feat: aggregate inference by question unit with pair AND

Reuse build_units/unit_correctness (pair contract single entry) in the
inference aggregation step: single questions count as one unit, AR pairs
collapse original+mirror into one unit scored by bidirectional AND. total/
correct/per_task_type are unit-grained; orphan pairs (missing one side) are
warned and dropped, not counted. Per-question predictions still land row by
row (traceability unchanged).

tests/unit/test_inference_pair_aggregate.py

@@ -0,0 +1,332 @@
+"""inference pair-level 双向 AND 聚合单元测试（Task 6）。
+
+覆盖 question-gen v3 Phase 1 Task 6 的核心契约：
+- 逐题推理不变：每条 GeneratedQuestion 照常各答一次、per-question prediction
+  仍逐题落 predictions 表（保留逐题溯源）。
+- pair 按 pair_id 收齐 original + mirror 后合成 1 条 unit-level 记录，
+  pair 正确 = (P.pred==P.answer) AND (Q.pred==Q.answer)（双向 AND）。
+- InferenceResult.total / correct / per_task_type 全部按 unit 粒度
+  （single 计 1，pair 计 1）。
+- 孤儿 pair（收不齐 2 条）→ 告警 + 剔除该 unit、不计入 total（不静默）。
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+from unittest.mock import AsyncMock
+
+import pytest
+from loguru import logger
+
+from app.harness.inference import _aggregate_results, run_inference
+from app.harness.log import HarnessLog
+from core.types import GeneratedQuestion, LLMResponse
+
+# ── 测试基础设施 ──────────────────────────────────────────────────
+
+
+def _make_question(
+    question_id: str,
+    *,
+    task_type: str = "Action Reasoning",
+    answer: str = "B",
+    pair_id: str | None = None,
+    question_role: str = "single",
+    flip_axis: str | None = None,
+    video_id: str = "v1",
+) -> GeneratedQuestion:
+    """构造测试题目；pair_id 非空时视为孪生对成员。"""
+    return GeneratedQuestion(
+        question_id=question_id,
+        video_id=video_id,
+        task_type=task_type,
+        question="测试问题",
+        options=("A. 选项A", "B. 选项B", "C. 选项C", "D. 选项D"),
+        answer=answer,
+        source_nodes=("L1_001",),
+        difficulty="medium",
+        pair_id=pair_id,
+        question_role=question_role,
+        flip_axis=flip_axis if pair_id else None,
+    )
+
+
+def _make_record(
+    question_id: str,
+    *,
+    prediction: str | None,
+    answer: str = "B",
+    task_type: str = "Action Reasoning",
+    steps_used: int = 2,
+    prompt_tokens: int = 100,
+    completion_tokens: int = 50,
+    stop_reason: str = "finished",
+) -> dict[str, Any]:
+    """构造与题目匹配的 prediction record（键与 _run_single_question 一致）。"""
+    return {
+        "question_id": question_id,
+        "prediction": prediction,
+        "answer": answer,
+        "task_type": task_type,
+        "steps_used": steps_used,
+        "prompt_tokens": prompt_tokens,
+        "completion_tokens": completion_tokens,
+        "stop_reason": stop_reason,
+    }
+
+
+def _make_llm_response(answer: str = "B") -> LLMResponse:
+    """构造 submit_answer 场景的 LLMResponse。"""
+    content = json.dumps(
+        {
+            "reflect": {"observation": "找到答案"},
+            "plan": {"next_step": "提交"},
+            "action": {
+                "tool": "submit_answer",
+                "args": {"answer": answer, "evidence": "证据", "reasoning": "推理"},
+            },
+        }
+    )
+    return LLMResponse(
+        content=content,
+        thinking="思考",
+        model="test-model",
+        provider="test",
+        prompt_tokens=100,
+        completion_tokens=50,
+        latency_ms=200,
+        ttft_ms=30.0,
+        max_inter_token_ms=5.0,
+        cache_hit=False,
+        call_id="test-call-001",
+    )
+
+
+async def _stub_tool_dispatch(
+    tool_name: str, args: dict[str, Any], *, context: dict[str, Any]
+) -> str:
+    """测试用工具调度函数。"""
+    if tool_name == "submit_answer":
+        return "答案已提交"
+    raise ValueError(f"未知工具: {tool_name}")
+
+
+def _stub_prompt_builder(qa: GeneratedQuestion) -> tuple[str, str]:
+    """测试用 prompt 构建函数。"""
+    return "系统提示词", f"用户问题: {qa.question}"
+
+
+@pytest.fixture
+def harness_log(tmp_path: Any, request: Any) -> HarnessLog:
+    """创建临时 HarnessLog 实例。"""
+    db_path = str(tmp_path / f"harness_{id(request)}.db")
+    log = HarnessLog(db_path, "test-run")
+    yield log
+    log.close()
+
+
