feat(report): LLM copy generation with mandatory post-validation and fallback

report_copy/v1 prompt（系统段零变量）+ 结构化 ReportCopy + 素材组装
（skipped 不进 prompt/超长裁剪）+ 贬损词表/一致性/绝对化三重校验 +
重试一次后降级规则直出模板（T-P3-10/11）。

- copy.py：ReportCopy(schema 宽进严出取舍见注释)/compose_prompt_material
  （ok 过滤+6000 字符裁剪，低优先级先丢、硬结论不丢，记录 dropped_keys）/
  render_fallback_copy 零 LLM 降级模板
- validation.py：PEJORATIVE_TERMS 35 词（源自合规 §3，初版内嵌注释）+
  一致性三规则（枚举同款含别名/色值在板内/建议句需规则支撑·宽松口径）+
  绝对化用语兜底
- copy_stage.py / validation_stage.py：任何出口 passed=true 或抛错，
  语义层重试一次（与网关解析层重试独立），仍违规降级模板并 fail-close 自证
- 单测 25 例（词表含变形/三规则正反/降级零违规）+ 集成 4 例
  （FakeProvider 全栈：干净通过/脏一次修正通过/持续脏降级且最终文案
  零贬损词——黄金断言）；全量 529 单测 + 49 集成全绿

偏差说明：
1. prompts/report_copy/v1.txt 非新建——LLM 批次已放置初版（user 段变量
   为 {payload}），按本任务契约改写为 {dimensions_json} 并补内容红线
   （该文件任务文档明确归 T17，无与他人冲突）。
2. check_conclusion_consistency 按契约签名 (copy, dimensions) 之外增加
   可选 rule_hits 关键字参数：规则③的支撑判定需兼容命中内嵌在维度/
   顶层列表两种 T16 契约形态，缺省时不追责（宽松口径）。
3. 绝对化用语未按单字「最」拦截，取营销语境复合词（最好/最合适/第一/
   百分百等 17 词）：单字会误杀「最近/最终」等中性词，红线范围不变。
4. 集成文件未挂 MEIDA_RUN_INTEGRATION skipif：用例为 FakeProvider 内存
   装配、不依赖真实 PG/Redis，按任务要求仍放 tests/integration/ 并带
   integration marker。

