fix: route None/degraded diagnoses to lapse; abort on high degrade rate

tests/unit/test_runner_diag_tree_inject.py

@@ -212,6 +212,32 @@ async def test_diagnosis_reads_traces_from_steps_json(
     assert traces[0]["tool_name"] == "search_tree"
 
 
+@pytest.mark.asyncio
+async def test_run_step_aborts_on_high_degrade_rate(
+    runner_with_real_store: Runner, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """诊断降级占比 > 50% 时 _run_step 中止（疑似 judge 基础设施故障）。"""
+    from unittest.mock import AsyncMock
+
+    batch = [_fake_question("q1", "vA"), _fake_question("q2", "vA")]
+
+    runner_with_real_store._rollout_batch = AsyncMock()
+    monkeypatch.setattr("app.harness.runner._apply_batch_correctness", lambda *a, **k: None)
+    runner_with_real_store._run_diagnosis = AsyncMock(
+        return_value=DiagnosisResult(run_id="r", degraded_count=2)
+    )
+    runner_with_real_store._gate_batch_skills = AsyncMock()
+
+    state = MagicMock()
+    state.correctness = {"q1": False, "q2": False}
+    state.gate_cooldown = {}
+    pools = MagicMock()
+    pools.baseline_run_id = "infer_adhoc"
+
+    with pytest.raises(RuntimeError, match="降级占比"):
+        await runner_with_real_store._run_step(1, 0, 10, batch, pools, state)
+
+
 @pytest.mark.asyncio
 async def test_run_diagnosis_full_scan_loads_all_video_trees(
     runner_with_real_store: Runner,