Tracked kernel package sources mirrored 1:1 from native-kernels/orbitquant-packed-matmul at 2d586f189ab8c44ca10685a64e71aaab10c61430. Adds the CPU backend (AVX2/AVX-512/NEON + stable torch ABI), the CUDA W4A4 int8 runtime ops with the cold-L2 pipelined mma64 path, and the Metal skinny-batch GEMV; no generated build files.
Source snapshot from OrbitQuant main commit 77c97cf. Includes tiled Metal packed matmul and benchmark reference comparison. No generated build artifacts, virtualenvs, bytecode, binaries, or benchmark outputs.
