Finite-Splitting Directional Attention MVP

Run:

/home/maxwelhelp/main/bin/python3 finite_splitting_attention.py

The script writes results.json and prints:

• exact geometric lacunarity-ratio verification;
• exact binary M-adic tree leaf/path split counts;
• dense 32-direction line responses;
• a proxy-guided top-2 sparse target approximation;
• random top-2 comparison;
• directional target-sample reduction and approximation error.

Important interpretation: the prototype uses a radius-1 line response as a routing proxy and a radius-2 response as the expensive target. The proxy currently evaluates all directions, so the 16x reduction is only in expensive target directional samples, not demonstrated end-to-end latency. In addition, with M=2, 32 distinct leaves require at least ceil(log2(32))=5 splitting vertices on a route; the requested N=3 and K=32 are incompatible for a full binary tree unless a larger branching factor or fewer leaves is used.