Initial standalone memabra release

tests/test_promotion.py

@@ -0,0 +1,112 @@
+from __future__ import annotations
+
+import pytest
+
+from memabra.promotion import PromotionDecision, PromotionPolicy
+from memabra.evaluator import EvaluationResult
+
+
+class TestPromotionPolicy:
+    def test_accepted_when_challenger_improves_on_all_metrics(self):
+        policy = PromotionPolicy(
+            min_reward_delta=0.01,
+            max_error_rate_increase=0.05,
+            max_latency_increase_ms=100.0,
+            required_task_count=2,
+        )
+        baseline = EvaluationResult(
+            task_count=2,
+            avg_reward=0.5,
+            error_rate=0.1,
+            avg_latency_ms=50.0,
+        )
+        challenger = EvaluationResult(
+            task_count=2,
+            avg_reward=0.6,
+            error_rate=0.05,
+            avg_latency_ms=45.0,
+        )
+
+        decision = policy.evaluate(baseline, challenger)
+
+        assert isinstance(decision, PromotionDecision)
+        assert decision.accepted is True
+        assert decision.reasons == []
+        assert decision.metrics["reward_delta"] == pytest.approx(0.1, abs=0.001)
+        assert decision.metrics["error_rate_delta"] == pytest.approx(-0.05, abs=0.001)
+        assert decision.metrics["latency_delta_ms"] == pytest.approx(-5.0, abs=0.001)
+
+    def test_rejected_when_reward_delta_below_minimum(self):
+        policy = PromotionPolicy(
+            min_reward_delta=0.1,
+            max_error_rate_increase=0.05,
+            max_latency_increase_ms=100.0,
+            required_task_count=2,
+        )
+        baseline = EvaluationResult(task_count=2, avg_reward=0.5, error_rate=0.1, avg_latency_ms=50.0)
+        challenger = EvaluationResult(task_count=2, avg_reward=0.55, error_rate=0.1, avg_latency_ms=50.0)
+
+        decision = policy.evaluate(baseline, challenger)
+
+        assert decision.accepted is False
+        assert any("reward" in r.lower() for r in decision.reasons)
+
+    def test_rejected_when_error_rate_increase_exceeds_max(self):
+        policy = PromotionPolicy(
+            min_reward_delta=0.01,
+            max_error_rate_increase=0.05,
+            max_latency_increase_ms=100.0,
+            required_task_count=2,
+        )
+        baseline = EvaluationResult(task_count=2, avg_reward=0.5, error_rate=0.1, avg_latency_ms=50.0)
+        challenger = EvaluationResult(task_count=2, avg_reward=0.6, error_rate=0.2, avg_latency_ms=50.0)
+
+        decision = policy.evaluate(baseline, challenger)
+
+        assert decision.accepted is False
+        assert any("error" in r.lower() for r in decision.reasons)
+
+    def test_rejected_when_latency_increase_exceeds_max(self):
+        policy = PromotionPolicy(
+            min_reward_delta=0.01,
+            max_error_rate_increase=0.05,
+            max_latency_increase_ms=10.0,
+            required_task_count=2,
+        )
+        baseline = EvaluationResult(task_count=2, avg_reward=0.5, error_rate=0.1, avg_latency_ms=50.0)
+        challenger = EvaluationResult(task_count=2, avg_reward=0.6, error_rate=0.1, avg_latency_ms=65.0)
+
+        decision = policy.evaluate(baseline, challenger)
+
+        assert decision.accepted is False
+        assert any("latency" in r.lower() for r in decision.reasons)
+
+    def test_rejected_when_task_count_below_required(self):
+        policy = PromotionPolicy(
+            min_reward_delta=0.01,
+            max_error_rate_increase=0.05,
+            max_latency_increase_ms=100.0,
+            required_task_count=5,
+        )
+        baseline = EvaluationResult(task_count=2, avg_reward=0.5, error_rate=0.1, avg_latency_ms=50.0)
+        challenger = EvaluationResult(task_count=2, avg_reward=0.6, error_rate=0.1, avg_latency_ms=50.0)
+
+        decision = policy.evaluate(baseline, challenger)
+
+        assert decision.accepted is False
+        assert any("task count" in r.lower() for r in decision.reasons)
+
+    def test_multiple_rejection_reasons_accumulate(self):
+        policy = PromotionPolicy(
+            min_reward_delta=0.2,
+            max_error_rate_increase=0.01,
+            max_latency_increase_ms=10.0,
+            required_task_count=10,
+        )
+        baseline = EvaluationResult(task_count=2, avg_reward=0.5, error_rate=0.1, avg_latency_ms=50.0)
+        challenger = EvaluationResult(task_count=2, avg_reward=0.55, error_rate=0.15, avg_latency_ms=70.0)
+
+        decision = policy.evaluate(baseline, challenger)
+
+        assert decision.accepted is False
+        assert len(decision.reasons) >= 3