Prioritize recall in failure-driven sampling

scripts/build_failure_driven_yolo_sampling.py

@@ -94,6 +94,15 @@ def build_sampling(
         and background["pure_empty_false_positives"]
         > gates["max_pure_empty_false_positives"]
     )
+    recall_dominant_regions = {
+        region
+        for region in weak_recall_regions & weak_precision_regions
+        if not background_failed
+        and (
+            regions[region]["recall"] / gates["min_region_recall"]
+            < regions[region]["precision"] / gates["min_region_precision"]
+        )
+    }
     weak_recall_contexts: set[tuple[str, str]] = set()
     weak_precision_contexts: set[tuple[str, str]] = set()
     for sample_slug, metrics in evaluation.get("samples", {}).items():
@@ -140,7 +149,11 @@ def build_sampling(
             # Precision-only correction still needs positive examples to avoid
             # shifting the classifier toward background and sacrificing recall.
             repeat = precision_positive_repeat
-        if tile["label_count"] == 0 and (background_failed or region in weak_precision_regions):
+        if (
+            tile["label_count"] == 0
+            and (background_failed or region in weak_precision_regions)
+            and region not in recall_dominant_regions
+        ):
             repeat = (
                 context_negative_repeat
                 if context_key in targeted_negative_contexts
@@ -197,6 +210,7 @@ def build_sampling(
         "failure_evidence_source": "test" if assessment.get("test") else "calibration",
         "weak_recall_regions": sorted(weak_recall_regions),
         "weak_precision_regions": sorted(weak_precision_regions),
+        "recall_dominant_regions": sorted(recall_dominant_regions),
         "weak_recall_contexts": [f"{region}:{context}" for region, context in sorted(weak_recall_contexts)],
         "weak_precision_contexts": [f"{region}:{context}" for region, context in sorted(weak_precision_contexts)],
         "targeted_negative_contexts": [