packages/report/copy.py

@@ -0,0 +1,250 @@
+"""报告文案的纯逻辑：LLM 素材组装与降级模板渲染（无 IO、无 LLM 调用）。
+
+对应架构 §7.1 步骤 17/18 中可独立单测的两块：
+
+- 素材组装：把 T16 的结构化诊断裁剪成 prompt 载荷。**skipped 维度不进
+  prompt**——LLM 看不到缺失的维度就不会编造它，这是防幻觉的第一道闸
+  （第二道闸在 validation.py 的生成后校验）。
+- 降级模板：持续违规时用规则直出拼一份零 LLM 的合规文案，是 T-P3-11
+  「绝不放行违规内容」的最终出口——有降级路径，fail-close 才不会变成
+  「永远失败」。
+"""
+
+from __future__ import annotations
+
+import json
+from collections.abc import Mapping, Sequence
+from typing import Any
+
+from pydantic import BaseModel
+
+from packages.shared.errors import AppError, ErrorCode
+
+#: 素材总长上限（字符）。REPORT_COPY 场景 max_tokens=3000，输入侧按架构
+#: §7.2 的裁剪思想控制：低重要度维度先丢，硬结论永不丢。
+MAX_MATERIAL_CHARS = 6000
+
+#: 硬结论维度：体/脸/季三大枚举是下游（记忆、Agent、换装推荐）共同依赖的
+#: 画像锚点，任何预算压力下都不裁——宁可 prompt 变长，不可结论缺失。
+HARD_DIMENSION_KEYS: frozenset[str] = frozenset({"body_shape", "face_shape", "color_season"})
+
+#: 维度裁剪优先级：数值越大越后裁。未登记的维度取默认优先级，新增维度
+#: 无需改这张表即可获得「先被裁」的保守待遇。
+_DIMENSION_PRIORITY: dict[str, int] = {
+    "body_shape": 100,
+    "face_shape": 100,
+    "color_season": 100,
+    "skin_tone": 80,
+    "color_palette": 80,
+    "style_impression": 60,
+    "figure_balance": 60,
+}
+_DEFAULT_PRIORITY = 50
+
+#: 外层 JSON 键（{"dimensions":[],"rule_hits":[]}）的固定序列化开销估计。
+_SERIALIZATION_OVERHEAD = 64
+
+
+class CopySection(BaseModel):
+    """报告正文章节。title 短语即可，文案主体放 paragraphs。"""
+
+    title: str
+    paragraphs: list[str]
+
+
+class ReportCopy(BaseModel):
+    """REPORT_COPY 场景的结构化输出 schema（宽进严出的取舍）。
+
+    字段越少，provider 的 JSON 约束成功率越高——解析失败会触发网关的
+    结构化重试，每次都是真金白银；文案的自由度全部放进 sections 的段落
+    文本，不靠 schema 枚举去硬卡。语义正确性不靠 schema：那属于生成后
+    校验（validation.py），两层各管各的。
+    """
+
+    summary: str
+    sections: list[CopySection]
+
+    def full_text(self) -> str:
+        """summary + 全部章节标题与段落，供校验器做全文扫描。"""
+        parts = [self.summary]
+        for section in self.sections:
+            parts.append(section.title)
+            parts.extend(section.paragraphs)
+        return "\n".join(parts)
+
+
+def coerce_parsed_copy(parsed: Any) -> ReportCopy:
+    """gateway ``Completion.parsed`` → ReportCopy；类型不符即网关契约破裂。
+
+    正常不会走到 raise（gateway 带 schema 调用成功即保证已过模型校验），
+    留这道闸是为了把意外的 TypeError 统一成 AppError，方便编排层按
+    错误码落任务失败与退款。
+    """
+    if not isinstance(parsed, ReportCopy):
+        msg = "REPORT_COPY 场景的 completion 未携带 ReportCopy 结构"
+        raise AppError(ErrorCode.LLM_STRUCTURED_OUTPUT_FAILED, message=msg)
+    return parsed
+
+
+def _compact_dimension(dim: Mapping[str, Any]) -> dict[str, Any]:
+    """维度瘦身到 prompt 需要的字段：key/value/evidence + 两位小数置信度。
+
+    confidence 保留（低置信维度要用不确定表述，模型必须能看到它），只把
+    视觉层带来的浮点尾数舍掉，省字符且避免无意义的精度噪音。
+    """
+    confidence = dim.get("confidence")
+    compact: dict[str, Any] = {
+        "key": dim.get("key"),
+        "value": dim.get("value"),
+        "evidence": list(dim.get("evidence") or []),
+    }
+    if isinstance(confidence, (int, float)):
+        compact["confidence"] = round(float(confidence), 2)
+    return compact
+
+
+def _compact_rule_hits(rule_hits: Sequence[Any]) -> list[dict[str, Any]]:
+    """规则命中瘦身：保留 key/维度/建议/解释/硬禁忌标记，其余字段不进 prompt。"""
+    compacted: list[dict[str, Any]] = []
+    for hit in rule_hits:
+        if not isinstance(hit, Mapping):
+            continue
+        compact: dict[str, Any] = {"rule_key": hit.get("rule_key") or hit.get("key")}
+        for field in ("dimension", "suggestion", "explanation", "is_hard"):
+            if hit.get(field) is not None:
+                compact[field] = hit[field]
+        compacted.append(compact)
+    return compacted
+
+
+def _compact_json(value: Any) -> str:
+    return json.dumps(value, ensure_ascii=False, separators=(",", ":"))
+
+
+def select_prompt_dimensions(
+    result_diagnosis: Mapping[str, Any],
+    *,
+    max_chars: int = MAX_MATERIAL_CHARS,
+) -> tuple[list[dict[str, Any]], list[str]]:
+    """取 status=ok 的维度并做超长裁剪；返回 (保留维度, 被裁掉的 key 列表)。
+
+    - skipped/error 维度直接不进素材（防幻觉第一道闸，见模块 docstring）；
+    - 超长时按优先级升序丢非硬结论维度（低重要度先丢，硬结论永不丢）；
+    - 若保留的全是硬结论仍超长，不再裁：硬结论缺位的代价远大于长 prompt，
+      由 provider 侧 max_tokens 兜底。此时 dropped 可能为空但载荷超长。
+    """
+    raw_dimensions = [
+        d for d in (result_diagnosis.get("dimensions") or []) if isinstance(d, Mapping)
+    ]
+    kept = [_compact_dimension(d) for d in raw_dimensions if d.get("status") == "ok"]
+    hits_len = len(_compact_json(_compact_rule_hits(result_diagnosis.get("rule_hits") or [])))
+    budget = max(max_chars - hits_len - _SERIALIZATION_OVERHEAD, 0)
+
+    droppable = sorted(
+        (d for d in kept if str(d.get("key")) not in HARD_DIMENSION_KEYS),
+        key=lambda d: (
+            _DIMENSION_PRIORITY.get(str(d.get("key")), _DEFAULT_PRIORITY),
+            str(d.get("key")),
+        ),
+    )
+    dropped: list[str] = []
+    while droppable and len(_compact_json(kept)) > budget:
+        victim = droppable.pop(0)
+        kept.remove(victim)
+        dropped.append(str(victim["key"]))
+    return kept, dropped
+
+
+def compose_prompt_material(
+    result_diagnosis: Mapping[str, Any],
+    *,
+    max_chars: int = MAX_MATERIAL_CHARS,
+) -> tuple[str, list[str]]:
+    """组装 {dimensions_json} 载荷，连同被裁掉的维度 key 一起返回。
+
+    CopyStage 一次调用拿全两者：载荷进 prompt，dropped_keys 落进
+    result["copy"] 供查询侧与排查用（用户看到的报告少了哪些维度）。
+    """
+    kept, dropped = select_prompt_dimensions(result_diagnosis, max_chars=max_chars)
+    payload = {
+        "dimensions": kept,
+        "rule_hits": _compact_rule_hits(result_diagnosis.get("rule_hits") or []),
+    }
+    return _compact_json(payload), dropped
+
+
+def build_prompt_material(result_diagnosis: Mapping[str, Any]) -> str:
+    """任务契约签名：只返回素材字符串（dropped_keys 用 compose_prompt_material 取）。"""
+    material, _dropped = compose_prompt_material(result_diagnosis)
+    return material
+
+
+#: 维度 key → 中文标签。模板文案与校验器共用的最小字典；T16 若在维度里
+#: 自带 label 字段则优先用它的。
+_KEY_LABELS: dict[str, str] = {
+    "body_shape": "体型",
+    "face_shape": "脸型",
+    "color_season": "季型",
+    "skin_tone": "肤色",
+    "color_palette": "色板",
+}
+
+_FALLBACK_SUMMARY = (
+    "以下内容根据本次分析的结构化结论整理，每条结论都来自规则引擎对诊断值的直接匹配。"
+)
+_FALLBACK_EMPTY_NOTE = "本次分析暂无可用结论，请稍后重新生成报告。"
+
+
+def dimension_label(dim: Mapping[str, Any]) -> str:
+    """维度的中文标签：优先维度自带 label，其次按 key 查表，兜底返回 key。"""
+    label = dim.get("label")
+    if isinstance(label, str) and label:
+        return label
+    key = str(dim.get("key") or "")
+    return _KEY_LABELS.get(key, key or "维度")
+
+
+def _evidence_text(dim: Mapping[str, Any]) -> str:
+    evidence = dim.get("evidence")
+    if isinstance(evidence, Sequence) and not isinstance(evidence, str) and evidence:
+        return "判定依据：" + "；".join(str(item) for item in evidence)
+    if isinstance(evidence, str) and evidence:
+        return f"判定依据：{evidence}"
+    return ""
+
+
+def render_fallback_copy(
+    dimensions: Sequence[Any],
+    rule_hits: Sequence[Any],
+) -> ReportCopy:
+    """降级模板：按维度 value/evidence + 规则 suggestion/explanation 拼文案，零 LLM。
+
+    持续违规（重试一次仍不过）时的最终出口。模板只做受控字符串拼接、
+    逐字复述诊断值与规则说明，不含任何自由生成文本，因此天然通过词表/
+    一致性/绝对化三重校验——ValidationStage 仍会对它跑一遍校验做
+    fail-close 自证（见 validation_stage.py）。
+    """
+    ok_dims = [d for d in dimensions if isinstance(d, Mapping) and d.get("status") == "ok"]
+
+    conclusion_paragraphs = []
+    for dim in ok_dims:
+        value = dim.get("value") or ""
+        para = f"你的{dimension_label(dim)}结论为「{value}」。{_evidence_text(dim)}".rstrip()
+        conclusion_paragraphs.append(para)
+
+    advice_paragraphs = []
+    for hit in rule_hits:
+        if not isinstance(hit, Mapping):
+            continue
+        parts = [str(hit[field]) for field in ("suggestion", "explanation") if hit.get(field)]
+        if parts:
+            advice_paragraphs.append("· " + "——".join(parts))
+
+    sections: list[CopySection] = []
+    if conclusion_paragraphs:
+        sections.append(CopySection(title="诊断结论", paragraphs=conclusion_paragraphs))
+    if advice_paragraphs:
+        sections.append(CopySection(title="搭配建议", paragraphs=advice_paragraphs))
+    if not sections:
+        sections.append(CopySection(title="说明", paragraphs=[_FALLBACK_EMPTY_NOTE]))
+    return ReportCopy(summary=_FALLBACK_SUMMARY, sections=sections)