+# ── unit 粒度聚合（_aggregate_results 直测） ─────────────────────────
+
+
+class TestUnitLevelAggregation:
+    """_aggregate_results 按 unit 粒度聚合测试。"""
+
+    def test_single_units_counted_per_question(self) -> None:
+        """全 single：total = single 数，correct 逐题判定。"""
+        questions = [
+            _make_question("s1", answer="B"),
+            _make_question("s2", answer="A"),
+        ]
+        records = [
+            _make_record("s1", prediction="B", answer="B"),
+            _make_record("s2", prediction="C", answer="A"),
+        ]
+        result = _aggregate_results(records, questions, "run-single")
+        assert result.total == 2
+        assert result.correct == 1
+        assert abs(result.accuracy - 0.5) < 1e-9
+
+    def test_pair_both_correct_is_one_correct_unit(self) -> None:
+        """pair 两题皆对 → 1 个 unit、correct=1。"""
+        questions = [
+            _make_question("po", answer="B", pair_id="p", question_role="pair_original"),
+            _make_question("pm", answer="A", pair_id="p", question_role="pair_mirror"),
+        ]
+        records = [
+            _make_record("po", prediction="B", answer="B"),
+            _make_record("pm", prediction="A", answer="A"),
+        ]
+        result = _aggregate_results(records, questions, "run-pair-ok")
+        assert result.total == 1
+        assert result.correct == 1
+        assert result.accuracy == 1.0
+
+    def test_pair_one_wrong_fails_by_and(self) -> None:
+        """pair 一题错 → 双向 AND 判 unit 错，correct=0。"""
+        questions = [
+            _make_question("po", answer="B", pair_id="p", question_role="pair_original"),
+            _make_question("pm", answer="A", pair_id="p", question_role="pair_mirror"),
+        ]
+        records = [
+            _make_record("po", prediction="B", answer="B"),  # 对
+            _make_record("pm", prediction="D", answer="A"),  # 错
+        ]
+        result = _aggregate_results(records, questions, "run-pair-half")
+        assert result.total == 1
+        assert result.correct == 0
+        assert result.accuracy == 0.0
+
+    def test_mixed_single_and_pair_unit_total(self) -> None:
+        """single + pair 混合：total = single 数 + pair 数（pair 计 1）。"""
+        questions = [
+            _make_question("s1", answer="B"),
+            _make_question("po", answer="B", pair_id="p", question_role="pair_original"),
+            _make_question("pm", answer="A", pair_id="p", question_role="pair_mirror"),
+        ]
+        records = [
+            _make_record("s1", prediction="B", answer="B"),
+            _make_record("po", prediction="B", answer="B"),
+            _make_record("pm", prediction="A", answer="A"),
+        ]
+        result = _aggregate_results(records, questions, "run-mixed")
+        assert result.total == 2  # 1 single + 1 pair
+        assert result.correct == 2
+
+    def test_per_task_type_by_unit(self) -> None:
+        """per_task_type 按 unit 计数：pair 归入其 task_type 计 1 个 unit。"""
+        questions = [
+            _make_question("s1", task_type="SP", answer="B"),
+            _make_question(
+                "po", task_type="AR", answer="B", pair_id="p", question_role="pair_original"
+            ),
+            _make_question(
+                "pm", task_type="AR", answer="A", pair_id="p", question_role="pair_mirror"
+            ),
+        ]
+        records = [
+            _make_record("s1", prediction="B", answer="B", task_type="SP"),
+            _make_record("po", prediction="B", answer="B", task_type="AR"),
+            _make_record("pm", prediction="C", answer="A", task_type="AR"),  # pair 错
+        ]
+        result = _aggregate_results(records, questions, "run-tt")
+        assert result.per_task_type["AR"]["total"] == 1  # pair 计 1 个 unit
+        assert result.per_task_type["AR"]["correct"] == 0
+        assert result.per_task_type["SP"]["total"] == 1
+        assert result.per_task_type["SP"]["correct"] == 1
+
+    def test_orphan_pair_dropped_and_warned(self) -> None:
+        """孤儿 pair（收不齐 2 条）→ 告警 + 剔除、不计入 total。"""
+        questions = [
+            _make_question("s1", answer="B"),
+            _make_question(
+                "po", answer="B", pair_id="orphan", question_role="pair_original"
