fix: fail loud when diagnosis video tree not covered (algo #7 input)

core/evolution/diagnose.py

@@ -2142,7 +2142,14 @@ async def run_diagnosis(
             key = (prediction.get("video_id", ""), prediction.get("question_id", ""))
             traces = traces_by_question.get(key, [])
             vid = prediction.get("video_id", "")
-            td = tree_data_by_video.get(vid, {})
+            if vid not in tree_data_by_video:
+                # P5 fail-loud：诊断需真实树，调用方须为每个诊断视频加载 tree_data；
+                # 静默回退空树会让 ground_truth 恒空、error_type 归因坍缩（本次修复的根因）。
+                raise ValueError(
+                    f"诊断视频树未覆盖: video_id={vid!r} 不在注入的 tree_data 中"
+                    "（调用方须为每个诊断视频加载树，P5 fail loud）"
+                )
+            td = tree_data_by_video[vid]
             skill_content = skill_cache.get(prediction.get("task_type", ""), "")
 
             try:

tests/integration/test_baseline_diagnosis.py

@@ -35,6 +35,36 @@ class _FakeRunLog:
         return []
 
 
+class _CoveragePredRunLog:
+    """RunLog 假实现：返回一条 correct=0、带 view_node 步的诊断 prediction。
+
+    用于验证 core fail-loud 护栏——该 prediction 的 video_id 故意不在注入的
+    tree_data 里，run_diagnosis 应 raise 而非静默回退空树。
+    """
+
+    def __init__(self, video_id: str, question_id: str) -> None:
+        self._video_id = video_id
+        self._question_id = question_id
+
+    async def get_predictions(self, run_id, *, question_ids=None):
+        return [
+            {
+                "video_id": self._video_id,
+                "question_id": self._question_id,
+                "task_type": "Counting Problem",
+                "prediction": "B",  # 与 answer 不同 → correct=0
+                "answer": "A",
+                "stop_reason": "answer_found",  # 非 INFRA，进 _process_question
+                "steps_json": [
+                    {"tool": "view_node", "args": {"node_id": "n0"}},
+                ],
+            }
+        ]
+
+    async def get_traces(self, run_id, *, question_ids=None):
+        return []
+
+
 class _FakeLLM: ...
 
 
@@ -96,6 +126,8 @@ def _deps(monkeypatch, calls):
         llm=_FakeLLM(),
         skill_store=_FakeSkillStore(),
         prompts=object(),
+        # 豁免 core 视频覆盖 fail-loud：本用例 monkeypatch 了 run_diagnosis 为
+        # 忽略 tree_data 的假实现，永不进入真实 _process_question 覆盖护栏，故留空。
         tree_data={},
         concurrency=2,
     )
@@ -221,6 +253,8 @@ async def test_infra_and_degraded_projection(tmp_path, monkeypatch):
         llm=_FakeLLM(),
         skill_store=_FakeSkillStore(),
         prompts=object(),
+        # 豁免 core 视频覆盖 fail-loud：本用例 monkeypatch 了 run_diagnosis 为
+        # 忽略 tree_data 的假实现，永不进入真实 _process_question 覆盖护栏，故留空。
         tree_data={},
         concurrency=2,
     )
@@ -308,6 +342,8 @@ async def test_degraded_overrides_attribution(tmp_path, monkeypatch):
         llm=_FakeLLM(),
         skill_store=_FakeSkillStore(),
         prompts=object(),
+        # 豁免 core 视频覆盖 fail-loud：本用例 monkeypatch 了 run_diagnosis 为
+        # 忽略 tree_data 的假实现，永不进入真实 _process_question 覆盖护栏，故留空。
         tree_data={},
         concurrency=2,
     )
@@ -352,3 +388,49 @@ async def test_degraded_overrides_attribution(tmp_path, monkeypatch):
     )
     assert md_path.exists()
     store.close()
+
+
+class _EmptySkillStore:
+    """SkillStore 假实现：无 skill 文件，_resolve_skill_file 回退空串。"""
+
+    def list_skill_files(self):
+        return []
+
+    def read_skill(self, filename):
+        return ""
+
+
+@pytest.mark.asyncio
+async def test_run_diagnosis_raises_when_video_tree_missing():
+    """诊断视频未被 tree_data 覆盖时 fail-loud（不静默回退、不走 judge 降级）。
+
+    构造一条 correct=0、带 view_node 步的 prediction，其 video_id="vMISS"
+    故意不在注入的 tree_data 里。core 护栏应在取 td 处 raise ValueError，
+    且该 raise 位于 compute_question_metrics 的 try 之前，不被 judge 降级吞掉。
+    """
+    from core.evolution.diagnose import run_diagnosis
+    from core.types import GeneratedQuestion
+
+    q = GeneratedQuestion(
+        question_id="vMISS-1",
+        video_id="vMISS",
+        task_type="Counting Problem",
+        question="",
+        options=("A", "B", "C", "D"),
+        answer="A",
+        source_nodes=(),
+        difficulty="easy",
+    )
+
+    with pytest.raises(ValueError, match="诊断视频树未覆盖"):
+        await run_diagnosis(
+            run_id="infer_adhoc",
+            questions=[q],
+            tree_data={"vOTHER": {"nodes": {}}},  # 故意不含 vMISS
+            llm=_FakeLLM(),
+            run_log=_CoveragePredRunLog("vMISS", "vMISS-1"),
+            skill_store=_EmptySkillStore(),
+            prompts=object(),
+            concurrency=1,
+            question_ids=["vMISS-1"],
+        )