test(harness): 黄金测试加非空护栏、补类型注解、端到端对照旧逻辑

回应 Codex 审查三项测试质量问题：
- C1: 每个 byte-identical 断言前加非空护栏（sum(len)>0 / len==samples>0），
  防空==空误通过。
- I1: TestBatchingByteIdentical._assert 补齐完整类型注解。
- I2: 端到端 checkpoint 用例增加 _reference_build_batches 对照，断言真实池划分
  →batching 与旧逐题逻辑逐字节一致（不只是 rebuilt==batches 自往返）。

tests/unit/test_non_ar_byte_identical.py

@@ -168,9 +168,19 @@ class TestPureSingleAreSizeOneUnits:
 class TestBatchingByteIdentical:
     """纯 single build_batches 与旧逐题算法逐字节一致。"""
 
-    def _assert(self, items, correctness, batch_size, min_cls, seed, ratio) -> None:
+    def _assert(
+        self,
+        items: list[GeneratedQuestion],
+        correctness: dict[str, bool],
+        batch_size: int,
+        min_cls: int,
+        seed: int,
+        ratio: float,
+    ) -> None:
         got, _ = build_batches(items, correctness, batch_size, min_cls, seed, ratio)
         ref = _reference_build_batches(items, correctness, batch_size, min_cls, seed, ratio)
+        # 非空护栏：确保对照是实质性非空比较（防空==空误通过）
+        assert sum(len(b) for b in ref) > 0
         assert _ids(got) == _ids(ref)
 
     def test_pure_errors(self) -> None:
@@ -249,6 +259,14 @@ class TestEndToEndByteIdentical:
         )
         batches, _ = build_batches(p.diagnosis, p.correctness, 6, 2, seed=3, correct_ratio=0.5)
 
+        # 端到端对照旧逻辑：真实池划分 → batching 与引入 QuestionUnit 前逐字节一致
+        reference = _reference_build_batches(
+            p.diagnosis, p.correctness, 6, 2, seed=3, correct_ratio=0.5
+        )
+        # 非空护栏：确保对照是实质性非空比较（防空==空误通过）
+        assert sum(len(b) for b in reference) > 0
+        assert _ids(batches) == _ids(reference)
+
         # checkpoint 折叠为 unit_id（纯 single 即 question_id）后恢复
         epoch_batches = [_batch_unit_ids(b) for b in batches]
         assert epoch_batches == _ids(batches)  # 纯非 AR：unit_id 序列 == question_id 序列
@@ -274,6 +292,7 @@ class TestMomentumPureNonARByteIdentical:
 
         candidates = [q for q in pool if q.task_type in allowed]
         ref = random.Random(epoch).sample(candidates, samples)
+        assert len(ref) == samples > 0  # 非空护栏：实质性抽样
         assert [q.question_id for q in got] == [q.question_id for q in ref]
 
     def test_ar_pairs_of_other_type_do_not_shift(self) -> None:
@@ -289,6 +308,7 @@ class TestMomentumPureNonARByteIdentical:
             po, pm = _pair(f"p{k}", task_type="AR")
             mixed.extend([po, pm])
         after = _sample_momentum_candidates(mixed, allowed, samples, epoch)
+        assert len(base) == samples > 0  # 非空护栏：实质性抽样
         assert [q.question_id for q in base] == [q.question_id for q in after]
 
     def test_fewer_candidates_than_samples_returns_all(self) -> None: