nx_nofloat_attn.nx

nx_nofloat_attn.nx -- SOVEREIGN NO-FLOAT ATTENTION (the transformer HEART), integer Q16, deterministic. The hard part of attention is softmax (needs exp); this builds a fixed-point exp via 2^x decomposition (|x|*log2e -> integer shift + quadratic 2^frac), then softmax (max-subtract for stability -> exp -> normalize) + attention (weighted sum of V). KEY: even with an APPROXIMATE exp, softmax NORMALIZES (each / sum), so the weights sum to ONE exactly and monotonicity holds -- all integer => bit-exact deterministic (the moat vs float/CUDA attention). T1 weights non-negative + sum ~= ONE. T2 monotonic (highest logit -> highest weight, order preserved). T3 (EXCEED) determinism bit-identical. T4 attention output = convex combo of V, dominated by the max-score key. expect_exit: 0 Sovereign: nx_syscalls.