packages/report/copy_stage.py

@@ -0,0 +1,63 @@
+"""流水线阶段：LLM 文案生成（架构 §7.1 步骤 17）。
+
+只生成、不校验——生成后校验是下一阶段（步骤 18）的独立职责。把两者
+塞进同一个阶段会让「校验不可跳过」失去结构保证：跳过一个阶段是配置
+错误（会被装配层拦下），跳过半段代码不是。
+"""
+
+from __future__ import annotations
+
+from packages.llm.types import Scene
+from packages.report.copy import ReportCopy, coerce_parsed_copy, compose_prompt_material
+from packages.report.pipeline import ReportContext
+from packages.shared.errors import AppError, ErrorCode
+from packages.shared.logging import get_logger
+
+logger = get_logger(__name__)
+
+
+class CopyStage:
+    """name="copy"：调 REPORT_COPY 场景生成结构化文案，产物落 result["copy"]。
+
+    产物键：summary / sections / model / prompt_version / dropped_keys。
+    dropped_keys 记录因超长被裁出 prompt 的维度——查询侧据此知道报告
+    少了哪些维度，避免「报告里没有」被当成「诊断说没有」。
+    """
+
+    name = "copy"
+
+    async def run(self, ctx: ReportContext) -> None:
+        diagnosis = ctx.report.result.get("diagnosis")
+        if not isinstance(diagnosis, dict):
+            # 没有 diagnosis 产物说明流水线装配顺序错了（copy 排在了 diagnosis 前面）
+            raise AppError(
+                ErrorCode.TASK_STATE_INVALID,
+                message="copy 阶段缺少 diagnosis 产物（流水线阶段顺序错误）",
+            )
+        if ctx.gateway is None:
+            raise AppError(ErrorCode.CONFIG_ERROR, message="报告流水线缺少 LLM 网关")
+
+        material, dropped_keys = compose_prompt_material(diagnosis)
+        completion = await ctx.gateway.complete(
+            scene=Scene.REPORT_COPY,
+            prompt_vars={"dimensions_json": material},
+            schema=ReportCopy,
+            user_id=ctx.report.user_id,
+            ref_type="report",
+            ref_id=str(ctx.report.report_id),
+        )
+        copy = coerce_parsed_copy(completion.parsed)
+
+        ctx.report.result["copy"] = {
+            "summary": copy.summary,
+            "sections": [section.model_dump() for section in copy.sections],
+            "model": completion.model,
+            "prompt_version": completion.prompt_version,
+            "dropped_keys": dropped_keys,
+        }
+        logger.info(
+            "report_copy_generated",
+            report_id=ctx.report.report_id,
+            model=completion.model,
+            dropped_dimensions=len(dropped_keys),
+        )