+            ),  # 缺 mirror
+        ]
+        records = [
+            _make_record("s1", prediction="B", answer="B"),
+            _make_record("po", prediction="B", answer="B"),
+        ]
+        captured: list[str] = []
+        sink_id = logger.add(captured.append, level="WARNING", format="{message}")
+        try:
+            result = _aggregate_results(records, questions, "run-orphan")
+        finally:
+            logger.remove(sink_id)
+
+        assert result.total == 1  # 仅 single，孤儿 pair 被剔除
+        assert result.correct == 1
+        assert any("orphan" in msg for msg in captured), "孤儿 pair 未告警（静默）"
+
+    def test_empty_records_returns_zero(self) -> None:
+        """空 records/questions → 零值结果。"""
+        result = _aggregate_results([], [], "run-empty")
+        assert result.total == 0
+        assert result.correct == 0
+        assert result.accuracy == 0.0
+        assert result.per_task_type == {}
+
+    def test_token_and_steps_span_all_records(self) -> None:
+        """token/steps 诊断字段覆盖全部 record（含 pair 两条）。"""
+        questions = [
+            _make_question("po", answer="B", pair_id="p", question_role="pair_original"),
+            _make_question("pm", answer="A", pair_id="p", question_role="pair_mirror"),
+        ]
+        records = [
+            _make_record("po", prediction="B", answer="B", steps_used=3, prompt_tokens=100),
+            _make_record("pm", prediction="A", answer="A", steps_used=1, prompt_tokens=200),
+        ]
+        result = _aggregate_results(records, questions, "run-diag")
+        assert result.token_usage["prompt_tokens"] == 300
+        assert abs(result.steps_mean - 2.0) < 1e-9  # (3+1)/2 record 粒度
+
+
+# ── run_inference 端到端：逐题溯源 + pair 聚合 ──────────────────────
+
+
+class TestRunInferencePairEndToEnd:
+    """run_inference pair 端到端：逐题落库不变 + unit 级聚合。"""
+
+    @pytest.mark.asyncio
+    async def test_pair_predictions_persisted_per_question(
+        self, harness_log: HarnessLog
+    ) -> None:
+        """pair 两题各自逐题落 predictions（保留逐题溯源），聚合按 unit。"""
+        llm = AsyncMock()
+        llm.chat.return_value = _make_llm_response(answer="B")
+
+        questions = [
+            _make_question("po", answer="B", pair_id="p", question_role="pair_original"),
+            _make_question("pm", answer="B", pair_id="p", question_role="pair_mirror"),
+        ]
+        result = await run_inference(
+            questions,
+            llm=llm,
+            tool_dispatch_fn=_stub_tool_dispatch,
+            prompt_builder=_stub_prompt_builder,
+            log=harness_log,
+            run_id="run-pair-e2e",
+            concurrency=2,
+            max_steps=10,
+            skill_mode="auto",
+        )
+
+        # unit 级：1 个 pair unit，两题皆对 → correct=1
+        assert result.total == 1
+        assert result.correct == 1
+
+        # 逐题溯源：predictions 表两条 record 都在
+        rows = harness_log.query("SELECT * FROM predictions WHERE run_id = ?", ("test-run",))
+        qids = {r["question_id"] for r in rows}
+        assert qids == {"po", "pm"}
+
+    @pytest.mark.asyncio
+    async def test_orphan_pair_excluded_single_survives(
+        self, harness_log: HarnessLog
+    ) -> None:
+        """run_inference 中孤儿 pair 被剔除、single 仍计入。"""
+        llm = AsyncMock()
+        llm.chat.return_value = _make_llm_response(answer="B")
+
+        questions = [
+            _make_question("s1", answer="B"),
+            _make_question(
+                "po", answer="B", pair_id="orphan", question_role="pair_original"
+            ),
+        ]
+        result = await run_inference(
+            questions,
+            llm=llm,
+            tool_dispatch_fn=_stub_tool_dispatch,
+            prompt_builder=_stub_prompt_builder,
+            log=harness_log,
+            run_id="run-orphan-e2e",
+            concurrency=2,
+            max_steps=10,
+            skill_mode="auto",
+        )
+
+        assert result.total == 1  # single 存活，孤儿剔除
+        # 逐题溯源：孤儿题仍逐题落库（推理不变）
+        rows = harness_log.query("SELECT * FROM predictions WHERE run_id = ?", ("test-run",))
+        assert {r["question_id"] for r in rows} == {"s1", "po"}