Require tile quality evidence for training

backend/tests/test_belgium_training_loop.py

@@ -98,6 +98,11 @@ def test_dry_run_can_gate_existing_checkpoint_without_training(tmp_path: Path) -
         "status": "needs_human_review", "failures": [],
         "manifest_immutable": True, "spatial_leakage_status": "ok",
     }))
+    quality = tmp_path / "quality.json"
+    quality.write_text(json.dumps({
+        "status": "ok", "low_variance_positive_tile_count": 0,
+        "label_stats": {"invalid_label_count": 0, "missing_label_file_count": 0},
+    }))
     result = subprocess.run(
         [
             sys.executable, str(SCRIPT),
@@ -105,6 +110,7 @@ def test_dry_run_can_gate_existing_checkpoint_without_training(tmp_path: Path) -
             "--train-yaml", str(tmp_path / "dataset.yaml"),
             "--train-summary", str(tmp_path / "train-summary.json"),
             "--dataset-audit", str(audit),
+            "--train-quality-audit", str(quality),
             "--calibration-summary", str(tmp_path / "cal.json"),
             "--test-summary", str(tmp_path / "test.json"),
             "--background-summary", str(tmp_path / "background.json"),
@@ -128,6 +134,11 @@ def test_existing_checkpoint_iteration_directory_can_be_created_without_yolo(tmp
 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}))
+    quality = tmp_path / "quality.json"
+    quality.write_text(json.dumps({
+        "status": "ok", "low_variance_positive_tile_count": 0,
+        "label_stats": {"invalid_label_count": 0, "missing_label_file_count": 0},
+    }))
     result = subprocess.run(
         [
             sys.executable,
@@ -140,6 +151,8 @@ def test_loop_refuses_failed_dataset_audit(tmp_path: Path) -> None:
             str(tmp_path / "train-summary.json"),
             "--dataset-audit",
             str(audit),
+            "--train-quality-audit",
+            str(quality),
             "--calibration-summary",
             str(tmp_path / "cal.json"),
             "--test-summary",
@@ -168,7 +181,11 @@ def test_pending_human_review_does_not_block_objective_training() -> None:
         "low_variance_positive_tile_count": 0,
         "review_complete": False,
     }
-    assert MODULE.dataset_audit_failures(audit) == []
+    quality = {
+        "status": "ok", "low_variance_positive_tile_count": 0,
+        "label_stats": {"invalid_label_count": 0, "missing_label_file_count": 0},
+    }
+    assert MODULE.dataset_audit_failures(audit, quality) == []
 
 
 def test_training_audit_still_fails_closed_on_automated_integrity_gates() -> None:
@@ -179,11 +196,30 @@ def test_training_audit_still_fails_closed_on_automated_integrity_gates() -> Non
         "spatial_leakage_status": "failed",
         "low_variance_positive_tile_count": 2,
     }
-    failures = MODULE.dataset_audit_failures(audit)
+    quality = {
+        "status": "failed", "low_variance_positive_tile_count": 2,
+        "label_stats": {"invalid_label_count": 1, "missing_label_file_count": 1},
+    }
+    failures = MODULE.dataset_audit_failures(audit, quality)
     assert "wallonia/test below minimum" in failures
     assert "corpus manifest is not immutable" in failures
     assert "spatial leakage audit is not ok" in failures
     assert "dataset contains blank/low-variance positive tiles" in failures
+    assert "train tile quality audit is not ok" in failures
+    assert "train tile quality audit contains invalid labels" in failures
+    assert "train tile quality audit contains missing label files" in failures
+
+
+def test_missing_tile_quality_evidence_fails_closed() -> None:
+    audit = {
+        "status": "needs_human_review", "failures": [],
+        "manifest_immutable": True, "spatial_leakage_status": "ok",
+    }
+    failures = MODULE.dataset_audit_failures(audit, {})
+    assert "train tile quality audit is not ok" in failures
+    assert "train tile quality audit contains invalid labels" in failures
+    assert "train tile quality audit contains missing label files" in failures
+    assert "dataset contains blank/low-variance positive tiles" in failures
 
 
 def test_calibration_failure_blocks_protected_evaluation() -> None: