A source-aligned CPU implementation core

This optional listing implements the two forward interfaces explicitly, preserves device and dtype consistency on CPU, registers structural masks and scars as persistent buffers, and supplies isolated evaluation streams.

It is a proposed replacement interface, not an executed regression. The source's default reset $h_0=0$ is retained. The actual supplied sequence length determines the number of steps.

The structural methods are exact-count, per-tensor operations. They do not claim equal per-gate allocation. Normal training must call enforce() after every optimizer step. Weights are stored at zero outside the mask, rather than multiplying an in-forward mask into autograd; ambient inactive derivatives therefore remain available to a separately declared regrowth scorer.