy": 0.99}, "fertilization"),
        ({"impregnation": 0.85, "pregnant": 0.99}, "fertilization"),
        ({"irrumatio": 0.9}, "fellatio"),
        ({"monster_girl": 0.9, "horns": 0.99, "wings": 0.98}, "monster_girl"),
        ({"slime_girl": 0.86}, "monster_girl"),
        ({"lamia": 0.9}, "monster_girl"),
        ({"scylla": 0.88}, "monster_girl"),
        ({"shark_girl": 0.85}, "monster_girl"),
        ({"moth_girl": 0.84}, "monster_girl"),
        ({"traditional_youkai": 0.9}, "monster_girl"),
        ({"monsterification": 0.8}, "monster_girl"),
        ({"cow_girl": 0.99}, None),
        ({"extra_eyes": 0.95}, None),
        ({"android": 0.9}, "android"),
        ({"robot_girl": 0.88}, "android"),
        ({"cyborg": 0.85}, "android"),
        ({"mecha_musume": 0.84}, "android"),
        ({"mecha": 0.99}, None),
        ({"mechanical_arms": 0.95}, None),
        ({"furry": 0.9, "animal_ears": 0.99}, "furry"),
        ({"pokemon_(creature)": 0.91, "pokemon_ears": 0.99}, "Pokemon"),
        # False-positive clusters must not route
        ({"lolita_fashion": 0.99, "gothic_lolita": 0.98, "flat_chest": 0.97}, None),
        ({"1boy": 0.99, "male_focus": 0.98, "otoko_no_ko": 0.9}, None),
        ({"animal_ears": 0.99, "fake_animal_ears": 0.9}, None),
        ({"siblings": 0.99}, None),
        ({"pregnant": 0.99}, None),
        # voyeurism is Voyeur/caught evidence (not NTR); caught/watching alone must not route
        ({"voyeurism": 0.99, "caught": 0.9}, "Voyeur/caught"),
        ({"watching": 0.99, "caught": 0.9}, None),
        ({"horns": 0.99, "wings": 0.98, "tail": 0.97}, None),
        ({"oral": 0.99}, None),
        ({"pantyshot": 0.9}, "Voyeur/panties"),
        ({"spread_pussy": 0.9}, "Voyeur/pussy"),
        # Surface cum is tease, not a hard act, so the pussy tease folder keeps it.
        ({"pussy": 0.88, "cum": 0.61}, "Voyeur/pussy"),
        ({"soles": 0.9, "feet": 0.85}, "Voyeur/feet"),
        ({"upskirt": 0.9, "loli": 0.6}, "loli"),
        # Voyeur sub-folder funnel
        ({"highleg_leotard": 0.9, "covered_nipples": 0.7}, "Voyeur/leotard"),
        ({"bikini": 0.95}, "Voyeur/swimsuit"),
        ({"nude": 0.9, "nipples": 0.88}, "Voyeur/nude"),
        ({"lingerie": 0.9}, "Voyeur/lingerie"),
        ({"pantyhose": 0.95, "zettai_ryouiki": 0.8}, "Voyeur/legwear"),
        ({"undressing": 0.9, "open_clothes": 0.7}, "Voyeur/undressing"),
        ({"see-through": 0.9, "nipples": 0.6}, "Voyeur/see_through"),
        ({"exhibitionism": 0.9}, "Voyeur/public"),
        ({"cameltoe": 0.9}, "Voyeur"),
        # see-through needs body exposure; crystalline characters must not match
        ({"see-through": 0.85, "androgynous": 0.9, "other_focus": 0.93}, None),
        # New act folders
        ({"sex": 0.95, "vaginal": 0.9, "hetero": 0.99}, "sex"),
        ({"paizuri": 0.9, "breasts": 0.99}, "paizuri"),
        ({"footjob": 0.9, "feet": 0.99}, "footjob"),
        ({"tentacle_sex": 0.9, "tentacles": 0.85}, "tentacles"),
        ({"consensual_tentacles": 0.88}, "tentacles"),
        ({"tentacles": 0.95}, None),
        ({"pussy": 0.9, "fingering": 0.9}, "sex"),
        ({"sex": 0.95, "fellatio": 0.7}, "fellatio"),
        ({"sex": 0.95, "cum_in_pussy": 0.6}, "nakadashi"),
        ({"sex": 0.95, "tentacle_sex": 0.6}, "tentacles"),
    ],
)
def test_taxonomy_routing_matrix(scores: dict[str, float], expected: str | None) -> None:
    folder, _score, _secondary = choose_best_destination(scores, SELECTED_ALL)
    assert folder == expected


