feat(report): 形象深度诊断全链路——POST /reports/{id}/full 派生 tier=full 报告，复用 basic 输入与 32 维结果；full 流水线（无视觉阶段）LLM 扩写四大块（十二季定位/规则色板/4场景穿搭/30天计划），三重校验+语义重试+规则兜底降级；删除终身一次硬校验（40920 下线，重复生成由配额兜底）；validation 枚举一致性校验放宽（十二季词表）；迁移 0003 SSOT 补 full_report 规则行

packages/report/validation.py

@@ -193,14 +193,15 @@ def check_absolute_claims(text: str) -> list[str]:
     return [term for term in ABSOLUTE_CLAIM_TERMS if term in text]
 
 
-def _find_dimension(
+def _find_dimensions(
     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
+) -> list[Mapping[str, Any]]:
+    """按 key 识别串收齐全部匹配维度（季型有 season_code 与 season_detailed 两个）。"""
+    return [
+        dim
+        for dim in dimensions
+        if any(token in str(dim.get("key") or "") for token in key_tokens)
+    ]
 
 
 def _check_enum_consistency(
@@ -211,34 +212,48 @@ def _check_enum_consistency(
     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。
+
+    两类放行：诊断值的「X型」写法（如「柔夏」→「柔夏型」，深度诊断的
+    十二季文案必经此路径）；词表 token 是某个放行词的子串时不再单独追责
+    （「夏型」⊂「柔夏型」——十二季细分命名无法全量枚举，见 TD-18）。
     """
     issues: list[str] = []
     for category in _ENUM_CATEGORIES:
-        dim = _find_dimension(dimensions, category.key_tokens)
-        if dim is None:
+        matched = _find_dimensions(dimensions, category.key_tokens)
+        if not matched:
             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()
+        allowed: set[str] = set()
+        for dim in matched:
+            value = dim.get("value")
+            if dim.get("status") == "ok" and isinstance(value, str) and value:
+                canonical = value
+                candidates = {canonical, f"{canonical}型"}
+                candidates.update(category.aliases.get(canonical, ()))  # 值本身可能就是别名键
+                # 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
+                    ):
+                        candidates.add(canonical_key)
+                        candidates.update(alias_values)
+                allowed.update(_normalize_for_match(c) for c in candidates)
         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
+            if normalized_token not in copy_text_normalized:
+                continue
+            if normalized_token in allowed:
+                continue
+            if any(normalized_token in entry for entry in allowed):
+                continue
+            observed = next(
+                (str(d.get("value")) for d in matched if d.get("status") == "ok"),
+                "未诊断",
+            )
+            issues.append(f"{category.label}用词「{token}」与诊断值「{observed}」不一致")
+            break
     return issues