feat: 完成 issue #6 LLM 网关 + RAG 模板化（混合网关主编排）

core/llm-gateway/tests/test_hallucination.py

@@ -0,0 +1,103 @@
+# -*- coding: utf-8 -*-
+"""幻觉/事实性校验（hallucination）单元测试：引用溯源 / 信度阈值 / 评测。"""
+import os
+import sys
+import unittest
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+import _bootstrap  # noqa: F401
+
+from llm_gateway.hallucination import HallucinationGuard  # noqa: E402
+
+
+class CitationCheckTest(unittest.TestCase):
+    def setUp(self):
+        self.guard = HallucinationGuard(default_threshold=0.8)
+
+    def test_declared_sources_are_verified(self):
+        verdict = self.guard.check(
+            answer="炉温偏高应降温。[来源: 沸腾氯化炉异常处置SOP]",
+            sources=["沸腾氯化炉异常处置SOP", "交接班规范"],
+        )
+        self.assertTrue(verdict.supported)
+        self.assertEqual(verdict.action, "pass")
+
+    def test_missing_source_is_unsupported(self):
+        verdict = self.guard.check(
+            answer="应停机。[来源: 不存在的文档]",
+            sources=["沸腾氯化炉异常处置SOP"],
+        )
+        self.assertFalse(verdict.supported)
+        self.assertEqual(verdict.action, "unsupported")
+        self.assertIn("不存在的文档", verdict.missing_sources)
+
+    def test_no_citation_is_supported(self):
+        # 无引用声明 = 不判幻觉（引用为强制项由 RAG 模板保证）
+        verdict = self.guard.check(answer="按操作规程执行。", sources=[])
+        self.assertTrue(verdict.supported)
+        self.assertEqual(verdict.action, "pass")
+
+
+class ConfidenceThresholdTest(unittest.TestCase):
+    def setUp(self):
+        self.guard = HallucinationGuard(default_threshold=0.8)
+
+    def test_low_confidence_high_stakes_human_review(self):
+        verdict = self.guard.check(
+            answer="建议立即停机。[来源: SOP]",
+            sources=["SOP"], confidence=0.55, high_stakes=True,
+        )
+        self.assertEqual(verdict.action, "human_review")
+
+    def test_high_confidence_high_stakes_passes(self):
+        verdict = self.guard.check(
+            answer="建议观察并记录。[来源: SOP]",
+            sources=["SOP"], confidence=0.95, high_stakes=True,
+        )
+        self.assertEqual(verdict.action, "pass")
+
+    def test_low_confidence_non_stakes_passes(self):
+        # 非高利害场景不启用阈值
+        verdict = self.guard.check(
+            answer="一般说明。[来源: 手册]", sources=["手册"],
+            confidence=0.3, high_stakes=False,
+        )
+        self.assertEqual(verdict.action, "pass")
+
+    def test_custom_threshold(self):
+        verdict = self.guard.check(
+            answer="处置建议。[来源: 手册]", sources=["手册"],
+            confidence=0.7, high_stakes=True, threshold=0.6,
+        )
+        self.assertEqual(verdict.action, "pass")
+
+
+class EvaluateTest(unittest.TestCase):
+    def setUp(self):
+        self.guard = HallucinationGuard()
+
+    def test_evaluate_rates(self):
+        samples = [
+            {"answer": "a[来源: X]", "sources": ["X"], "confidence": 0.9, "high_stakes": True},
+            {"answer": "b[来源: Y]", "sources": ["Z"], "confidence": 0.9},
+            {"answer": "c[来源: X]", "sources": ["X"], "confidence": 0.4, "high_stakes": True},
+        ]
+        report = self.guard.evaluate(samples)
+        self.assertEqual(report["total"], 3)
+        # 支持 2 条（a/c），human_review 1 条（c）
+        self.assertEqual(report["supported_rate"], round(2 / 3, 4))
+        self.assertEqual(report["human_review_rate"], round(1 / 3, 4))
+
+    def test_empty_samples(self):
+        report = self.guard.evaluate([])
+        self.assertEqual(report["total"], 0)
+
+    def test_audit_records(self):
+        self.guard.check("x[来源: A]", sources=["A"])
+        records = self.guard.drain_audit()
+        self.assertEqual(len(records), 1)
+        self.assertEqual(records[0]["action"], "pass")
+
+
+if __name__ == "__main__":
+    unittest.main()