Promote focused small-building detector

scripts/README.md

@@ -531,8 +531,27 @@ Current Tower audit status:
   `0.000694274766`, small-box share `0.3832694151486098`, no invalid labels and
   no missing label files. The balanced visual pass rendered 40 tiles across all
   19 source samples that retained at least one tile, with no invalid labels,
-  missing images or low-variance selections. A new candidate may be trained,
-  but remains inactive until positive and split-background promotion gates pass.
+  missing images or low-variance selections. Its promoted model remains the
+  higher-precision legacy `0.15` operator profile.
+- `yolo-building-aoi1024-smallbld-minpx3vis035`: focused small-building corpus
+  exported from an explicit 23-sample subset. Beerse, Rijkevorsel, Hoogstraten
+  and Vorselaar extend training; Vosselaar and Grobbendonk are validation-only;
+  Turnhout, Retie and Westerlo remain external operation-level holdouts. The
+  Tower export retained 198 tiles and 58,820 labels. Its small-object-aware
+  audit passed with no invalid/missing labels, and the 48-tile balanced visual
+  review contained no missing, invalid or low-variance selections. The trained
+  `geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt` candidate passed
+  seven positive-AOI and three pure-empty background gates at tile `512`,
+  overlap `64`, threshold `0.15` and QA match IoU `0.25`. Mean F1 is `0.5825`
+  and all pure-empty samples remain at zero detections. Persisted comparison
+  found 1,571 fewer false negatives than the previous balanced model, with a
+  lower mean precision and therefore a higher operator review load.
+
+Use `--samples` or `OPERATOR_YOLO_SAMPLES` to make an experimental corpus
+membership explicit. Dataset summaries preserve the complete manifest count,
+selected sample slugs and excluded sample slugs. Split validation still applies
+after filtering, so a manifest-backed holdout cannot be selected as training by
+omitting it from `--val-samples`.
 
 After rebuilding the all-in-one image, the operator scripts are available inside
 the container at `/app/scripts/...`. Before rebuilding, use the host checkout or
@@ -765,7 +784,10 @@ python scripts/audit_detection_false_negative_evidence.py \
 ```
 
 The audit reports false-negative rates and area buckets per AOI/model, plus
-reference buildings missed by every compared portfolio. Stable
+reference buildings missed by every compared portfolio. It writes the combined
+`persistent_false_negatives.geojson`, records geodetic persistent-miss area and
+adds persistent area buckets so operators can inspect the shared misses on a
+map instead of relying only on counts. Stable
 `source_feature_id` values are preferred; a normalized geometry fingerprint is
 used only when source IDs are absent. Invalid or missing geometry fails the
 audit instead of being silently skipped. The tools do not run inference,