auto-sync 2026-07-04T06:15:18Z workspace (part 4)

workspace/tests/test_selector_diagnostic_sweep.py

@@ -85,3 +85,44 @@ def test_selector_diagnostic_sweep_exports_best_by_family(tmp_path: Path, monkey
         "table.tex",
     ):
         assert (out_dir / filename).exists()
+
+
+def test_selector_diagnostic_sweep_keeps_utility_energy_as_own_family(
+    tmp_path: Path, monkeypatch
+) -> None:
+    monkeypatch.chdir(tmp_path)
+    utility = tmp_path / "runs/ctt_dominance_utility_energy_val_to_test_seed2/metrics.json"
+    utility.parent.mkdir(parents=True, exist_ok=True)
+    utility.write_text(
+        json.dumps(
+            {
+                "report_type": "dominance_calibrated_selector_eval",
+                "k": 16,
+                "score_source": "checkpoint",
+                "tau_mode": "auto",
+                "data_hash": "data-utility",
+                "split_hash": "split-utility",
+                "eval_summary": {
+                    "base_success": 0.29,
+                    "selected_success": 0.28,
+                    "proposal_oracle_success": 0.57,
+                    "coverage": 0.17,
+                    "success_support_gap": 0.20,
+                    "success_selector_gap": 0.31,
+                    "outcome_ptr": 0.55,
+                    "pairwise_causal_calibration_ece": 0.02,
+                },
+            }
+        )
+        + "\n"
+    )
+
+    out_dir = tmp_path / "runs/ctt_selector_diagnostic_sweep"
+    assert sweep.main(["--metrics", str(utility), "--out-dir", str(out_dir)]) == 0
+
+    payload = json.loads((out_dir / "metrics.json").read_text())
+    by_family = {row["family"]: row for row in payload["best_by_family"]}
+    assert by_family["K16 env_clip utility-energy"]["seed"] == "2"
+    assert "checkpoint utility energy" in by_family["K16 env_clip utility-energy"]["selector"]
+    by_seed = json.loads((out_dir / "metrics_by_seed.json").read_text())
+    assert by_seed["2"]["selected_success"] == 0.28