fix: address whole-impl review (INFRA T0 rows, reproducible manifest, evolution_target report, dead config, canonical DRY)

C-1: persist_infra_t0_rows 补 INFRA/空预测错题的 T0 信号行（不进诊断故须单独落库），run_pipeline 加 Phase 0，dry-run 用假数据走通。
C-2: CLI 加 --generated-at，真实运行默认盖真实 UTC now，可显式固定以字节级复现 manifest。
I-1: coverage_report 增 evolution_target_distribution（T2 信号按 tool/skill/system 计数）。
I-2: 删除 PoolConfig 死字段 n_trainval/floor_k/epsilon/report_floor/val_wrong_min（grep 确认无消费者，视频级切分用独立 VideoSplitConfig/SplitBuildConfig/SelectConfig）。
I-3: 抽共享 load_canonical_predictions(db_path, run_id)，CLI 与 build_split 共用；消除 canonical 取行 + correct 判定重复。
M-1: build_split docstring 注明 val_wrong_min-agnostic 契约（McNemar 护栏由 CLI 冻结后执行，Task 11 契约）。

tests/unit/test_video_split_cli.py

@@ -102,10 +102,18 @@ def test_check_mcnemar_power_zero_threshold_skips():
     assert cli.check_mcnemar_power(pools, val_wrong_min=0) == 0
 
 
-def test_run_pipeline_orders_two_phases(monkeypatch, tmp_path):
-    """run_pipeline 先跑 Phase 1 诊断、后跑 Phase 2 build_split（按序）。"""
+def test_run_pipeline_orders_three_phases(monkeypatch, tmp_path):
+    """run_pipeline 先补 Phase 0 INFRA T0、再 Phase 1 诊断、后 Phase 2 build_split（按序）。"""
     calls: list[str] = []
 
+    class _SpyStore:
+        def __init__(self):
+            self.t0_rows: list = []
+
+        def upsert(self, row):
+            calls.append("t0_upsert")
+            self.t0_rows.append(row)
+
     async def fake_diag(**kwargs):
         calls.append("diagnosis")
         assert kwargs["diag_fingerprint"] == "fp"
@@ -120,24 +128,111 @@ def test_run_pipeline_orders_two_phases(monkeypatch, tmp_path):
     monkeypatch.setattr(cli, "run_baseline_diagnosis", fake_diag)
     monkeypatch.setattr(cli, "build_split", fake_build_split)
 
+    # 一条 INFRA 空预测错题 → Phase 0 应补一行 T0（在诊断/切分之前）。
+    canonical_preds = [
+        {
+            "question_id": "q_infra",
+            "video_id": "v9",
+            "task_type": "Counting Problem",
+            "prediction": "",
+            "answer": "A",
+            "stop_reason": "error",
+            "correct": False,
+        }
+    ]
+    store = _SpyStore()
+
     result = asyncio.run(
         cli.run_pipeline(
             config=_config(),
             fingerprint="fp",
             diagnosis_deps=object(),
-            signal_store=object(),
+            signal_store=store,
             wrong_ids=["q1"],
             questions={},
+            canonical_preds=canonical_preds,
             harness_db=tmp_path / "h.db",
             questions_dir=tmp_path,
             out_dir=tmp_path / "out",
             generated_at="2026-07-15T00:00:00Z",
         )
     )
-    assert calls == ["diagnosis", "build_split"]
+    assert calls == ["t0_upsert", "diagnosis", "build_split"]  # Phase 0 先于诊断与切分
+    assert len(store.t0_rows) == 1
+    assert store.t0_rows[0].tier == "T0" and store.t0_rows[0].infra is True
     assert result.pools.validation == []
 
 
+def test_select_diagnosable_wrong_ids_excludes_infra_and_correct():
+    """可诊断错题筛选：排除 INFRA / 空预测 / 正确题，保留非空非 INFRA 错题（保序）。"""
+    preds = [
+        {"question_id": "ok", "stop_reason": "finished", "prediction": "B", "correct": True},
+        {"question_id": "wrong", "stop_reason": "finished", "prediction": "C", "correct": False},
+        {"question_id": "infra", "stop_reason": "error", "prediction": "", "correct": False},
+        {"question_id": "parse", "stop_reason": "parse_error", "prediction": "x", "correct": False},
+        {"question_id": "empty", "stop_reason": "finished", "prediction": "", "correct": False},
+    ]
+    assert cli.select_diagnosable_wrong_ids(preds) == ["wrong"]
+
+
+def test_persist_infra_t0_rows_persists_only_infra_or_empty_wrong(tmp_path):
+    """INFRA / 空预测错题落 T0（infra=True，error_type/target=None）；正确 / 可诊断错题不落。"""
+    from adapters.baseline_diagnosis_store import SqliteDiagnosisSignalStore
+
+    preds = [
+        {
+            "question_id": "ok",
+            "video_id": "v1",
+            "task_type": "Counting Problem",
+            "prediction": "A",
+            "answer": "A",
+            "stop_reason": "finished",
+            "correct": True,
+        },
+        {
+            "question_id": "diag_wrong",
+            "video_id": "v2",
+            "task_type": "Counting Problem",
+            "prediction": "C",
+            "answer": "A",
+            "stop_reason": "finished",
+            "correct": False,
+        },
+        {
+            "question_id": "infra_err",
+            "video_id": "v3",
+            "task_type": "OCR Problems",
+            "prediction": "",
+            "answer": "A",
+            "stop_reason": "error",
+            "correct": False,
+        },
+        {
+            "question_id": "parse_err",
+            "video_id": "v4",
+            "task_type": "Counting Problem",
+            "prediction": "",
+            "answer": "A",
+            "stop_reason": "parse_error",
+            "correct": False,
+        },
+    ]
+    store = SqliteDiagnosisSignalStore(str(tmp_path / "h.db"))
+    n = cli.persist_infra_t0_rows(store, preds, "infer_adhoc", "fp")
+    assert n == 2  # 只有两条 INFRA 空预测错题
+    rows = {r.question_id: r for r in store.load("infer_adhoc", "fp")}
+    assert set(rows) == {"infra_err", "parse_err"}
+    for r in rows.values():
+        assert r.tier == "T0" and r.infra is True
+        assert r.error_type is None and r.evolution_target is None and r.cause_category is None
+        assert r.degraded is False
+
+    # 幂等：重复调用同 PK 覆盖，行数不变。
+    assert cli.persist_infra_t0_rows(store, preds, "infer_adhoc", "fp") == 2
+    assert len(store.load("infer_adhoc", "fp")) == 2
+    store.close()
+
+
 def test_dry_run_computes_fingerprint_without_llm(monkeypatch, tmp_path, capsys):
     """--dry-run：diag_fingerprint 被调用、Phase 1 走空错题早返回、不真调 LLM。"""
     fp_calls: list[tuple[str, str, str]] = []