refactor: fail-loud unit grading and stricter orphan pair drop

Address review: replace bare KeyError with a contextual ValueError invariant
check in _grade_unit (fail-loud, no catch/skip/fallback); tighten
_drop_orphan_pairs to require exactly one original + one mirror with no extra
illegal-role records (total==2); clarify InferenceResult docstring to unit-grained
semantics. Add tests for missing-prediction descriptive error and extra-illegal-role
pair drop.

tests/unit/test_inference_pair_aggregate.py

@@ -238,6 +238,45 @@ class TestUnitLevelAggregation:
         assert result.correct == 1
         assert any("orphan" in msg for msg in captured), "孤儿 pair 未告警（静默）"
 
+    def test_pair_with_extra_illegal_role_dropped(self) -> None:
+        """pair_id 下混入额外非法 role 记录（total>2）→ 整对剔除、不计入 total。"""
+        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"),
+            # 共享 pair_id 的额外非法记录（重复 original 角色）
+            _make_question("px", answer="C", pair_id="p", question_role="pair_original"),
+        ]
+        records = [
+            _make_record("s1", prediction="B", answer="B"),
+            _make_record("po", prediction="B", answer="B"),
+            _make_record("pm", prediction="A", answer="A"),
+            _make_record("px", prediction="C", answer="C"),
+        ]
+        captured: list[str] = []
+        sink_id = logger.add(captured.append, level="WARNING", format="{message}")
+        try:
+            result = _aggregate_results(records, questions, "run-illegal-role")
+        finally:
+            logger.remove(sink_id)
+
+        assert result.total == 1  # 仅 single 存活，非法配对整对剔除
+        assert result.correct == 1
+        assert any("total=3" in msg for msg in captured), "非法配对未告警（静默）"
+
+    def test_unit_missing_prediction_raises_descriptive_error(self) -> None:
+        """unit 缺 prediction → 描述性 ValueError（fail-loud，非静默、非裸 KeyError）。
+
+        故意破坏聚合不变量（questions 含 s2 但 records 无 s2），验证带上下文报错。
+        """
+        questions = [
+            _make_question("s1", answer="B"),
+            _make_question("s2", answer="A"),
+        ]
+        records = [_make_record("s1", prediction="B", answer="B")]  # 缺 s2 的 record
+        with pytest.raises(ValueError, match=r"缺 prediction.*聚合不变量被破坏"):
+            _aggregate_results(records, questions, "run-broken-invariant")
+
     def test_empty_records_returns_zero(self) -> None:
         """空 records/questions → 零值结果。"""
         result = _aggregate_results([], [], "run-empty")