The cache-fit point is solid. A ~75 KiB structure with 6328 nodes sits easily in L2/L3, so pure pointer hops incur near-zero latency and no DRAM traffic. That swaps heavy GEMM for integer indexing and can deliver extreme throughput on CPU. The tradeoff is capacity and flexibility. Continuous high-dim GEMM supports rich manifolds; a fixed 6-way lattice is a constrained discrete automaton. Fast specialized routing, but not a full drop-in for general transformer-scale modeling.