Function compile_to_cuda_matmul 

pub fn compile_to_cuda_matmul(
    source: &str,
) -> Result<(MatmulSpec, String), String>

Genuinely LOWER an ordinary LOGOS matmul program to an optimized CUDA GEMM kernel — the CUDA analog of compile_to_tenstorrent_matmul. Recognizes the GEMM idiom with the SAME backend-agnostic crate::tensor_recognize::recognize_matmul, extracts (M,N,K) from the source, and emits a CUDA program whose dims are the program’s (a function of the source, not a canned template); a non-matmul is REFUSED (typed). Returns the recognized (spec, kernel_cu); the emitted program’s checksum equals the VM’s Show byte-for-byte.