fix: Codex 全量审查修正

Critical:
- C1: assemble_mode 'plain' → 'ids'（合法枚举值）
- C2: question_id 加入 task_type slug 避免跨题型冲突

Important/Minor:
- generate_one 移除未用的 embed_fn/similarity_threshold 参数
- config.py 注释 11→12 同步
- 测试 question_id 断言更新

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

app/harness/config.py

@@ -267,7 +267,7 @@ def _validate_minibatch(config: RunConfig) -> None:
 
     关键实现细节:
         val_size 必须 >= eval_min_per_class * _VIDEO_MME_TASK_TYPE_COUNT，保证验证池
-        能为 Video-MME 的全部 11 个任务类型各保底 eval_min_per_class 题。
+        能为 Video-MME 的全部 12 个任务类型各保底 eval_min_per_class 题。
     """
     if config.batch_size <= 0:
         raise ValueError(f"batch_size 必须 > 0，实际: {config.batch_size}")

app/question_gen/synthesizer.py

@@ -624,7 +624,7 @@ def parse_vlm_response(
         raise ValueError(f"answer 必须是 A/B/C/D 之一，实际 '{answer}': {raw[:200]}")
 
     return {
-        "question_id": f"gen-{video_id}-{seq:03d}",
+        "question_id": f"gen-{video_id}-{task_type.lower().replace(' ', '_')}-{seq:03d}",
         "question": data["question"],
         "options": list(options),
         "answer": answer,
@@ -669,7 +669,6 @@ def is_duplicate(
 
 async def generate_one(
     vlm: VLMProvider,
-    embed_fn: Callable[[str | list[str]], np.ndarray],
     tree: TreeIndex,
     video_id: str,
     task_type: str,
@@ -678,7 +677,6 @@ async def generate_one(
     exemplars: list[GeneratedQuestion],
     used_node_ids: set[str],
     max_retries: int,
-    similarity_threshold: float,
     rng: random.Random,
     session_id: str,
 ) -> GeneratedQuestion | None:
@@ -695,7 +693,6 @@ async def generate_one(
 
     参数:
         vlm: VLM 调用端口。
-        embed_fn: 文本嵌入函数（本函数内未使用，由调用方统一去重）。
         tree: 三层树索引。
         video_id: 所属视频标识。
         task_type: 12 种 Video-MME 题型之一。
@@ -703,7 +700,6 @@ async def generate_one(
         exemplars: 少样本示例列表。
         used_node_ids: 已用节点 ID 集合。
         max_retries: 最大重试次数。
-        similarity_threshold: 余弦相似度阈值（本函数内未使用）。
         rng: 可控随机数生成器。
         session_id: 会话 ID（传递给 VLM 遥测）。
 

tests/unit/test_synthesizer.py

@@ -346,14 +346,14 @@ class TestParseVlmResponse:
         assert result["question"] == "What?"
         assert result["answer"] == "A"
         assert len(result["options"]) == 4
-        assert result["question_id"] == "gen-vid1-001"
+        assert result["question_id"] == "gen-vid1-object_recognition-001"
 
     def test_json_in_code_block(self) -> None:
         """从 markdown 代码块中提取 JSON。"""
         raw = '```json\n{"question": "Q?", "options": ["A. 1", "B. 2", "C. 3", "D. 4"], "answer": "B"}\n```'
         result = parse_vlm_response(raw, "vid1", "Object Recognition", 2)
         assert result["question"] == "Q?"
-        assert result["question_id"] == "gen-vid1-002"
+        assert result["question_id"] == "gen-vid1-object_recognition-002"
 
     def test_invalid_json_raises(self) -> None:
         """非 JSON 文本应抛出 ValueError。"""
@@ -382,7 +382,7 @@ class TestParseVlmResponse:
         """seq 应按 3 位零填充格式化到 question_id 中。"""
         raw = '{"question": "Q?", "options": ["A. 1", "B. 2", "C. 3", "D. 4"], "answer": "C"}'
         result = parse_vlm_response(raw, "video_abc", "Action Reasoning", 42)
-        assert result["question_id"] == "gen-video_abc-042"
+        assert result["question_id"] == "gen-video_abc-action_reasoning-042"
 
