runDnnRemote() decoded the raw softmax*s field but skipped dnnGhostbust, so the untrained corner-velocity sidelobes (which the local CPU path zeros when curt_dnn_vmax>0) survived as background noise (~0.06 vs ~5e-4 on the MAX-all-v layer). Apply dnnGhostbust to the ROI field, mirroring the local path -> verified: ImageJ subtract of -HYPER-RECT MAX slices vs the CPU (17_UAS_REFACTORED) is exact zero. (Full-frame -OFFSET s is not yet ghostbusted - Java lacks the full 121-field; a DGX-side decode ghostbuster would cover that.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>