Specialise by sequence length: single-token execution is at the memory-bandwidth ceiling, multi-token is compute-bound. Fuse the per-layer algebra at build time - combine projections sharing an input, fold the norm scale into the next projection, fold the attention scale into the queries, and pre-compose value and output projections at length one. Carry activations transposed so the weight is the streamed operand. The largest remaining gains are a hand-written vector kernel with pre-packed weights, a persistent pool pinned one thread per physical core, and cutting bytes per weight on the bandwidth-bound path.