Percepta-Core/transformer-vm

Compile programs directly into transformer weights. Includes a 2D convex-hull KV cache with O(log n) inference.