feat/ 导出开发容器内的数据用于评估

api/tests/evaluation/test_gating_service.py

@@ -0,0 +1,54 @@
+from unittest.mock import MagicMock
+
+from app.features.evaluation.gating_service import compute_gate
+
+
+def _case(cid: str, protected: bool = False):
+    c = MagicMock()
+    c.id = cid
+    c.title = None
+    c.is_protected = protected
+    return c
+
+
+def _run(case_id: str, side: str, composite: float, status: str = "completed"):
+    r = MagicMock()
+    r.case_id = case_id
+    r.side = side
+    r.status = status
+    r.composite_score = composite
+    return r
+
+
+def test_gate_passes_when_mean_up_and_no_protected_regression() -> None:
+    cases = [_case("1"), _case("2")]
+    runs = [
+        _run("1", "baseline", 50),
+        _run("1", "candidate", 60),
+        _run("2", "baseline", 40),
+        _run("2", "candidate", 55),
+    ]
+    g = compute_gate(cases=cases, runs=runs, regression_threshold=2.0)
+    assert g.passed
+    assert g.mean_delta > 0
+
+
+def test_gate_fails_on_protected_regression() -> None:
+    cases = [_case("1", protected=True)]
+    runs = [
+        _run("1", "baseline", 80.0),
+        _run("1", "candidate", 75.0),
+    ]
+    g = compute_gate(cases=cases, runs=runs, regression_threshold=2.0)
+    assert not g.passed
+    assert len(g.protected_regressions) == 1
+
+
+def test_gate_fails_when_mean_not_higher() -> None:
+    cases = [_case("1")]
+    runs = [
+        _run("1", "baseline", 70.0),
+        _run("1", "candidate", 69.0),
+    ]
+    g = compute_gate(cases=cases, runs=runs, regression_threshold=2.0)
+    assert not g.passed