Previous onnx/coreml exports never loaded resnet50_MixVPR_4096_channels(1024)_rows(4).ckpt — they shipped an ImageNet backbone with a randomly initialized aggregator, and the CosSim checks passed because the PyTorch reference was the same unloaded model. Symptom: same-image 90%-crop cosine ~0.48 (should be ~0.77); indoor relocalization recall@2m 59.5% -> 94.8% after the fix.
Export scripts now load the ckpt inside the model constructor with a strict key check and refuse to run without it.
