meida_backend/tests/unit/test_report_validation.py

"""生成后校验纯逻辑：贬损词表（含变形）、一致性三条规则、绝对化用语、降级路径。"""

from __future__ import annotations

from typing import Any

from packages.report.copy import CopySection, ReportCopy, render_fallback_copy
from packages.report.validation import (
    check_absolute_claims,
    check_conclusion_consistency,
    check_pejorative,
    validate_copy,
)


def _copy(summary: str = "", *paragraphs: str) -> ReportCopy:
    return ReportCopy(
        summary=summary,
        sections=[CopySection(title="正文", paragraphs=list(paragraphs))],
    )


def _dim(
    key: str,
    value: str = "",
    *,
    status: str = "ok",
    rule_hits: tuple[str, ...] = (),
    evidence: tuple[str, ...] = (),
) -> dict[str, Any]:
    return {
        "key": key,
        "status": status,
        "value": value,
        "confidence": 0.82,
        "evidence": list(evidence),
        "rule_hits": list(rule_hits),
    }


# ---------------------------------------------------------------------------
# 贬损词表
# ---------------------------------------------------------------------------


def test_pejorative_hits_variants_without_stemming() -> None:
    # 「有点胖」「微胖」都命中「胖」——子串匹配即变形处理，不做词干
    assert check_pejorative("整体轮廓有点胖，建议宽松款") == ["胖"]
    assert check_pejorative("微胖体型适合直筒剪裁") == ["胖"]


def test_pejorative_returns_every_matched_term_in_table_order() -> None:
    assert check_pejorative("配色显老又土气，显得臃肿") == ["臃肿", "土气", "显老"]


def test_pejorative_clean_text_passes() -> None:
    assert check_pejorative("肩腰臀比例均衡，收腰设计能强化这一优势。") == []


def test_pejorative_terms_count_stays_within_initial_band() -> None:
    """初版词表约定 20~40 词：太少护不住红线，太多误杀（正式版归 T-P0-07）。"""
    from packages.report.validation import PEJORATIVE_TERMS

    assert 20 <= len(PEJORATIVE_TERMS) <= 40


# ---------------------------------------------------------------------------
# 一致性规则①：体型/脸型/季型枚举词
# ---------------------------------------------------------------------------


def test_enum_word_mismatching_diagnosis_is_flagged() -> None:
    dimensions = [_dim("body_shape", "X型")]
    copy = _copy("你的整体印象", "你的体型属于 A 型，肩线舒展。")

    issues = check_conclusion_consistency(copy, dimensions)

    assert any("体型" in issue and "A型" in issue for issue in issues)


def test_enum_word_matching_diagnosis_or_its_alias_passes() -> None:
    dimensions = [_dim("body_shape", "X型")]
    same_value = _copy("你的整体印象", "你的体型结论为 X 型，比例均衡。")
    alias = _copy("你的整体印象", "接近沙漏型的轮廓，腰线明显。")

    assert check_conclusion_consistency(same_value, dimensions) == []
    assert check_conclusion_consistency(alias, dimensions) == []


def test_enum_word_for_skipped_dimension_is_treated_as_fabrication() -> None:
    dimensions = [_dim("face_shape", status="skipped")]
    copy = _copy("你的整体印象", "你的圆脸轮廓柔和，适合纵向线条。")

    issues = check_conclusion_consistency(copy, dimensions)

    assert any("脸型" in issue for issue in issues)


def test_season_enum_mismatch_is_flagged() -> None:
    dimensions = [_dim("color_season", "浅暖柔和型")]
    copy = _copy("色彩结论", "你属于典型的夏季型，适合冷调色彩。")

    issues = check_conclusion_consistency(copy, dimensions)

    assert any("季型" in issue for issue in issues)


# ---------------------------------------------------------------------------
# 一致性规则②：色值必须在色板集合内
# ---------------------------------------------------------------------------


def test_hex_color_outside_palette_is_flagged() -> None:
    dimensions = [_dim("color_palette", "#A52A2A", evidence=("#F5F5DC",))]
    inside = _copy("色彩方案", "主色可用 #A52A2A，辅色 #f5f5dc 提亮。")
    outside = _copy("色彩方案", "主色建议 #AABBCC。")

    assert check_conclusion_consistency(inside, dimensions) == []
    issues = check_conclusion_consistency(outside, dimensions)
    assert any("#AABBCC" in issue for issue in issues)