packages/report/validation.py

@@ -0,0 +1,352 @@
+"""生成后校验的纯逻辑（架构 §7.1 步骤 18，T-P3-10 / T-P3-11 🔴）。
+
+prompt 红线管「生成前」，本模块管「生成后」：LLM 输出永远不可信任，
+三重校验（贬损词表 / 文案-结论一致性 / 绝对化用语）全过才放行，
+ValidationStage 据此决定重试与降级。与 gateway 的结构化重试是两层独立
+机制：那层管 JSON 解析合不合法，这层管内容合不合规，互不共享计数——
+解析成功但内容违规，网关不会替我们发现。
+"""
+
+from __future__ import annotations
+
+import re
+from collections.abc import Mapping, Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+from packages.report.copy import ReportCopy, dimension_label
+
+#: 贬损词表（初版内嵌版，正式版归合规词库任务 T-P0-07，届时改为读库+后台可维护）。
+#: 来源：《合规与安全方案》§3「内容伦理红线」——把「描述而非评判」「不使用
+#: 缺陷框架」翻译成具体词表。命中按子串匹配：「有点胖」「微胖」都命中「胖」，
+#: 不做词干（中文无明显词干，初版宁多命中不放过）。
+PEJORATIVE_TERMS: tuple[str, ...] = (
+    # 身材/外貌贬损（§3 反例：「你的脸型有点大，不好看」「腿短是你的缺点」）
+    "胖",
+    "肥",
+    "臃肿",
+    "丑",
+    "难看",
+    "土气",
+    "显老",
+    "老气",
+    "矮",
+    "腿短",
+    "短腿",
+    "水桶腰",
+    "大饼脸",
+    "虎背熊腰",
+    # 缺陷框架（§3：不使用缺陷框架——特征是比例事实，不是缺点）
+    "缺陷",
+    "缺点",
+    "瑕疵",
+    "短板",
+    "硬伤",
+    "失败",
+    "悲剧",
+    "灾难",
+    # 评判性观感
+    "廉价",
+    "地摊货",
+    "邋遢",
+    "呆板",
+    "死板",
+    "笨拙",
+    # 医疗/身材干预越界（§3：不做医疗判断 → 一律转向穿搭与色彩建议）
+    "减肥",
+    "减重",
+    "瘦身",
+    "整形",
+    "整容",
+    "就医",
+    "皮肤病",
+)
+
+#: 绝对化用语（广告法红线；system prompt 已禁，这里是兜底——prompt 不是
+#: 约束保证，只是愿望）。「最」不取单字：「最近/最终」等中性词误杀率高，
+#: 取营销语境的高频复合词，误报代价见 check_unsupported_advice 的取舍注释。
+ABSOLUTE_CLAIM_TERMS: tuple[str, ...] = (
+    "最好",
+    "最佳",
+    "最优",
+    "最合适",
+    "最适合",
+    "最完美",
+    "最显瘦",
+    "第一",
+    "唯一",
+    "彻底",
+    "百分百",
+    "百分之百",
+    "绝对",
+    "完美",
+    "永不过时",
+    "无可替代",
+    "独一无二",
+)
+
+#: 校验类别 → 展平违规清单时的人读前缀。
+CHECK_PREFIXES: dict[str, str] = {
+    "pejorative": "贬损词",
+    "consistency": "结论不一致",
+    "absolute": "绝对化用语",
+}
+
+_HEX_COLOR_RE = re.compile(r"#[0-9a-fA-F]{6}\b")
+_ADVICE_MARKERS = ("建议", "推荐")
+
+
+@dataclass(frozen=True, slots=True)
+class EnumCategory:
+    """体型/脸型/季型三类枚举维度：key 识别串 + 全量枚举词 + 同义别名。
+
+    一致性规则①的依据：这三类的结论词是全量枚举（需求文档 §4 维度字典），
+    文案里出现「同类的另一个枚举值」即可断定与诊断矛盾。
+    """
+
+    label: str
+    key_tokens: tuple[str, ...]
+    vocabulary: tuple[str, ...]
+    #: 诊断值 → 同义说法（诊断说 X 型、文案写沙漏型不算矛盾）。
+    aliases: Mapping[str, tuple[str, ...]] = field(default_factory=dict)
+
+
+_ENUM_CATEGORIES: tuple[EnumCategory, ...] = (
+    EnumCategory(
+        label="体型",
+        key_tokens=("body_shape", "体型"),
+        vocabulary=(
+            "A型",
+            "H型",
+            "O型",
+            "T型",
+            "X型",
+            "苹果型",
+            "梨型",
+            "沙漏型",
+            "倒三角型",
+            "矩形型",
+        ),
+        aliases={
+            "O型": ("苹果型",),
+            "A型": ("梨型",),
+            "X型": ("沙漏型",),
+            "T型": ("倒三角型",),
+            "H型": ("矩形型", "直筒型"),
+        },
+    ),
+    EnumCategory(
+        label="脸型",
+        key_tokens=("face_shape", "脸型"),
+        # 只收「X脸」复合词：单字（圆/方/长）会把「轮廓偏圆润」这类合规
+        # 描述性用语（合规 §3 的正例）误杀。
+        vocabulary=(
+            "鹅蛋脸",
+            "圆形脸",
+            "圆脸",
+            "方形脸",
+            "方脸",
+            "长形脸",
+            "长脸",
+            "心形脸",
+            "菱形脸",
+            "梨形脸",
+        ),
+        aliases={
+            "鹅蛋": ("鹅蛋脸", "鹅蛋型"),
+            "圆": ("圆脸", "圆形脸"),
+            "方": ("方脸", "方形脸"),
+            "长": ("长脸", "长形脸"),
+            "心形": ("心形脸",),
+            "菱形": ("菱形脸",),
+            "梨形": ("梨形脸",),
+        },
+    ),
+    EnumCategory(
+        label="季型",
+        key_tokens=("color_season", "season", "季型"),
+        # 十二季细分是自研描述性命名（TD-18），无法全量枚举，只收四季
+        # 粗分词；细分红利是值本身进 allowed 集合，复述永远合法。
+        vocabulary=("春季型", "夏季型", "秋季型", "冬季型", "春型", "夏型", "秋型", "冬型"),
+        aliases={
+            "春": ("春型", "春季型"),
+            "夏": ("夏型", "夏季型"),
+            "秋": ("秋型", "秋季型"),
+            "冬": ("冬型", "冬季型"),
+        },
+    ),
+)
+
+
+def _normalize_for_match(text: str) -> str:
+    """去空白 + 拉丁字母大写：「x 型」「X型」归一成同一写法再比较。"""
+    return re.sub(r"\s+", "", text).upper()
+
+
+def check_pejorative(text: str) -> list[str]:
+    """全文扫描贬损词表，返回命中的词（按词表顺序）。"""
+    return [term for term in PEJORATIVE_TERMS if term in text]
+
+
+def check_absolute_claims(text: str) -> list[str]:
+    """绝对化用语兜底检查，返回命中的词。"""
+    return [term for term in ABSOLUTE_CLAIM_TERMS if term in text]
+
+
+def _find_dimension(
+    dimensions: Sequence[Mapping[str, Any]], key_tokens: Sequence[str]
+) -> Mapping[str, Any] | None:
+    for dim in dimensions:
+        key = str(dim.get("key") or "")
+        if any(token in key for token in key_tokens):
+            return dim
+    return None
+
+
+def _check_enum_consistency(
+    copy_text_normalized: str, dimensions: Sequence[Mapping[str, Any]]
+) -> list[str]:
+    """规则①：体型/脸型/季型枚举词必须与诊断值同款（含同义别名）。
+
+    维度 skipped/无值时，文案出现该类任何枚举词都算编造——LLM 没看到
+    这个维度却写出了结论，正是素材组装要防的幻觉从校验侧兜底。
+    每类只报第一条命中，避免同一问题刷屏冲淡修正指令。
+    """
+    issues: list[str] = []
+    for category in _ENUM_CATEGORIES:
+        dim = _find_dimension(dimensions, category.key_tokens)
+        if dim is None:
+            continue
+        value = dim.get("value")
+        if dim.get("status") == "ok" and value:
+            canonical = str(value)
+            allowed = {_normalize_for_match(canonical)}
+            for alias in category.aliases.get(canonical, ()):  # 值本身可能就是别名键
+                allowed.add(_normalize_for_match(alias))
+            # 诊断值也可能是别名列里的写法（如存了「沙漏型」），反查一次
+            for canonical_key, alias_values in category.aliases.items():
+                if _normalize_for_match(canonical) == _normalize_for_match(canonical_key) or any(
+                    _normalize_for_match(canonical) == _normalize_for_match(a) for a in alias_values
+                ):
+                    allowed.add(_normalize_for_match(canonical_key))
+                    allowed.update(_normalize_for_match(a) for a in alias_values)
+        else:
+            allowed = set()
+        for token in category.vocabulary:
+            normalized_token = _normalize_for_match(token)
+            if normalized_token in copy_text_normalized and normalized_token not in allowed:
+                issues.append(
+                    f"{category.label}用词「{token}」与诊断值「{value or '未诊断'}」不一致"
+                )
+                break
+    return issues
+
+
+def _check_hex_colors(copy_text: str, dimensions: Sequence[Mapping[str, Any]]) -> list[str]:
+    """规则②：文案中的 #RRGGBB 必须来自诊断维度（色板）的取值集合。
+
+    色值集合从全部维度的 value/evidence 里递归收集而非只认某个 key——
+    T16 的色板维度 key 未定死，宽收集合只可能放宽校验，不会误杀合规色值。
+    """
+    used = {m.group(0).upper() for m in _HEX_COLOR_RE.finditer(copy_text)}
+    if not used:
+        return []
+    dimension_blob = str([dict(d) for d in dimensions])
+    allowed = {m.group(0).upper() for m in _HEX_COLOR_RE.finditer(dimension_blob)}
+    return [f"色值「{color}」不在诊断给出的色板取值范围内" for color in sorted(used - allowed)]
+
+
+def _check_unsupported_advice(
+    copy_text: str,
+    dimensions: Sequence[Mapping[str, Any]],
+    rule_hits: Sequence[Mapping[str, Any]] | None,
+) -> list[str]:
+    """规则③（宽松口径）：建议/推荐句式必须有规则命中支撑。
+
+    取舍：误报的代价是一次无谓重试甚至降级到模板（文案质量明显变差、
+    多花一次 LLM 调用），漏报的底线由贬损词表与枚举一致性兜住，因此只
+    标记两种可确定的情形：
+    ① 文案给出建议但规则引擎零命中（建议完全无据）；
+    ② 建议句点名了某维度，而该维度既无内嵌 rule_hits、也无顶层命中指向它。
+    其余（句式识别不全、支撑关系模糊）一律放行——宁漏勿误杀。
+    """
+    sentences = re.split(r"[。！？!?;\n]", copy_text)
+    advice_sentences = [s for s in sentences if any(marker in s for marker in _ADVICE_MARKERS)]
+    if not advice_sentences:
+        return []
+
+    top_level_dims = {
+        str(hit.get("dimension")) for hit in (rule_hits or []) if hit.get("dimension")
+    }
+    embedded_exists = any(dim.get("rule_hits") for dim in dimensions)
+
+    if not top_level_dims and not embedded_exists:
+        return ["文案包含建议/推荐表述，但规则引擎没有任何命中来支撑该建议"]
+
+    issues: list[str] = []
+    for sentence in advice_sentences:
+        snippet = sentence.strip()[:30]
+        for dim in dimensions:
+            if dim.get("status") != "ok":
+                continue
+            key = str(dim.get("key") or "")
+            label = dimension_label(dim)
+            if not label or label not in sentence:
+                continue
+            supported = bool(dim.get("rule_hits")) or key in top_level_dims
+            if not supported:
+                issues.append(f"建议句「{snippet}」提到{label}，但该维度没有规则命中支撑")
+    return issues
+
+
+def check_conclusion_consistency(
+    copy: ReportCopy,
+    dimensions: list[dict[str, Any]],
+    *,
+    rule_hits: Sequence[Mapping[str, Any]] | None = None,
+) -> list[str]:
+    """文案-结论一致性（T-P3-10，初版三条规则）。返回不一致清单。
+
+    ① 体型/脸型/季型枚举词必须与诊断值同款；维度未诊断时不得出现该类词；
+    ② 具体色值 #RRGGBB 必须在诊断（色板维度）取值集合内；
+    ③ 建议/推荐句式必须有规则命中支撑（宽松口径，见 _check_unsupported_advice）。
+
+    ``rule_hits`` 为可选参数：规则③的支撑判定兼容两种契约形态（命中内嵌
+    在维度里 / 顶层列表），两者都没有时按宽松口径不追责。
+    """
+    dict_dimensions = [d for d in dimensions if isinstance(d, Mapping)]
+    full_text = copy.full_text()
+    issues = _check_enum_consistency(_normalize_for_match(full_text), dict_dimensions)
+    issues.extend(_check_hex_colors(full_text, dict_dimensions))
+    issues.extend(_check_unsupported_advice(full_text, dict_dimensions, rule_hits))
+    return issues
+
+
+def collect_violations(
+    copy: ReportCopy,
+    dimensions: list[dict[str, Any]],
+    *,
+    rule_hits: Sequence[Mapping[str, Any]] | None = None,
+) -> dict[str, list[str]]:
+    """三重校验的汇总视图，键固定为 pejorative / consistency / absolute。"""
+    full_text = copy.full_text()
+    return {
+        "pejorative": check_pejorative(full_text),
+        "consistency": check_conclusion_consistency(copy, dimensions, rule_hits=rule_hits),
+        "absolute": check_absolute_claims(full_text),
+    }
+
+
+def validate_copy(
+    copy: ReportCopy,
+    dimensions: list[dict[str, Any]],
+    *,
+    rule_hits: Sequence[Mapping[str, Any]] | None = None,
+) -> list[str]:
+    """展平的违规清单，条目带类别前缀——可直接作为重试的修正指令。"""
+    breakdown = collect_violations(copy, dimensions, rule_hits=rule_hits)
+    return [
+        f"{CHECK_PREFIXES[key]}：{item}"
+        for key in ("pejorative", "consistency", "absolute")
+        for item in breakdown[key]
+    ]

