Gate completed checkpoints through training loop

backend/tests/test_belgium_training_loop.py

@@ -76,6 +76,31 @@ def test_failed_iteration_builds_train_only_sampling_for_next_checkpoint(tmp_pat
     assert command[command.index("--output-dir") + 1].endswith("failure-driven-training")
 
 
+def test_dry_run_can_gate_existing_checkpoint_without_training(tmp_path: Path) -> None:
+    audit = tmp_path / "audit.json"
+    audit.write_text(json.dumps({
+        "status": "needs_human_review", "failures": [],
+        "manifest_immutable": True, "spatial_leakage_status": "ok",
+    }))
+    result = subprocess.run(
+        [
+            sys.executable, str(SCRIPT),
+            "--initial-model", str(tmp_path / "candidate.pt"),
+            "--train-yaml", str(tmp_path / "dataset.yaml"),
+            "--train-summary", str(tmp_path / "train-summary.json"),
+            "--dataset-audit", str(audit),
+            "--calibration-summary", str(tmp_path / "cal.json"),
+            "--test-summary", str(tmp_path / "test.json"),
+            "--background-summary", str(tmp_path / "background.json"),
+            "--corpus-manifest", str(tmp_path / "manifest.json"),
+            "--output-dir", str(tmp_path / "output"),
+            "--evaluate-initial-model", "--dry-run",
+        ], capture_output=True, text=True, check=False,
+    )
+    assert result.returncode == 0
+    assert json.loads(result.stdout) == {"training_command": None, "evaluate_existing": True}
+
+
 def test_loop_refuses_failed_dataset_audit(tmp_path: Path) -> None:
     audit = tmp_path / "audit.json"
     audit.write_text(json.dumps({"status": "needs_attention", "low_variance_positive_tile_count": 4}))