Data: GPU-CorruptNet head features and BatchNorm input

Exported by model/corruptnet_export.py from GPU-CorruptNet's fine-tuned ResNet-50, on the same seen-content (1,560) and unseen-content (2,600) test frames it reports on. Recomputing the head from these features gives macro-F1 0.9112 (seen) and 0.8756 (unseen), matching the published 0.911 / 0.876.

Contents: 2,048 pooled features per frame, fp32 reference logits, labels, head weight and bias, calibration temperature, and the input to layer4.2.bn3 for 32 frames (1,568 x 2,048, channels last) with that BatchNorm folded into a per-channel scale and shift.

sha256: 605de0de817389f9362d9c0821189816cfe4ef41c67062587129c26c535073dc