packages/report/validation_stage.py

@@ -0,0 +1,173 @@
+"""流水线阶段：生成后校验（架构 §7.1 步骤 18，T-P3-10 / T-P3-11 🔴）。
+
+出口只有两种：``passed=true``（干净通过 / 重试后通过 / 降级模板），
+或抛错让整条流水线失败。**绝不放行未过校验的文案**——宁可任务失败退款，
+也不让一条贬损文案到达用户（T-P3-11 的硬性禁忌违反数必须为 0）。
+
+重试语义：这里的「重试一次」是**语义层**修正（把违规清单作为修正指令
+追加进 messages 重新生成），与 gateway 的结构化重试（解析层，JSON 合不
+合法）互不共享计数——解析成功但内容违规，网关不会替我们发现。
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from pydantic import ValidationError
+
+from packages.llm.types import Completion, Message, Scene
+from packages.report.copy import (
+    ReportCopy,
+    coerce_parsed_copy,
+    compose_prompt_material,
+    render_fallback_copy,
+)
+from packages.report.pipeline import ReportContext
+from packages.report.validation import validate_copy
+from packages.shared.errors import AppError, ErrorCode
+from packages.shared.logging import get_logger
+
+logger = get_logger(__name__)
+
+#: 语义层修正的最大重试次数（架构 §7.1：校验失败最多重试后降级）。
+#: 再多就是花钱买降级——持续违规说明模型在该素材上不可救。
+_MAX_SEMANTIC_RETRIES = 1
+
+#: 降级模板的 model 标记：查询侧/排查侧凭它知道这份文案没经过 LLM。
+_FALLBACK_MODEL = "fallback-template"
+
+
+def _clean_checks() -> dict[str, int]:
+    """全过时的 checks 计数（最终放行文案的违规数必然全零）。"""
+    return {"pejorative": 0, "consistency": 0, "absolute": 0}
+
+
+def _copy_payload(
+    copy: ReportCopy, base: dict[str, Any], model: str, prompt_version: str | None
+) -> dict[str, Any]:
+    """组装 result["copy"]：文案主体换新，dropped_keys 等诊断类元数据沿用。"""
+    return {
+        **base,
+        "summary": copy.summary,
+        "sections": [section.model_dump() for section in copy.sections],
+        "model": model,
+        "prompt_version": prompt_version,
+    }
+
+
+class ValidationStage:
+    """name="validation"：三重校验 → 违规则重试一次 → 仍违规则降级模板。
+
+    产物落 result["validation"]：passed 恒为 true（或本阶段抛错），
+    degraded=true 表示最终文案是规则直出模板；violations_first_pass
+    留存第一轮违规清单，供报告查询侧与质量分析回看降级原因。
+    """
+
+    name = "validation"
+
+    async def run(self, ctx: ReportContext) -> None:
+        result = ctx.report.result
+        diagnosis = result.get("diagnosis")
+        copy_payload = result.get("copy")
+        if not isinstance(diagnosis, dict) or not isinstance(copy_payload, dict):
+            raise AppError(
+                ErrorCode.TASK_STATE_INVALID,
+                message="validation 阶段缺少 copy/diagnosis 产物（流水线阶段顺序错误）",
+            )
+        try:
+            copy = ReportCopy.model_validate(copy_payload)
+        except ValidationError as exc:
+            raise AppError(
+                ErrorCode.TASK_STATE_INVALID,
+                message=f"copy 产物结构不合法：{exc.error_count()} 处校验错误",
+            ) from exc
+
+        dimensions = [d for d in (diagnosis.get("dimensions") or []) if isinstance(d, dict)]
+        rule_hits = [h for h in (diagnosis.get("rule_hits") or []) if isinstance(h, dict)]
+
+        violations = validate_copy(copy, dimensions, rule_hits=rule_hits)
+        if not violations:
+            result["validation"] = {"passed": True, "checks": _clean_checks()}
+            return
+
+        logger.warning(
+            "report_copy_violations_found",
+            report_id=ctx.report.report_id,
+            violations=violations,
+        )
+        retry = await self._retry_with_fixes(ctx, diagnosis, violations)
+        if retry is not None:
+            retry_copy, retry_completion = retry
+            retry_violations = validate_copy(retry_copy, dimensions, rule_hits=rule_hits)
+            if not retry_violations:
+                result["copy"] = _copy_payload(
+                    retry_copy,
+                    copy_payload,
+                    model=retry_completion.model,
+                    prompt_version=retry_completion.prompt_version,
+                )
+                result["validation"] = {
+                    "passed": True,
+                    "retried": True,
+                    "violations_first_pass": violations,
+                    "checks": _clean_checks(),
+                }
+                return
+            logger.warning(
+                "report_copy_retry_still_violating",
+                report_id=ctx.report.report_id,
+                violations=retry_violations,
+            )
+
+        # 降级：规则直出模板，零 LLM。模板由受控拼接生成，理论必过校验；
+        # 仍不过（词表演进/诊断值自带违规词）则 fail-close：抛错让流水线
+        # 失败并退款——这正是「任何出口要么 passed 要么抛错」的含义。
+        fallback = render_fallback_copy(dimensions, rule_hits)
+        residual = validate_copy(fallback, dimensions, rule_hits=rule_hits)
+        if residual:
+            raise AppError(
+                ErrorCode.INTERNAL_ERROR,
+                message=f"降级模板文案未通过校验，拒绝放行：{residual}",
+            )
+        result["copy"] = _copy_payload(
+            fallback, copy_payload, model=_FALLBACK_MODEL, prompt_version=None
+        )
+        result["validation"] = {
+            "passed": True,
+            "degraded": True,
+            "violations_first_pass": violations,
+            "checks": _clean_checks(),
+        }
+        logger.warning(
+            "report_copy_degraded_to_template",
+            report_id=ctx.report.report_id,
+            first_pass_violations=len(violations),
+        )
+
+    async def _retry_with_fixes(
+        self,
+        ctx: ReportContext,
+        diagnosis: dict[str, Any],
+        violations: list[str],
+    ) -> tuple[ReportCopy, Completion] | None:
+        """带修正指令重新生成一次；网关层故障原样抛出（属基础设施错误，
+        不该被降级模板掩盖——那是内容问题出口，不是可用性出口）。"""
+        if ctx.gateway is None:
+            raise AppError(ErrorCode.CONFIG_ERROR, message="报告流水线缺少 LLM 网关")
+        material, _dropped = compose_prompt_material(diagnosis)
+        correction = (
+            "上一版报告文案未通过内容校验，违规清单如下：\n"
+            + "\n".join(f"- {item}" for item in violations)
+            + "\n请严格遵守系统提示词的内容红线，重新生成整份文案：只复述给定诊断值，"
+            "删除全部违规表述，不要解释修改过程。"
+        )
+        completion = await ctx.gateway.complete(
+            scene=Scene.REPORT_COPY,
+            messages=[Message(role="user", content=correction)],
+            prompt_vars={"dimensions_json": material},
+            schema=ReportCopy,
+            user_id=ctx.report.user_id,
+            ref_type="report",
+            ref_id=str(ctx.report.report_id),
+        )
+        return coerce_parsed_copy(completion.parsed), completion