 
 # ---------------------------------------------------------------------------
@@ -448,14 +448,9 @@ class TestGenerateOne:
             content='{"question":"Q?","options":["A. 1","B. 2","C. 3","D. 4"],"answer":"A"}',
         )
 
-        def embed_fn(t: str | list[str]) -> np.ndarray:
-            shape = (1, 4) if isinstance(t, str) else (len(t), 4)
-            return np.zeros(shape, dtype=np.float32)
-
         tree, vid = self._load_test_tree()
         result = await generate_one(
             vlm=vlm,
-            embed_fn=embed_fn,
             tree=tree,
             video_id=vid,
             task_type="Object Recognition",
@@ -463,12 +458,11 @@ class TestGenerateOne:
             exemplars=[],
             used_node_ids=set(),
             max_retries=3,
-            similarity_threshold=0.85,
             rng=random.Random(42),
             session_id="test",
         )
         assert result is not None
-        assert result.question_id == f"gen-{vid}-001"
+        assert result.question_id == f"gen-{vid}-object_recognition-001"
         assert result.task_type == "Object Recognition"
         assert result.source_nodes  # non-empty
         assert result.difficulty == "medium"
@@ -479,13 +473,9 @@ class TestGenerateOne:
         vlm = AsyncMock()
         vlm.chat_with_images.return_value = MagicMock(content="invalid")
 
-        def embed_fn(t: str | list[str]) -> np.ndarray:
-            return np.zeros((1, 4), dtype=np.float32)
-
         tree, vid = self._load_test_tree()
         result = await generate_one(
             vlm=vlm,
-            embed_fn=embed_fn,
             tree=tree,
             video_id=vid,
             task_type="Object Recognition",
@@ -493,7 +483,6 @@ class TestGenerateOne:
             exemplars=[],
             used_node_ids=set(),
             max_retries=2,
-            similarity_threshold=0.85,
             rng=random.Random(42),
             session_id="test",
         )

tools/generate_questions.py

@@ -594,7 +594,7 @@ async def _run_inference_for_questions(
                 ocr=None,
                 verify_vision=False,
                 anchor=False,
-                assemble_mode="plain",
+                assemble_mode="ids",
             )
             result = await run_inference(
                 group,
@@ -861,7 +861,6 @@ async def _run_generate(args: argparse.Namespace) -> None:
 
     async def _generate_with_sem(
         vlm_client,
-        embed_fn_inner,
         tree,
         video_id,
         task_type,
@@ -870,7 +869,6 @@ async def _run_generate(args: argparse.Namespace) -> None:
         exemplars,
         used_node_ids,
         max_retries_inner,
-        similarity_threshold_inner,
         rng_inner,
         session_id,
     ):
@@ -878,7 +876,6 @@ async def _run_generate(args: argparse.Namespace) -> None:
         async with sem:
             return await generate_one(
                 vlm_client,
-                embed_fn_inner,
                 tree,
                 video_id,
                 task_type,
@@ -886,7 +883,6 @@ async def _run_generate(args: argparse.Namespace) -> None:
                 exemplars=exemplars,
                 used_node_ids=used_node_ids,
                 max_retries=max_retries_inner,
-                similarity_threshold=similarity_threshold_inner,
                 rng=rng_inner,
                 session_id=session_id,
             )
@@ -929,7 +925,6 @@ async def _run_generate(args: argparse.Namespace) -> None:
             for _attempt in range(max_retries):
                 candidate = await _generate_with_sem(
                     vlm,
-                    embed_fn,
                     tree,
                     video_id,
                     task_type,
@@ -937,7 +932,6 @@ async def _run_generate(args: argparse.Namespace) -> None:
                     exemplars=exemplars,
                     used_node_ids=used_node_ids,
                     max_retries_inner=1,
-                    similarity_threshold_inner=similarity_threshold,
                     rng_inner=rng,
                     session_id=session_id,
                 )