fix: gate orchestrator materialize leak + cancel-drain on abort (algo #6)

Codex 质量审 2 Critical:
- C001: 候选物化移入 try、成功一个登记一个,第 N 个题型物化失败时
  finally 仍清理前 N-1 个已建目录,不泄漏
- C002: gather 首异常(护栏 raise)后显式取消其余任务并排水,确保
  finally 删除候选目录时无在飞任务访问、事件循环无 pending task 警告;
  护栏中止整轮语义不变
- 新增 2 测试:部分物化失败清理 / 护栏 raise 取消收束不悬挂(wait_for 5s)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

tests/unit/test_gate_concurrent.py

@@ -165,3 +165,95 @@ async def test_all_infra_raises(tmp_path, monkeypatch) -> None:
             log=log,
             concurrency=8,
         )
+
+
+@pytest.mark.asyncio
+async def test_partial_materialize_failure_cleans_up(tmp_path, monkeypatch) -> None:
+    """第 2 个题型物化失败:OSError 传播,且第 1 个已物化目录被清理不泄漏。"""
+    spec_a = _mk_spec("Action Reasoning", "action-reasoning", 1)
+    spec_b = _mk_spec("Counting Problem", "counting-problem", 1)
+    made: list[Path] = []
+
+    def _mat(workspace_dir, base_skills_version, target_file, content):
+        if made:  # 第 2 次调用:模拟磁盘错误
+            raise OSError("第 2 个题型物化失败(模拟)")
+        d = tmp_path / "cand_a"
+        d.mkdir()
+        made.append(d)
+        return d
+
+    monkeypatch.setattr("app.harness.validate.materialize_candidate_skill", _mat)
+
+    async def _never_called(questions, *, run_id, skills_dir):
+        raise AssertionError("物化失败后不应发起任何推理")
+
+    with pytest.raises(OSError):
+        await validate_skills_concurrent(
+            workspace_dir=tmp_path,
+            base_skills_version="v1",
+            specs=[spec_a, spec_b],
+            gate_params=_PARAMS,
+            gate_guard_err=0.10,
+            baseline_cache=BaselineCache(tmp_path / "bc.json"),
+            prompts_version="v1",
+            run_inference=_never_called,
+            log=_FakeLog(),
+            concurrency=8,
+        )
+    assert len(made) == 1
+    assert not made[0].exists()
+
+
+@pytest.mark.asyncio
+async def test_guard_raise_cancels_remaining_tasks(tmp_path, monkeypatch) -> None:
+    """护栏 raise 后其余在飞任务被取消收束:整体在超时内返回,不悬挂。
+
+    A 型 12 单元推理全 INFRA(stop_reason="error"),分母 ≥10 后错误率 1.0
+    超护栏 0.01 → RuntimeError;B 型推理挂在永不 set 的 Event 上,若无
+    取消收束,validate 将悬挂,wait_for 超时即为回归。
+    """
+    spec_a = _mk_spec("Action Reasoning", "action-reasoning", 12)
+    spec_b = _mk_spec("Counting Problem", "counting-problem", 2)
+    log = _FakeLog()
+    hang = asyncio.Event()  # 永不 set:B 型推理只能靠取消收束
+
+    class _R:
+        def __init__(self, run_id, total):
+            self.run_id, self.total = run_id, total
+
+    async def _run(questions, *, run_id, skills_dir):
+        if "counting-problem" in run_id:
+            await hang.wait()
+        for q in questions:
+            log.rows.append(
+                {
+                    "run_id": run_id,
+                    "question_id": q.question_id,
+                    "prediction": "",
+                    "answer": "A",
+                    "stop_reason": "error",
+                    "steps_json": "[]",
+                }
+            )
+        return _R(run_id, len(questions))
+
+    monkeypatch.setattr(
+        "app.harness.validate.materialize_candidate_skill",
+        lambda *a, **k: tmp_path / "cand",
+    )
+    with pytest.raises(RuntimeError):
+        await asyncio.wait_for(
+            validate_skills_concurrent(
+                workspace_dir=tmp_path,
+                base_skills_version="v1",
+                specs=[spec_a, spec_b],
+                gate_params=_PARAMS,
+                gate_guard_err=0.01,
+                baseline_cache=BaselineCache(tmp_path / "bc.json"),
+                prompts_version="v1",
+                run_inference=_run,
+                log=log,
+                concurrency=8,
+            ),
+            timeout=5,
+        )