prompts/report_copy/v1.txt

@@ -1,17 +1,20 @@
 [system]
-你为「美搭」的美学报告撰写文案。输入是结构化诊断结论与规则命中结果，输出是给用户阅读的报告正文。
+你是「美搭」的形象顾问，为用户的美学报告撰写文案。输入是结构化诊断结论与规则命中结果，输出是给用户阅读的报告正文。
 
-硬性约束：
-- 只能使用输入中给出的诊断值与知识条目，不得引入任何未提供的事实。
-- 不得违反输入中标记为硬禁忌的规则，一条都不行。
+内容红线（违反任何一条，整篇文案即作废重写）：
+- 结论只能复述输入中给出的诊断值，不得新增、改写或推断任何维度结论；输入未提供的维度一律不提。
+- 描述而非评判：客观陈述特征与搭配路径，不评判外貌。
+- 不出现任何贬损性词汇，例如胖、丑、矮、土气、显老、缺陷一类表述。
+- 不使用绝对化用语，例如最好、第一、彻底、百分百。
 - 不做医疗、减重、整形、命理相关的结论或暗示。
-- 不评判外貌，不使用贬损性词汇；描述而非评价。
-- 置信度低的维度必须使用不确定表述，不得断言。
+- 置信度低的维度使用不确定表述，不得断言。
 
 写作要求：
-- 简体中文，平实具体，避免营销腔与绝对化用语。
+- 简体中文，平实具体，避免营销腔。
 - 每个结论后面跟一句「为什么」，依据来自输入的规则命中说明。
 - 专业术语首次出现时补白话解释。
 
 [user]
-{payload}
+以下是结构化诊断素材（仅含已成功诊断的维度与规则命中结果），请据此生成报告文案：
+
+{dimensions_json}

tests/integration/test_report_copy_stage.py

@@ -0,0 +1,268 @@
+"""报告 copy / validation 两阶段的集成验证（FakeProvider 替换真实 provider）。
+
+不依赖真实 PG/Redis（本组用例只读写内存中的 RpReport.result），但按批次
+约定放 integration：这里组装的是真实网关栈（PromptRegistry 读仓库里的
+report_copy/v1.txt、路由、预算、结构化解析），验证「prompt 模板 →
+gateway 结构化输出 → 阶段产物 → 语义重试/降级」的完整咬合，纯单测的桩
+盖不住这条链。不测整条流水线编排（T15 已覆盖），只直接构造 ctx 调阶段。
+"""
+
+from __future__ import annotations
+
+import itertools
+import json
+from collections.abc import Sequence
+from decimal import Decimal
+from pathlib import Path
+from typing import Any, cast
+
+import pytest
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from packages.llm.budget import DailyBudgetGuard, InMemoryCostCounter
+from packages.llm.circuit import CircuitBreakerRegistry
+from packages.llm.cost import InMemoryAICallRecorder, ModelRate, RateTable
+from packages.llm.gateway import LLMGateway
+from packages.llm.prompts import PromptRegistry
+from packages.llm.routing import SceneRouter, build_scene_configs
+from packages.llm.types import ProviderRequest, ProviderResponse, Usage
+from packages.report.copy import ReportCopy
+from packages.report.copy_stage import CopyStage
+from packages.report.pipeline import ReportContext
+from packages.report.validation import check_pejorative, validate_copy
+from packages.report.validation_stage import ValidationStage
+from packages.shared.config import get_settings
+from packages.shared.errors import AppError, ErrorCode
+from packages.shared.models import RpInput, RpReport
+
+pytestmark = [pytest.mark.integration]
+
+REPO_PROMPTS = Path(__file__).resolve().parents[2] / "prompts"
+
+
+class QueuedProvider:
+    """按队列返回预置响应，耗尽后重复最后一个——最后一条即「持续脏」语义。"""
+
+    def __init__(self, outcomes: Sequence[ProviderResponse]) -> None:
+        self._outcomes = list(outcomes)
+        self.requests: list[ProviderRequest] = []
+
+    @property
+    def name(self) -> str:
+        return "p1"
+
+    @property
+    def default_model(self) -> str:
+        return "fake-model"
+
+    async def complete(self, request: ProviderRequest) -> ProviderResponse:
+        self.requests.append(request)
+        return self._outcomes[min(len(self.requests) - 1, len(self._outcomes) - 1)]
+
+    async def aclose(self) -> None:
+        return None
+
+
+_RATE = ModelRate(
+    input_cents_per_mtok=Decimal("100"),
+    cached_input_cents_per_mtok=Decimal("10"),
+    output_cents_per_mtok=Decimal("200"),
+)
+
+
+def _build_gateway(provider: QueuedProvider) -> LLMGateway:
+    return LLMGateway(
+        providers={"p1": provider},
+        router=SceneRouter(build_scene_configs(primary_provider="p1", backup_provider=None)),
+        prompts=PromptRegistry(REPO_PROMPTS),
+        rates=RateTable({"fake-model": _RATE}, default=_RATE),
+        recorder=InMemoryAICallRecorder(),
+        budget=DailyBudgetGuard(
+            InMemoryCostCounter(), daily_budget_cents=Decimal("100000"), enabled=True
+        ),
+        breakers=CircuitBreakerRegistry(),
+        id_factory=itertools.count(1).__next__,
+    )
+
+
+def _diagnosis() -> dict[str, Any]:
+    return {
+        "dimensions": [
+            {
+                "key": "body_shape",
+                "status": "ok",
+                "value": "X型",
+                "confidence": 0.76,
+                "evidence": ["shoulder_hip_ratio:1.02", "waist_hip_ratio:0.71"],
+                "rule_hits": ["R-BODY-0007"],
+            },
+            {
+                "key": "face_shape",
+                "status": "ok",
+                "value": "鹅蛋",
+                "confidence": 0.81,
+                "evidence": ["face_ratio:1.35"],
+                "rule_hits": [],
+            },
+            {
+                "key": "color_season",
+                "status": "ok",
+                "value": "浅暖柔和型",
+                "confidence": 0.82,
+                "evidence": ["contrast:low"],
+                "rule_hits": ["R-COLOR-0021"],
+            },
+            {
+                # skipped 维度：不应出现在任何一次请求的素材里（防幻觉得到结构保证）
+                "key": "posture_check",
+                "status": "skipped",
+                "value": "",
+                "confidence": 0.0,
+                "evidence": [],
+                "rule_hits": [],
+            },
+        ],
+        "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",
+            },
+            {
+                "rule_key": "R-COLOR-0021",
+                "dimension": "color_season",
+                "suggestion": "主色选择低饱和暖调",
+                "explanation": "浅暖柔和型对比弱，低饱和暖调更显匀净",
+            },
+        ],
+    }
+
+
+CLEAN_COPY: dict[str, Any] = {
+    "summary": "本次分析完成：体型、脸型与色彩结论均来自照片与问卷的结构化测算。",
+    "sections": [
+        {
+            "title": "体型与比例",
+            "paragraphs": [
+                "你的体型结论为 X 型，肩腰臀比例均衡。",
+                "建议优先选择收腰版型，突出腰部线条。",
+            ],
+        },
+        {"title": "脸型与领线", "paragraphs": ["脸型轮廓流畅，接近鹅蛋脸，多数领型都能协调。"]},
+        {"title": "色彩方向", "paragraphs": ["色彩结论为浅暖柔和型，建议主色选择低饱和暖调。"]},
+    ],
+}
+
+DIRTY_COPY: dict[str, Any] = {
+    "summary": "整体看身材偏微胖，需要靠版型修饰，收腰款最合适。",
+    "sections": [
+        {"title": "体型小结", "paragraphs": ["你的体型属于 X 型，轮廓数据完整。"]},
+    ],
+}
+
+
+def _response(payload: dict[str, Any]) -> ProviderResponse:
+    return ProviderResponse(
+        text=json.dumps(payload, ensure_ascii=False),
+        model="fake-model",
+        usage=Usage(tokens_in=1200, tokens_out=400, tokens_cached=900),
+    )
+
+
+def _make_ctx(gateway: LLMGateway, *, result: dict[str, Any] | None = None) -> ReportContext:
+    report = RpReport(
+        report_id=910001,
+        user_id=42,
+        result=result if result is not None else {"diagnosis": _diagnosis()},
+    )
+    report_input = RpInput(input_id=910002, report_id=910001)
+    return ReportContext(
+        report=report,
+        report_input=report_input,
+        # 本组用例的阶段只读写内存中的 result，不触库；session 仅为满足契约形状
+        session=cast("AsyncSession", None),
+        settings=get_settings(),
+        gateway=gateway,
+    )
+
+
+async def test_clean_llm_copy_passes_both_stages() -> None:
+    provider = QueuedProvider([_response(CLEAN_COPY)])
+    ctx = _make_ctx(_build_gateway(provider))
+
+    await CopyStage().run(ctx)
+    await ValidationStage().run(ctx)
+
+    copy_payload = ctx.report.result["copy"]
+    assert copy_payload["model"] == "fake-model"
+    assert copy_payload["prompt_version"] == "v1"
+    assert copy_payload["dropped_keys"] == []
+    assert ctx.report.result["validation"] == {
+        "passed": True,
+        "checks": {"pejorative": 0, "consistency": 0, "absolute": 0},
+    }
+    assert len(provider.requests) == 1  # 干净文案不需要任何重试
+    # skipped 维度不出现在发给 LLM 的任何消息里（防幻觉第一道闸）
+    all_text = "".join(message.content for message in provider.requests[0].messages)
+    assert "posture_check" not in all_text
+    assert "X型" in all_text
+
+
+async def test_dirty_once_copy_passes_after_one_semantic_retry() -> None:
+    provider = QueuedProvider([_response(DIRTY_COPY), _response(CLEAN_COPY)])
+    ctx = _make_ctx(_build_gateway(provider))
+
+    await CopyStage().run(ctx)
+    await ValidationStage().run(ctx)
+
+    validation = ctx.report.result["validation"]
+    assert validation["passed"] is True
+    assert validation["retried"] is True
+    assert "degraded" not in validation
+    assert validation["violations_first_pass"]  # 第一轮违规清单非空（微胖/最合适）
+    # 重试通过后放行的是修正版文案
+    assert ctx.report.result["copy"]["summary"] == CLEAN_COPY["summary"]
+    # 恰好两次调用：初稿 + 一次语义修正（与网关解析层重试互不叠加）
+    assert len(provider.requests) == 2
+    retry_text = "".join(message.content for message in provider.requests[1].messages)
+    assert "贬损词" in retry_text  # 修正指令带上了违规清单
+
+
+async def test_persistently_dirty_llm_degrades_to_zero_pejorative_copy() -> None:
+    provider = QueuedProvider([_response(DIRTY_COPY)])  # 恒返脏文案
+    ctx = _make_ctx(_build_gateway(provider))
+
+    await CopyStage().run(ctx)
+    await ValidationStage().run(ctx)
+
+    validation = ctx.report.result["validation"]
+    assert validation["passed"] is True
+    assert validation["degraded"] is True
+    assert validation["violations_first_pass"]
+    copy_payload = ctx.report.result["copy"]
+    assert copy_payload["model"] == "fallback-template"
+    assert copy_payload["prompt_version"] is None
+    # 黄金断言（T-P3-11）：最终触达用户的文案零贬损词、零违规
+    final_copy = ReportCopy.model_validate(copy_payload)
+    assert check_pejorative(final_copy.full_text()) == []
+    diagnosis = ctx.report.result["diagnosis"]
+    assert (
+        validate_copy(final_copy, diagnosis["dimensions"], rule_hits=diagnosis["rule_hits"]) == []
+    )
+    # 初稿 + 一次重试，之后停止烧钱、改走零 LLM 模板
+    assert len(provider.requests) == 2
+
+
+async def test_stages_fail_fast_when_predecessor_output_is_missing() -> None:
+    gateway = _build_gateway(QueuedProvider([_response(CLEAN_COPY)]))
+
+    ctx = _make_ctx(gateway, result={})
+    with pytest.raises(AppError) as copy_exc:
+        await CopyStage().run(ctx)
+    assert copy_exc.value.code == ErrorCode.TASK_STATE_INVALID
+
+    ctx_without_copy = _make_ctx(gateway, result={"diagnosis": _diagnosis()})
+    with pytest.raises(AppError) as validation_exc:
+        await ValidationStage().run(ctx_without_copy)
+    assert validation_exc.value.code == ErrorCode.TASK_STATE_INVALID

