Fix: replace flawed yolo11m-seg@640 checkpoint (89 FP/80% recall at real inference settings) with a clean retrain under corrected checkpoint-selection code and corrected val labels. See inference_config.json for the calibrated conf/iou/max_det this checkpoint needs.

models/yolo/best.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:799a8d8917ad2388105d70397e4026b75bc5a40a6e700908118b06d35f1fdb02
+oid sha256:4b13308f937444558906eff96dd3f34c17ec2a2e05493f6c8bde36cabdaa820f
 size 45299375