# ---------------------------------------------------------------------------
# 一致性规则③：建议句式必须有规则命中支撑
# ---------------------------------------------------------------------------


def test_advice_with_no_rule_hits_at_all_is_flagged() -> None:
    dimensions = [_dim("body_shape", "X型", rule_hits=())]
    copy = _copy("总体", "建议优先选择收腰版型，突出腰部线条。")

    issues = check_conclusion_consistency(copy, dimensions)

    assert any("没有任何命中" in issue for issue in issues)


def test_advice_on_dimension_without_support_is_flagged() -> None:
    dimensions = [
        _dim("body_shape", "X型", rule_hits=("R-BODY-0007",)),
        _dim("color_season", "浅暖柔和型", rule_hits=()),
    ]
    copy = _copy("总体", "关于季型，建议主色选择高饱和冷调。")

    issues = check_conclusion_consistency(copy, dimensions)

    assert any("季型" in issue and "没有规则命中支撑" in issue for issue in issues)


def test_advice_supported_by_rule_hits_passes() -> None:
    dimensions = [_dim("body_shape", "X型", rule_hits=("R-BODY-0007",))]
    rule_hits = [{"rule_key": "R-BODY-0007", "dimension": "body_shape", "suggestion": "收腰"}]
    copy = _copy("总体", "关于体型，建议优先选择收腰版型，突出腰部线条。")

    assert check_conclusion_consistency(copy, dimensions, rule_hits=rule_hits) == []


# ---------------------------------------------------------------------------
# 绝对化用语
# ---------------------------------------------------------------------------


def test_absolute_claims_are_flagged() -> None:
    assert "最合适" in check_absolute_claims("这款收腰连衣裙最合适不过了")
    assert "第一" in check_absolute_claims("你是本周第一位、也是第一优选用户")
    assert "百分百" in check_absolute_claims("百分百提升气质")


def test_absolute_claims_tolerate_neutral_words_with_the_same_head() -> None:
    # 「最」单字不拦：最近/最终是中性词，拦了会把正常文案打进重试
    assert check_absolute_claims("根据最近一次分析，你的体型结论保持稳定。") == []


# ---------------------------------------------------------------------------
# 降级路径的纯函数级验证：脏文案违规 → 模板文案零违规
# ---------------------------------------------------------------------------


def test_validate_copy_aggregates_all_three_check_categories() -> None:
    dimensions = [_dim("body_shape", "X型")]
    dirty = _copy("小结", "整体偏微胖，收腰款最合适。", "你的体型属于 A 型。")

    violations = validate_copy(dirty, dimensions)

    assert any(v.startswith("贬损词：") for v in violations)
    assert any(v.startswith("结论不一致：") for v in violations)
    assert any(v.startswith("绝对化用语：") for v in violations)


def test_degraded_fallback_is_clean_when_llm_copy_stays_dirty() -> None:
    """重试后仍违规 → 降级模板：最终文案零违规（T-P3-11 黄金断言的纯函数版）。"""
    dimensions = [
        _dim(
            "body_shape", "X型", evidence=("shoulder_hip_ratio:1.02",), rule_hits=("R-BODY-0007",)
        ),
        _dim("face_shape", "鹅蛋", evidence=("face_ratio:1.35",)),
    ]
    rule_hits = [
        {
            "rule_key": "R-BODY-0007",
            "dimension": "body_shape",
            "suggestion": "优先选择收腰版型突出腰线",
            "explanation": "X 型体型肩腰臀比例均衡，收腰设计能强化这一优势",
        }
    ]
    dirty = _copy("小结", "整体偏微胖，收腰款最合适。")
    assert validate_copy(dirty, dimensions, rule_hits=rule_hits) != []

    fallback = render_fallback_copy(dimensions, rule_hits)

    assert validate_copy(fallback, dimensions, rule_hits=rule_hits) == []
    assert check_pejorative(fallback.full_text()) == []