Drive corpus sampling from calibration failures

backend/tests/test_failure_driven_yolo_sampling.py

@@ -52,3 +52,31 @@ def test_sampling_repeats_only_failed_region_train_tiles() -> None:
     assert not any("protected" in path for path in paths)
     assert metadata["protected_samples_in_training"] == []
     assert metadata["weak_recall_regions"] == ["flanders"]
+
+
+def test_sampling_can_use_calibration_before_test_is_opened() -> None:
+    manifest = {"samples": [{"sample_slug": "train-fl", "split": "train", "region": "flanders"}]}
+    summary = {
+        "tiles": [
+            {"sample_slug": "train-fl", "split": "train", "label_count": 1, "image_path": "/tmp/fl.png"}
+        ]
+    }
+    assessment = {
+        "status": "continue_training_loop",
+        "gates": {
+            "min_region_f1": 0.45,
+            "min_region_precision": 0.5,
+            "min_region_recall": 0.4,
+            "max_pure_empty_false_positives": 0,
+        },
+        "calibration": {
+            "regions": {"flanders": {"f1": 0.4, "precision": 0.6, "recall": 0.35}}
+        },
+        "test": None,
+        "background": None,
+    }
+    paths, metadata = MODULE.build_sampling(
+        summary=summary, manifest=manifest, assessment=assessment
+    )
+    assert len(paths) == 3
+    assert metadata["failure_evidence_source"] == "calibration"