fix: tolerate backfill under-delivery, fix predict None-fill, drop dead session_id

app/question_gen/adversarial_filter.py

@@ -756,9 +756,18 @@ async def run_adversarial_rounds(
         if deficit <= 0 or round_no + 1 >= config.adversarial_max_rounds:
             break
         new_qs = await backfill(deficit, round_no, all_questions)
-        assert len(new_qs) == deficit, (
+        # 真实 backfill 经 run_pipeline_v2 只返回过门的题：部分题被 gate 拒 = 常态欠产。
-            f"backfill 应产出 {deficit} 题，实得 {len(new_qs)}（缺额驱动契约）"
+        # 容忍 len(new_qs) <= deficit——欠产先并入，deficit 下轮重算继续补，max_rounds 兜底
+        # 防死循环。多产（> deficit）才是契约被破坏的 bug，保留 sanity 断言。
+        assert len(new_qs) <= deficit, (
+            f"backfill 多产: 需 {deficit} 实得 {len(new_qs)}（超额 = 契约被破坏）"
         )
+        if len(new_qs) < deficit:
+            logger.warning(
+                "backfill 欠产: 需 {} 实得 {}，本轮先并入，后续轮次继续补",
+                deficit,
+                len(new_qs),
+            )
         for q in new_qs:
             all_questions[q.question_id] = q
         pending = new_qs  # 只对新补的题重新过滤（已判题走续跑）
@@ -843,7 +852,12 @@ class _RealAgentRunner:
         rows = await RunLogImpl(self._db_path).get_predictions(
             run_id, question_ids=[q.question_id for q in questions]
         )
-        return {r["question_id"]: r["prediction"] for r in rows}
+        # 契约：每个入参 qid 都必须有键。先全填 None，再用回读行覆盖——回读缺行
+        # （agent 未落库/异常）的 qid 保持 None，不静默丢键。
+        preds: dict[str, str | None] = {q.question_id: None for q in questions}
+        for r in rows:
+            preds[r["question_id"]] = r["prediction"]
+        return preds
 
 
 # ---------------------------------------------------------------------------
@@ -873,7 +887,6 @@ def build_backfill(
     store: QuestionGenStore,
     pipeline_config: PipelineConfig,
     filter_task_types: tuple[str, ...],
-    session_id: str,
 ) -> BackfillFn:
     """组装真实 backfill 回调 — 缺额驱动 run_pipeline_v2，返回新增 AR 题。
 
@@ -883,6 +896,7 @@ def build_backfill(
       - initial_embed_pool = 已有题 embedding（跨 run 去重）；
       - seq_offset = 已有题最大 seq（新题续编，防撞 question_id）。
     仅补 filter_task_types（当前锁定为 AR 单类），返回 PipelineResult.accepted。
+    session/run_id 由 run_pipeline_v2 内部自生成，故本层不接受 session_id。
 
     参数:
         trees: video_id → 三层树索引（生成素材来源）。
@@ -892,7 +906,6 @@ def build_backfill(
         store: 出题持久化。
         pipeline_config: ar30 原配置（除 per_type 外全部继承）。
         filter_task_types: 补生成的题型（仅这些）。
-        session_id: 遥测会话 ID（透传给管线子调用）。
 
     返回:
         BackfillFn 闭包。

tests/integration/test_adversarial_filter_e2e.py

@@ -275,3 +275,27 @@ async def test_real_agent_runner_predict_roundtrips_predictions(tmp_path: Path)
         "smoke_r0", question_ids=["smoke"]
     )
     assert rows and rows[0]["prediction"] == "B"
+
+
+@pytest.mark.asyncio
+async def test_real_agent_runner_predict_fills_none_for_missing(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """predict 对 predictions 表缺失的 qid 返回 None（契约：每个入参 qid 都有键）。"""
+
+    async def _noop_run_inference(*args: Any, **kwargs: Any) -> None:
+        """空跑：不写 predictions 表，模拟回读缺行。"""
+        return None
+
+    monkeypatch.setattr("app.harness.inference.run_inference", _noop_run_inference)
+    runner = _RealAgentRunner(
+        llm=_MockLLM(answer="B"),
+        tool_dispatch_fn=lambda *a, **k: None,
+        prompt_builder=lambda q: ("s", "u"),
+        db_path=str(tmp_path / "harness.db"),
+        concurrency=1,
+        skill_mode="none",
+        model="mock",
+    )
+    preds = await runner.predict([_smoke_q("missing")], max_steps=1, run_id="r_none")
+    assert preds == {"missing": None}  # 缺回读行 → 该 qid 键存在且为 None

tests/unit/test_adversarial_iteration.py

@@ -163,3 +163,29 @@ async def test_rounds_backfill_then_stop_at_max(tmp_path):
     ids = {d["question_id"] for d in json.loads(final_path.read_text(encoding="utf-8"))}
     assert "q0" in ids and any(x.startswith("bf0_") for x in ids)
     store.close()
+
+
+@pytest.mark.asyncio
+async def test_rounds_tolerate_backfill_under_delivery(tmp_path):
+    """backfill 欠产（实得 < deficit，被 gate 拒常态）→ 不崩，并入实得题，达上限停。"""
+    store = QuestionGenStore(str(tmp_path / "q.db"))
+    agent = _FakeAgent(pred="B")
+    # 每轮只产出 deficit-1 道（模拟部分题被 gate 拒的真实欠产）
+    backfill = _CountingBackfill(lambda d, r: [_q(f"bf{r}_{i}") for i in range(max(d - 1, 0))])
+    final_path = tmp_path / "accepted_questions_final.json"
+    await run_adversarial_rounds(
+        [_q("q0")],
+        agent=agent,
+        vlm=object(),
+        store=store,
+        trees={},
+        config=AdversarialFilterConfig(adversarial_max_rounds=3),
+        final_path=final_path,
+        target=99,  # 永远达不到 → 靠 max_rounds 终止，不因欠产崩溃或死循环
+        backfill=backfill,
+        session_id="s",
+    )
+    assert backfill.calls == 2  # round0/round1 各补一次；round2 达上限 break
+    ids = {d["question_id"] for d in json.loads(final_path.read_text(encoding="utf-8"))}
+    assert "q0" in ids and any(x.startswith("bf0_") for x in ids)  # 实得欠产题已并入
+    store.close()

tools/generate_questions.py

@@ -1318,7 +1318,6 @@ async def _run_adversarial_filter(args: argparse.Namespace) -> None:
         store=store,
         pipeline_config=pipeline_config,
         filter_task_types=filter_config.filter_task_types,
-        session_id=args.session_id,
     )
 
     # Phase 6: 运行对抗过滤