fix: harden mirror material rebuild (empty/invalid frames, dedup, observability)

app/question_gen/adversarial_filter.py

@@ -244,10 +244,12 @@ def _rebuild_material(tree: TreeIndex, source_nodes: tuple[str, ...]) -> Materia
         _collect_subtitle_sentences,
     )
 
-    subtitles = _collect_subtitle_sentences(tree, source_nodes)
+    # 保序去重：source_nodes 若同时含父子节点，两者子树重叠会重复采集，去冗余再喂 VLM。
+    subtitles = list(dict.fromkeys(_collect_subtitle_sentences(tree, source_nodes)))
     frames: list[str] = []
     for nid in source_nodes:
         frames.extend(_collect_frame_paths(tree, nid))
+    frames = list(dict.fromkeys(frames))
     return _MaterialContext(
         anchor=None,  # 镜像 prompt 不用 anchor
         source_nodes=source_nodes,
@@ -286,6 +288,34 @@ def _parse_mirror(raw: str) -> dict | None:
     return mirror if isinstance(mirror, dict) else None
 
 
+def _existing_frames(frame_paths: list[str]) -> list[str]:
+    """过滤掉磁盘上不存在的帧路径，保序返回有效帧列表（防失效帧喂给 VLM）。"""
+    return [p for p in frame_paths if Path(p).exists()]
+
+
+def _build_mirror_question(
+    question: GeneratedQuestion, mirror: dict
+) -> GeneratedQuestion | None:
+    """从解析出的 mirror dict 构造镜像题；字段缺失/类型错误 → None。"""
+    try:
+        options = tuple(str(o) for o in mirror["options"])
+        answer = str(mirror["answer"]).strip().upper()
+        m_question = str(mirror["question"])
+    except (KeyError, TypeError):
+        return None
+    return GeneratedQuestion(
+        question_id=f"{question.question_id}_mirror",
+        video_id=question.video_id,
+        task_type=question.task_type,
+        question=m_question,
+        options=options,
+        answer=answer,
+        source_nodes=question.source_nodes,
+        difficulty=question.difficulty,
+        sub_pattern=question.sub_pattern,
+    )
+
+
 async def generate_mirror_question(
     question: GeneratedQuestion,
     *,
@@ -305,8 +335,15 @@ async def generate_mirror_question(
 
     返回:
         镜像 GeneratedQuestion（question_id 加 "_mirror" 后缀，不进题库）；
-        无法造出有效对（null / 正解相同 / 解析失败）返回 None。
+        素材为空 / null / 正解相同 / 解析失败均返回 None（上层记 flip_skipped）。
     """
+    # 素材防御：无字幕 AND 无有效帧 → 不调 VLM，直接 flip_skipped（避免空素材空转 VLM）。
+    valid_frames = _existing_frames(list(material.frame_paths))
+    if not material.subtitle_sentences and not valid_frames:
+        logger.warning(
+            "镜像素材为空（source_nodes 可能缺失或帧失效），跳过: {}", question.question_id
+        )
+        return None
     system = (_PROMPTS_DIR / "ar_mirror_question.md").read_text(encoding="utf-8")
     subs = "\n".join(f"  - {s}" for s in material.subtitle_sentences)
     user = (
@@ -317,32 +354,20 @@ async def generate_mirror_question(
         f"## Subtitles\n{subs}\n"
     )
     messages = [{"role": "system", "content": system}, {"role": "user", "content": user}]
-    resp = await vlm.chat_with_images(
-        messages, list(material.frame_paths), session_id=session_id
-    )
+    resp = await vlm.chat_with_images(messages, valid_frames, session_id=session_id)
     mirror = _parse_mirror(resp.content)
     if mirror is None:
         return None
-    try:
-        options = tuple(str(o) for o in mirror["options"])
-        answer = str(mirror["answer"]).strip().upper()
-        m_question = str(mirror["question"])
-    except (KeyError, TypeError):
+    mirror_q = _build_mirror_question(question, mirror)
+    if mirror_q is None:
         return None
-    mirror_q = GeneratedQuestion(
-        question_id=f"{question.question_id}_mirror",
-        video_id=question.video_id,
-        task_type=question.task_type,
-        question=m_question,
-        options=options,
-        answer=answer,
-        source_nodes=question.source_nodes,
-        difficulty=question.difficulty,
-        sub_pattern=question.sub_pattern,
-    )
     # 镜像正解字面校验：canonical(P) 必须 != canonical(Q)（按选项文本比较，非字母）
     p_text = canonical_answer_text(question.options, question.answer)
-    q_text = canonical_answer_text(mirror_q.options, answer)
+    q_text = canonical_answer_text(mirror_q.options, mirror_q.answer)
     if p_text is None or q_text is None or p_text.strip() == q_text.strip():
+        logger.info(
+            "镜像正解与原题正解无区分（或无效），跳过翻转门: question_id={}",
+            question.question_id,
+        )
         return None
     return mirror_q