feat: add greedy joint-constrained split selector

两阶段贪心视频级切分：Floor 硬约束优先满足 + 多样性 submodular 覆盖最大化，
两阶段均带 ε 守护保 test 代表性；不可行 fail loud（InfeasibleSplitError），
多样性欠额记 loguru warning 不静默。derive_reportable_types 落地长尾报告门限。
确定性：Random(seed) 预洗牌打破等增益/等槽数平局，同 config 同 videos 同解。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

tests/unit/test_split_selection.py

@@ -1,6 +1,98 @@
+import sqlite3
+from collections import Counter
+from itertools import cycle
+from pathlib import Path
+
 import pytest
 
-from app.harness.split_selection import cell_of, evolution_target_of, score_signal
+from app.harness.split_selection import (
+    build_video_records,
+    cell_of,
+    evolution_target_of,
+    score_signal,
+)
+
+# T2 信号构造用的错误类别循环源（4 类真实 error_type，制造多样格子）。
+_ERROR_TYPES = ["extraction_failure", "search_failure", "reasoning_failure", "mixed"]
+# 二次构造 T2 信号覆盖的高信号题型（floor / diversity 需有料）。
+_SIGNAL_TYPES = {"Counting Problem", "Action Reasoning"}
+_HARNESS_DB = Path(__file__).resolve().parents[2] / "workspaces" / "default" / "harness.db"
+
+
+def _load_adhoc_predictions() -> list[dict]:
+    """从真实 harness.db 的 infer_adhoc run 读取 900 条预测（按 question_id 去重）。
+
+    返回:
+        preds 列表，每行含 video_id / question_id / task_type / correct，
+        correct = (prediction 非空 且 prediction == answer)。
+    """
+    con = sqlite3.connect(str(_HARNESS_DB))
+    try:
+        cur = con.execute(
+            "SELECT question_id, video_id, task_type, prediction, answer "
+            "FROM predictions WHERE run_id = 'infer_adhoc'"
+        )
+        seen: set[str] = set()
+        preds: list[dict] = []
+        for question_id, video_id, task_type, prediction, answer in cur.fetchall():
+            if question_id in seen:
+                continue
+            seen.add(question_id)
+            preds.append(
+                {
+                    "video_id": video_id,
+                    "question_id": question_id,
+                    "task_type": task_type,
+                    "correct": prediction is not None and prediction == answer,
+                }
+            )
+        return preds
+    finally:
+        con.close()
+
+
+def _real_shaped_video_records() -> list:
+    """用真实预测结构 + 二次构造 T2 信号构建 300 个 VideoRecord。
+
+    题型 / 难度画像取自真实 infer_adhoc 预测；因真实诊断尚未跑，T2 信号是二次构造：
+    对高信号题型（Counting / Action Reasoning）的错题标 T2 defect，error_type 循环取 4 类，
+    使 floor 约束有料、多样性格子有区分度。
+
+    返回:
+        全部 300 个真实结构的 VideoRecord 列表。
+    """
+    preds = _load_adhoc_predictions()
+    err = cycle(_ERROR_TYPES)
+    signal_rows: list[dict] = []
+    for pred in preds:
+        if pred["task_type"] in _SIGNAL_TYPES and not pred["correct"]:
+            signal_rows.append(
+                {
+                    "question_id": pred["question_id"],
+                    "task_type": pred["task_type"],
+                    "error_type": next(err),
+                    "tier": "T2",
+                }
+            )
+    return build_video_records(preds, signal_rows)
+
+
+def _count_questions_by_type(videos: list) -> dict:
+    """统计各 task_type 的视频承载数（含该题型的视频数），供 derive_reportable_types。
+
+    VideoRecord 只留题型去重集 type_set，故这里按"含该题型的视频数"聚合，
+    作为长尾报告门限的代理度量。
+
+    参数:
+        videos: VideoRecord 列表。
+
+    返回:
+        {task_type: 含该题型的视频数}。
+    """
+    counter: Counter = Counter()
+    for video in videos:
+        counter.update(video.type_set)
+    return dict(counter)
 
 
 def test_evolution_target_mapping():
@@ -61,3 +153,58 @@ def test_build_video_records_covers_all_videos_with_difficulty_and_types():
     assert v1.cells == {("Counting Problem", "search_failure")}
     assert v1.type_set == {"Counting Problem", "Action Reasoning", "OCR Problems"}
     assert v1.wrong_by_type == {"Counting Problem": 1}  # T2 计数供 floor
+
+
+def test_select_split_video_disjoint_and_floor_and_deterministic():
+    from app.harness.split_selection import (
+        SelectConfig,
+        derive_reportable_types,
+        select_split,
+    )
+
+    videos = _real_shaped_video_records()
+    total_by_type = _count_questions_by_type(videos)
+    cfg = SelectConfig(
+        n_trainval=100,
+        floor_k={"Counting Problem": 3},
+        epsilon=0.1,
+        reportable_types=derive_reportable_types(total_by_type, report_floor=27),
+        seed=7,
+    )
+    a = select_split(videos, config=cfg)
+    b = select_split(videos, config=cfg)
+    assert set(a.trainval) & set(a.test) == set()  # 视频级不相交
+    assert set(a.trainval) | set(a.test) == {v.video_id for v in videos}  # 补集覆盖全集
+    assert a.trainval == b.trainval  # 同 seed 同 config → 同解
+    trainval_ids = set(a.trainval)
+    counting_defects = sum(
+        v.wrong_by_type.get("Counting Problem", 0) for v in videos if v.video_id in trainval_ids
+    )
+    assert counting_defects >= 3  # floor 硬约束满足
+
+
+def test_infeasible_floor_vs_epsilon_raises():
+    from app.harness.split_selection import (
+        InfeasibleSplitError,
+        SelectConfig,
+        select_split,
+    )
+
+    videos = _real_shaped_video_records()
+    with pytest.raises(InfeasibleSplitError):  # 极小 ε + 高 floor → 死锁
+        select_split(
+            videos,
+            config=SelectConfig(
+                n_trainval=2,
+                floor_k={"OCR Problems": 50},
+                epsilon=0.001,
+                reportable_types=set(),
+                seed=1,
+            ),
+        )
+
+
+def test_derive_reportable_types():
+    from app.harness.split_selection import derive_reportable_types
+
+    assert derive_reportable_types({"A": 30, "B": 10}, report_floor=27) == {"A"}