feat(training): pair evaluations across physical sides

training/test_evaluate.py

@@ -0,0 +1,62 @@
+"""Offline regression checks for evaluate.py's paired side-swap logic."""
+
+import unittest
+from unittest.mock import patch
+
+import evaluate
+
+
+class EvaluatePairTests(unittest.TestCase):
+    @patch("evaluate.run_half")
+    def test_replays_same_seed_with_models_and_masks_swapped(self, run_half) -> None:
+        run_half.side_effect = [
+            {"episodes": 4, "goals_a": 3, "goals_b": 1, "draws": 0},
+            {"episodes": 4, "goals_a": 2, "goals_b": 1, "draws": 1},
+        ]
+
+        record = evaluate.evaluate_pair(
+            "godot", "candidate.json", "reference.json", 8, 16, 42,
+            grounded_a=False, grounded_b=True,
+        )
+
+        self.assertEqual(run_half.call_args_list[0].args, ("godot", "candidate.json", "reference.json", 4, 16, 42))
+        self.assertEqual(run_half.call_args_list[1].args, ("godot", "reference.json", "candidate.json", 4, 16, 42))
+        self.assertEqual(
+            run_half.call_args_list[0].kwargs,
+            {"grounded_a": False, "grounded_b": True},
+        )
+        self.assertEqual(
+            run_half.call_args_list[1].kwargs,
+            {"grounded_a": True, "grounded_b": False},
+        )
+        self.assertEqual(record["wins_a"], 4)
+        self.assertEqual(record["wins_b"], 3)
+        self.assertEqual(record["draws"], 1)
+        self.assertEqual(record["physical_team_wins"], {"team_0": 5, "team_1": 2})
+        self.assertEqual(record["side_results"]["a_team_0"]["wins_a"], 3)
+        self.assertEqual(record["side_results"]["a_team_1"]["wins_a"], 1)
+
+    def test_rejects_unpaired_episode_counts(self) -> None:
+        for episodes in (0, 1, 3, 99):
+            with self.subTest(episodes=episodes):
+                with self.assertRaisesRegex(ValueError, "even number"):
+                    evaluate.evaluate_pair("godot", "a", "b", episodes, 16, 1)
+
+    @patch("evaluate.run_half")
+    def test_identical_policy_results_cancel_physical_side_bias(self, run_half) -> None:
+        # Replaying the same deterministic matchup must produce the same
+        # physical-team result. Model A receives opposite sides in the two
+        # halves, so even a large team-0 advantage cancels exactly.
+        physical_result = {"episodes": 10, "goals_a": 8, "goals_b": 1, "draws": 1}
+        run_half.side_effect = [physical_result, physical_result]
+
+        record = evaluate.evaluate_pair("godot", "same.json", "same.json", 20, 16, 7)
+
+        self.assertEqual(record["wins_a"], 9)
+        self.assertEqual(record["wins_b"], 9)
+        self.assertEqual(record["draws"], 2)
+        self.assertEqual(record["win_rate_a"], 0.45)
+
+
+if __name__ == "__main__":
+    unittest.main()