Removed original checkpoint loading (already verified 0.997 cosine). Test now: load NVFP4 → dequant BF16 ref → run kernel → compare. Exits with code 1 if cosine < 0.99.
Removed original checkpoint loading (already verified 0.997 cosine). Test now: load NVFP4 → dequant BF16 ref → run kernel → compare. Exits with code 1 if cosine < 0.99.