fix: idempotent _run_step (DELETE stale) + checkpoint after gate save

tests/unit/test_runner_diag_tree_inject.py

@@ -212,6 +212,47 @@ async def test_diagnosis_reads_traces_from_steps_json(
     assert traces[0]["tool_name"] == "search_tree"
 
 
+@pytest.mark.asyncio
+async def test_run_step_deletes_stale_rows_before_rerun(
+    runner_with_real_store: Runner, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """同 run_id 重跑前先清 predictions/traces，避免重复行双计（幂等）。"""
+    from unittest.mock import AsyncMock
+
+    from app.harness.inference import PREDICTIONS_SCHEMA, TRACES_SCHEMA
+    from app.harness.log import HarnessLog
+
+    run_id = "infer_adhoc_e1_s0"
+    # 预置该 step run_id 的旧 predictions/traces 行
+    with HarnessLog(str(runner_with_real_store._paths.db_path), run_id) as log:
+        log.create_table("predictions", PREDICTIONS_SCHEMA)
+        log.create_table("traces", TRACES_SCHEMA)
+        log.insert("predictions", {"video_id": "vA", "question_id": "q1", "prediction": "A"})
+        log.insert("traces", {"video_id": "vA", "question_id": "q1", "step": 0})
+
+    batch = [_fake_question("q1", "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=run_id))
+    runner_with_real_store._gate_batch_skills = AsyncMock()
+
+    state = MagicMock()
+    state.correctness = {"q1": True}
+    state.gate_cooldown = {}
+    pools = MagicMock()
+    pools.baseline_run_id = "infer_adhoc"
+
+    await runner_with_real_store._run_step(1, 0, 10, batch, pools, state)
+
+    with HarnessLog(
+        str(runner_with_real_store._paths.db_path), run_id, register_run=False
+    ) as log:
+        preds = log.query("SELECT * FROM predictions WHERE run_id=?", (run_id,))
+        traces = log.query("SELECT * FROM traces WHERE run_id=?", (run_id,))
+    assert preds == []
+    assert traces == []
+
+
 @pytest.mark.asyncio
 async def test_run_step_aborts_on_high_degrade_rate(
     runner_with_real_store: Runner, monkeypatch: pytest.MonkeyPatch