@pytest.mark.parametrize(
    "scores,expected",
    [
        ({"1girl": 0.99, "solo": 0.98, "smile": 0.9}, "SFW"),
        ({"comic": 0.9, "speech_bubble": 0.8, "1girl": 0.85}, "comic"),
        ({"scenery": 0.8, "no_humans": 0.95}, "scenery"),
        # Fallbacks yield to any real destination.
        ({"1girl": 0.99, "pantyshot": 0.4}, "Voyeur/panties"),
        ({"1girl": 0.99, "loli": 0.3}, "loli"),
        ({"1girl": 0.99, "nipples": 0.9}, "Voyeur/nude"),
        # Explicit content with no destination goes to review, never to SFW.
        ({"1girl": 0.99, "censored": 0.6, "penis": 0.2}, None),
    ],
)
def test_fallback_routing_matrix(scores: dict[str, float], expected: str | None) -> None:
    folder, _score, _secondary = choose_best_destination(
        scores, SELECTED_ALL | SELECTED_FALLBACKS
    )
    assert folder == expected


def test_classify_gate_clears_weak_taxonomy_winner(tmp_path: Path) -> None:
    result = _classify_from_scores(
        tmp_path / "x.jpg",
        {"loli": 0.2, "1girl": 0.99},
        {"loli", "shota"},
        confidence_threshold=0.6,
    )
    assert result.primary_tag is None
    assert result.needs_review is True
    assert result.secondary[0]["tag"] == "loli"


def test_classify_keeps_mid_character_primary_for_review(tmp_path: Path) -> None:
    """Character hits below confidence stay as primary (needs_review), not cleared."""
    result = _classify_from_scores(
        tmp_path / "x.jpg",
        {"loli": 0.42, "1girl": 0.99},
        {"loli", "shota", "NTR", "fellatio"},
        confidence_threshold=0.6,
    )
    assert result.primary_tag == "loli"
    assert result.primary_score == 0.42
    assert result.needs_review is True
    assert "threshold" in (result.reason or "").lower()


def test_classify_still_clears_mid_act_primary(tmp_path: Path) -> None:
    result = _classify_from_scores(
        tmp_path / "x.jpg",
        {"fellatio": 0.42, "1girl": 0.99},
        {"loli", "fellatio"},
        confidence_threshold=0.6,
    )
    assert result.primary_tag is None
    assert result.needs_review is True
    assert result.secondary[0]["tag"] == "fellatio"


def test_classify_assigns_strong_taxonomy_winner(tmp_path: Path) -> None:
    result = _classify_from_scores(
        tmp_path / "x.jpg",
        {"loli": 0.91, "flat_chest": 0.95, "shota": 0.4},
        {"loli", "shota", "furry"},
        confidence_threshold=0.6,
    )
    assert result.primary_tag == "loli"
    assert result.primary_score == 0.91
    assert result.needs_review is False


def test_priority_fertilization_over_nakadashi_on_tie(tmp_path: Path) -> None:
    result = _classify_from_scores(
        tmp_path / "x.jpg",
        {"fertilization": 0.8, "internal_cumshot": 0.8},
        {"fertilization", "nakadashi"},
        confidence_threshold=0.6,
    )
    assert result.primary_tag == "fertilization"


def test_model_smoke_nonempty_scores_all_taggers(tmp_path: Path) -> None:
    """Smoke: each tagger returns a non-empty score dict on a tiny RGB image.

    This is not semantic accuracy (no labels); it catches broken sessions /
    empty-output regressions across models.
    """
    pytest.importorskip("imgutils")
    from app.services import extract_scores

    img = tmp_path / "probe.png"
    Image.new("RGB", (448, 448), color=(180, 120, 160)).save(img)

    models = ["ml_danbooru", "wd_swinv2_v3", "wd_eva02_large"]
    for model in models:
        try:
            scores = extract_scores(img, tagger_model=model, wd_general_threshold=0.35)
        except Exception as err:  # pragma: no cover - environment-specific
            pytest.skip(f"tagger {model} unavailable: {err}")
        assert isinstance(scores, dict)
        assert len(scores) > 0, f"{model} returned empty scores"
        assert all(0.0 <= float(v) <= 1.0 for v in scores.values())