tests/unit/test_report_copy.py

@@ -0,0 +1,211 @@
+"""报告文案纯逻辑：素材组装（skipped 不进 / 超长裁剪 / 硬结论不裁）与降级模板。"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from packages.report.copy import (
+    build_prompt_material,
+    compose_prompt_material,
+    render_fallback_copy,
+)
+from packages.report.validation import validate_copy
+
+
+def _dim(
+    key: str,
+    value: str = "X型",
+    *,
+    status: str = "ok",
+    confidence: float = 0.82,
+    evidence: tuple[str, ...] = ("shoulder_hip_ratio:1.02",),
+    rule_hits: tuple[str, ...] = ("R-BODY-0007",),
+) -> dict[str, Any]:
+    return {
+        "key": key,
+        "status": status,
+        "value": value,
+        "confidence": confidence,
+        "evidence": list(evidence),
+        "rule_hits": list(rule_hits),
+    }
+
+
+def _diagnosis(
+    *dimensions: dict[str, Any], rule_hits: list[dict[str, Any]] | None = None
+) -> dict[str, Any]:
+    return {"dimensions": list(dimensions), "rule_hits": rule_hits or []}
+
+
+# ---------------------------------------------------------------------------
+# 素材组装：skipped 不进 prompt（防幻觉第一道闸）
+# ---------------------------------------------------------------------------
+
+
+def test_build_prompt_material_excludes_skipped_dimensions() -> None:
+    diagnosis = _diagnosis(
+        _dim("body_shape", "X型"),
+        _dim("face_shape", "鹅蛋"),
+        _dim("occasion_preference", status="skipped", value=""),
+    )
+
+    material = build_prompt_material(diagnosis)
+    payload = json.loads(material)
+
+    keys = [d["key"] for d in payload["dimensions"]]
+    assert keys == ["body_shape", "face_shape"]
+    # skipped 维度连同它的 key/value 都不能出现在载荷字符串里——LLM 看不到
+    # 缺失维度就不会编造它
+    assert "occasion_preference" not in material
+
+
+def test_build_prompt_material_rounds_confidence() -> None:
+    diagnosis = _diagnosis(_dim("body_shape", "X型", confidence=0.876543))
+
+    payload = json.loads(build_prompt_material(diagnosis))
+
+    assert payload["dimensions"][0]["confidence"] == 0.88
+
+
+def test_build_prompt_material_keeps_key_value_evidence_and_rule_hits() -> None:
+    hits = [
+        {
+            "rule_key": "R-BODY-0007",
+            "dimension": "body_shape",
+            "suggestion": "优先选择收腰版型",
+            "explanation": "X 型体型肩腰臀比例均衡",
+            "is_hard": False,
+            "internal_field": "不该进 prompt 的字段",
+        }
+    ]
+    diagnosis = _diagnosis(
+        _dim("body_shape", "X型", evidence=("ratio:1.02", "ratio:0.71")), rule_hits=hits
+    )
+
+    material = build_prompt_material(diagnosis)
+    payload = json.loads(material)
+
+    assert payload["dimensions"][0] == {
+        "key": "body_shape",
+        "value": "X型",
+        "evidence": ["ratio:1.02", "ratio:0.71"],
+        "confidence": 0.82,
+    }
+    assert payload["rule_hits"] == [
+        {
+            "rule_key": "R-BODY-0007",
+            "dimension": "body_shape",
+            "suggestion": "优先选择收腰版型",
+            "explanation": "X 型体型肩腰臀比例均衡",
+            "is_hard": False,
+        }
+    ]
+    assert "internal_field" not in material
+
+
+# ---------------------------------------------------------------------------
+# 素材组装：超长裁剪
+# ---------------------------------------------------------------------------
+
+
+def test_overlong_material_drops_low_priority_dimensions_and_records_keys() -> None:
+    big_evidence = ("x" * 200,) * 3  # 每个维度约 600+ 字符
+    diagnosis = _diagnosis(
+        _dim("body_shape", "X型", evidence=big_evidence),  # 硬结论，优先级 100
+        _dim("style_impression", evidence=big_evidence),  # 优先级 60
+        _dim("occasion_preference", evidence=big_evidence),  # 未登记 → 默认优先级 50
+    )
+
+    material, dropped = compose_prompt_material(diagnosis, max_chars=1000)
+
+    # 默认优先级先丢，其次 style_impression；硬结论保留
+    assert dropped == ["occasion_preference", "style_impression"]
+    payload = json.loads(material)
+    assert [d["key"] for d in payload["dimensions"]] == ["body_shape"]
+    assert len(material) <= 1000
+
+
+def test_hard_dimensions_are_never_dropped_even_when_over_budget() -> None:
+    huge_evidence = ("x" * 400,) * 5
+    diagnosis = _diagnosis(
+        _dim("body_shape", "X型", evidence=huge_evidence),
+        _dim("face_shape", "鹅蛋", evidence=huge_evidence),
+        _dim("color_season", "浅暖柔和型", evidence=huge_evidence),
+    )
+
+    material, dropped = compose_prompt_material(diagnosis, max_chars=200)
+
+    assert dropped == []
+    payload = json.loads(material)
+    assert [d["key"] for d in payload["dimensions"]] == ["body_shape", "face_shape", "color_season"]
+
+
+# ---------------------------------------------------------------------------
+# 降级模板
+# ---------------------------------------------------------------------------
+
+
+def test_render_fallback_copy_includes_values_evidence_and_rule_explanations() -> None:
+    dimensions = [
+        _dim(
+            "body_shape", "X型", evidence=("shoulder_hip_ratio:1.02",), rule_hits=("R-BODY-0007",)
+        ),
+        _dim("color_season", "浅暖柔和型", evidence=("contrast:low",), rule_hits=("R-COLOR-0021",)),
+    ]
+    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",
+        }
+    ]
+
+    fallback = render_fallback_copy(dimensions, rule_hits)
+
+    titles = [s.title for s in fallback.sections]
+    assert titles == ["诊断结论", "搭配建议"]
+    conclusion = "\n".join(fallback.sections[0].paragraphs)
+    assert "「X型」" in conclusion and "shoulder_hip_ratio:1.02" in conclusion
+    assert "「浅暖柔和型」" in conclusion
+    advice = "\n".join(fallback.sections[1].paragraphs)
+    assert "优先选择收腰版型突出腰线" in advice
+    # 有 rule_hits 时必须含 explanation——解释链是模板文案的最低质量线
+    assert "收腰设计能强化这一优势" in advice
+
+
+def test_render_fallback_copy_without_hits_has_no_advice_section() -> None:
+    dimensions = [_dim("body_shape", "X型")]
+    fallback = render_fallback_copy(dimensions, [])
+
+    assert [s.title for s in fallback.sections] == ["诊断结论"]
+
+
+def test_render_fallback_copy_with_no_usable_input_explains_itself() -> None:
+    fallback = render_fallback_copy([_dim("occasion_preference", status="skipped", value="")], [])
+
+    assert [s.title for s in fallback.sections] == ["说明"]
+    assert fallback.sections[0].paragraphs
+
+
+def test_fallback_copy_passes_the_full_validation_suite() -> None:
+    """降级模板必须天然通过三重校验——降级路径的黄金性质在纯函数级先锁一遍。"""
+    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",
+        }
+    ]
+
+    fallback = render_fallback_copy(dimensions, rule_hits)
+
+    assert validate_copy(fallback, dimensions, rule_hits=rule_hits) == []

tests/unit/test_report_validation.py

@@ -0,0 +1,211 @@
+"""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()) == []