fix: load real tree_data for training-loop diagnosis (runner)

算法保真 §4.7#12（训练循环编排）：仅将 _run_diagnosis 内 tree_data={} 替换为
按 batch question_ids 涉及 video 加载真实树注入 run_diagnosis，未改动三级嵌套/
慢更新/断点续训/gate 等编排逻辑。覆盖 question_ids=None 全量诊断与
question_ids=[...] batch 两条路径。

app/harness/runner.py

@@ -2174,10 +2174,19 @@ class Runner:
         skill_store = VersionedSkillStore(self._paths.skills_dir)
         diagnose_prompts = self._load_diagnose_prompts()
 
+        from app.harness.tree_nodes import load_tree_data_for_videos
+
+        if question_ids is not None:
+            qid_set = set(question_ids)
+            video_ids = [q.video_id for q in questions if q.question_id in qid_set]
+        else:
+            video_ids = [q.video_id for q in questions]
+        tree_data = load_tree_data_for_videos(Path(self._config.store_dir), video_ids)
+
         return await run_diagnosis(
             run_id=run_id,
             questions=questions,
-            tree_data={},  # tree_data 由诊断管线内部按需加载
+            tree_data=tree_data,
             llm=self._llm,
             run_log=run_log,
             skill_store=skill_store,

tests/unit/test_runner_diag_tree_inject.py

@@ -0,0 +1,179 @@
+"""训练循环诊断注入：_run_diagnosis 按 batch question_ids 加载真实树注入 run_diagnosis。
+
+覆盖诊断 tree_data 断链修复的训练循环侧（Task 4）：验证 runner._run_diagnosis 会把
+batch 涉及 video 的真实树作为 tree_data 传给 core.run_diagnosis，而非空 dict。
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path  # noqa: TC003 — 运行时 tmp_path 标注使用
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from app.harness.config import RunConfig
+from app.harness.runner import Runner
+from core.evolution.types import DiagnosisResult
+
+# 真实样本：question 604-2 属于 video 0RxMZBLeqRI（111 节点树）。
+_REAL_QUESTION_ID = "604-2"
+_REAL_VIDEO_ID = "0RxMZBLeqRI"
+
+
+def _empty_diagnosis_result(run_id: str = "infer_adhoc") -> DiagnosisResult:
+    """构造仅含 run_id 的空诊断结果（其余字段走 dataclass 默认值）。"""
+    return DiagnosisResult(run_id=run_id)
+
+
+def _base_config(workspace_dir: Path, store_dir: Path) -> RunConfig:
+    """构造 diagnose 模式 RunConfig，所有必填字段给测试默认值。
+
+    参数:
+        workspace_dir: 已写入 manifest 的 workspace 根目录。
+        store_dir: 真实 store 根目录（含 questions/ 与 videos/）。
+    """
+    return RunConfig(
+        workspace_dir=workspace_dir,
+        store_dir=store_dir,
+        mode="diagnose",
+        concurrency=1,
+        max_steps=5,
+        skill_mode="none",
+        n_samples=0,
+        questions="benchmarks/Video-MME",
+        skills_version="v1",
+        prompts_version="v1",
+        epochs=1,
+        diag_size=10,
+        diag_correct_ratio=0.5,
+        val_size=24,
+        val_correct_ratio=0.5,
+        edit_budget_start=5,
+        edit_budget_end=2,
+        batch_size=5,
+        min_class_per_batch=2,
+        eval_min_per_class=2,
+        early_stop_patience=3,
+        test_size=10,
+        use_slow_momentum=False,
+        gate_e_confirm=20.0,
+        gate_e_provisional=3.0,
+        gate_w_net_min=2,
+        gate_delta_min=0.02,
+        gate_lambda_dir=-0.642,
+        gate_e_rollback=10.0,
+        gate_block=8,
+        gate_n_max=40,
+        gate_p_low=0.05,
+        gate_p_high=0.95,
+        gate_probe_quota=0.2,
+        gate_gamma_decay=0.9,
+        gate_cooldown_steps=2,
+        gate_guard_err=0.10,
+        skill_update_mode="patch",
+        appendix_consolidate_threshold=6,
