diff --git a/app/application/classification/execution.py b/app/application/classification/execution.py
index 41dc5e672..3b6c42ae3 100644
--- a/app/application/classification/execution.py
+++ b/app/application/classification/execution.py
@@ -314,9 +314,12 @@ class ClassificationExecutionService:
         *,
         effective_override: ClassificationSelection | None,
     ) -> ClassificationSubject:
-        """应用纯求值结果和人工覆盖，并更新兼容目录分类。"""
+        """应用纯求值结果、降级保护和人工覆盖，并更新兼容目录分类。"""
         evaluation = evaluate_classification_facts(policy, facts)
         result = evaluation.result.model_copy(deep=True)
+        carried = _degraded_automatic_selection(finalized, policy, result)
+        if carried is not None:
+            result.effective = carried
         if effective_override:
             result.effective = effective_override.model_copy(deep=True)
         finalized.classification = result
@@ -537,6 +540,38 @@ def _explicit_effective_override(
     return deepcopy(effective)
 
 
+def _degraded_automatic_selection(
+    media: ClassificationSubject,
+    policy: ClassificationPolicy,
+    result: ClassificationResult,
+) -> ClassificationSelection | None:
+    """事实缺失把同策略结果打成兜底时，返回可复用的既有自动分类选择。
+
+    补充元数据的入口会先裁剪媒体字段再重新分类，此时缺失事实只会让结果更差，
+    不会带来新的判断依据。因此仅在既有快照来自同一策略修订且命中过规则，而本次
+    求值同时出现事实缺失和兜底化时才复用；事实读齐后给出的兜底结果仍然生效，
+    策略调整和真正的重新分类不会被掩盖。
+    """
+    if result.state != "partial":
+        return None
+    proposed = result.effective or result.recommended
+    if proposed is not None and proposed.source != "fallback":
+        return None
+    carried = getattr(media, "classification", None)
+    if not isinstance(carried, ClassificationResult):
+        return None
+    if carried.policy_revision != policy.revision:
+        return None
+    if carried.state not in {"complete", "partial"}:
+        return None
+    selection = carried.effective or carried.recommended
+    if selection is None or selection.source != "automatic":
+        return None
+    if not str(selection.category_id or "").strip():
+        return None
+    return selection.model_copy(deep=True)
+
+
 def _enum_text(value: object) -> str:
     """把字符串枚举或普通值转换为去空白文本。"""
     if isinstance(value, Enum):
diff --git a/tests/test_media_classification_execution.py b/tests/test_media_classification_execution.py
index 1bbc517e2..a53fa31b3 100644
--- a/tests/test_media_classification_execution.py
+++ b/tests/test_media_classification_execution.py
@@ -334,6 +334,69 @@ def test_execution_refreshes_same_revision_after_auxiliary_facts_change() -> Non
     assert refreshed.classification.effective.rule_id == "rule.movie.jp"
 
 
+def test_execution_preserves_automatic_category_when_supplement_drops_facts() -> None:
+    """裁剪过字段的媒体再次分类时，同策略已命中的自动分类不得被兜底覆盖。"""
+    migration = migrate_legacy_category_config(
+        {
+            "movie": {},
+            "tv": {
+                "国产剧": {"origin_country": "CN,TW,HK"},
+                "未分类": None,
+            },
+        }
+    )
+    domestic = next(
+        category.id
+        for category in migration.policy.categories
+        if category.name == "国产剧"
+    )
+    service = ClassificationExecutionService(_Runtime(migration.policy))
+    source = MediaInfo(
+        media_source=MediaSource.Douban,
+        media_id="35593344",
+        type=MediaType.TV,
+        title="测试剧",
+        production_countries=[{"name": "中国大陆"}],
+    )
+    classified = service.finalize(source)
+    assert classified.classification.effective.category_id == domestic
+    assert classified.classification.state == "complete"
+
+    classified.clear()
+    supplemented = service.finalize(classified)
+
+    assert supplemented.classification is not None
+    assert supplemented.classification.effective.category_id == domestic
+    assert supplemented.classification.effective.source == "automatic"
+    assert supplemented.library_category == "国产剧"
+    # 本次求值确实缺事实，推荐结果和状态仍如实记录，供预览解释差异。
+    assert supplemented.classification.state == "partial"
+    assert supplemented.classification.recommended.source == "fallback"
+
+
+def test_execution_keeps_fallback_when_complete_facts_no_longer_match() -> None:
+    """事实读齐后不再命中规则时，兜底结果必须生效而不是复用旧自动分类。"""
+    service = ClassificationExecutionService(_Runtime(_policy()))
+    source = MediaInfo(
+        media_source="douban",
+        media_id="1",
+        type=MediaType.MOVIE,
+        origin_country=["JP"],
+        genres=[{"name": "动画"}],
+    )
+    classified = service.finalize(source)
+    assert classified.classification.effective.category_id == "movie.jp"
+
+    classified.origin_country = ["US"]
+    reclassified = service.finalize(classified)
+
+    assert reclassified.classification is not None
+    assert reclassified.classification.state == "complete"
+    assert reclassified.classification.effective.category_id == "movie.other"
+    assert reclassified.classification.effective.source == "fallback"
+    assert reclassified.library_category == "其它电影"
+
+
 def test_execution_applies_manual_effective_override_without_losing_recommendation() -> None:
     """订阅或目录人工覆盖只替换 effective，仍保留自动推荐供 UI 解释。"""
     source = MediaInfo(