v8: yolo11s+353 validator-pseudo frames (mAP50 0.84 on 73-img holdout)

5x bigger train set (353 unique validator frames vs 70 in v7). Same yolo11s base, fp16 ONNX, same miner.py with auto-detect dtype.

best.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8beaa959e6efb3b0f82ac0876fe14bfe1d8d5dfaeecb39ca31d3430473bd3ed9
+oid sha256:5582c14a3771ab2bc44d9a66e0fe50620f98c1cac06b4ba766002714410a1054
 size 19271662