+        run_id="infer_adhoc",
+    )
+
+
+@pytest.fixture
+def runner_with_real_store(tmp_path: Path) -> Runner:
+    """构造 diagnose 模式 runner，questions/videos 指向真实 store。
+
+    manifest.store 写真实 store 绝对路径，使 resolve_paths 的 questions_dir
+    命中含 604-2 的题库、store_dir 命中 0RxMZBLeqRI 的真实 tree.json。
+    """
+    store_dir = Path(__file__).resolve().parents[2] / "store"
+    ws = tmp_path / "ws"
+    ws.mkdir()
+    (ws / "skills" / "v1").mkdir(parents=True)
+    manifest = {
+        "name": "ws",
+        "created_at": "",
+        "store": store_dir.as_posix(),
+        "current": {
+            "videos": "videos",
+            "questions": "questions/benchmarks/Video-MME",
+            "skills": "skills/v1",
+            "prompts": "prompts/v1",
+        },
+        "history": [],
+    }
+    (ws / "manifest.json").write_text(json.dumps(manifest))
+
+    return Runner(
+        _base_config(ws, store_dir),
+        llm=MagicMock(),
+        evolve_llm=MagicMock(),
+        vlm=MagicMock(),
+        telemetry=MagicMock(),
+    )
+
+
+@pytest.mark.asyncio
+async def test_run_diagnosis_injects_tree_data(runner_with_real_store: Runner) -> None:
+    """_run_diagnosis 把 batch question_ids 对应 video 的真实树注入 run_diagnosis。"""
+    captured: dict[str, object] = {}
+
+    async def _fake_run_diagnosis(**kwargs: object) -> DiagnosisResult:
+        captured["tree_data"] = kwargs["tree_data"]
+        return _empty_diagnosis_result()
+
+    with patch(
+        "core.evolution.diagnose.run_diagnosis",
+        new=AsyncMock(side_effect=_fake_run_diagnosis),
+    ):
+        await runner_with_real_store._run_diagnosis("infer_adhoc", question_ids=[_REAL_QUESTION_ID])
+
+    tree_data = captured["tree_data"]
+    assert _REAL_VIDEO_ID in tree_data
+    assert tree_data[_REAL_VIDEO_ID]["nodes"]
+
+
+def _fake_question(question_id: str, video_id: str) -> object:
+    """构造仅设置 question_id/video_id 的 GeneratedQuestion（其余字段给占位默认）。"""
+    from core.types import GeneratedQuestion
+
+    return GeneratedQuestion(
+        question_id=question_id,
+        video_id=video_id,
+        task_type="Action Reasoning",
+        question="问题",
+        options=("A", "B", "C", "D"),
+        answer="A",
+        source_nodes=(),
+        difficulty="medium",
+    )
+
+
+@pytest.mark.asyncio
+async def test_run_diagnosis_full_scan_loads_all_video_trees(
+    runner_with_real_store: Runner,
+) -> None:
+    """question_ids=None（全量诊断）路径：加载全部 questions 涉及 video 的树。
+
+    轻量验证：patch load_benchmark 返回两个不同 video 的假题，patch
+    load_tree_data_for_videos 捕获 video_ids，断言全量路径覆盖全部 video。
+    """
+    fake_qs = [
+        _fake_question("q-a", "vA"),
+        _fake_question("q-b", "vB"),
+    ]
+    captured: dict[str, object] = {}
+
+    def _cap(store_dir: Path, video_ids: list[str]) -> dict[str, object]:
+        captured["video_ids"] = list(video_ids)
+        return {v: {"nodes": {}} for v in video_ids}
+
+    with (
+        patch("app.question_gen.load_benchmark", return_value=fake_qs),
+        patch("app.harness.tree_nodes.load_tree_data_for_videos", side_effect=_cap),
+        patch(
+            "core.evolution.diagnose.run_diagnosis",
+            new=AsyncMock(return_value=_empty_diagnosis_result()),
+        ),
+    ):
+        await runner_with_real_store._run_diagnosis("infer_adhoc", question_ids=None)
+
+    assert set(captured["video_ids"]) == {"vA", "vB"}