feat: add online evolution loop and 5% gray risk-policy rollout

tests/test_evolution_mvp.py

@@ -0,0 +1,54 @@
+import unittest
+from unittest.mock import patch
+
+from evolution.mvp import Finding, Sample, can_publish_candidate, evaluate_samples
+
+
+class EvolutionMvpTest(unittest.TestCase):
+    def test_evaluate_detects_risk_without_transfer(self):
+        samples = [
+            Sample(
+                customer_id="c1",
+                acc_id="shop",
+                in_ts="2026-02-28 10:00:00",
+                in_text="我要投诉并退款，你们骗人",
+                out_ts="2026-02-28 10:00:10",
+                out_text="这个我不清楚，稍后再说",
+                latency_sec=10,
+            )
+        ]
+        findings = evaluate_samples(samples)
+        kinds = {f.kind for f in findings}
+        self.assertIn("risk_not_transferred", kinds)
+        self.assertIn("weak_reply", kinds)
+
+    def test_publish_gate(self):
+        samples = [
+            Sample(
+                customer_id=f"c{i}",
+                acc_id="shop",
+                in_ts="2026-02-28 10:00:00",
+                in_text="你好",
+                out_ts="2026-02-28 10:00:05",
+                out_text="您好",
+                latency_sec=5,
+            )
+            for i in range(35)
+        ]
+        findings: list[Finding] = []
+        policy = {
+            "publish_gate": {
+                "min_sample_count": 30,
+                "max_high_findings_rate": 0.1,
+                "max_ai_fail_rate": 5.0,
+                "max_transfer_rate": 45.0,
+            }
+        }
+        with patch("utils.metrics_tracker.get_runtime_summary", return_value={"rates": {"ai_fail_rate": 1.0, "transfer_rate": 10.0}}):
+            ok, report = can_publish_candidate(samples, findings, runtime_hours=24, policy=policy)
+        self.assertTrue(ok)
+        self.assertEqual(report["sample_count"], 35)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)