nx_nofloat_layernorm_gate.nx

nx_nofloat_layernorm_gate.nx -- the THIRD no-float transformer primitive: integer rsqrt -> LayerNorm. Same from-the-hardware-rung-up SYSTEM as the exp/softmax one: range-reduce + Newton iterate in PURE INTEGER -> bit-exact DETERMINISTIC. isqrt(v)=floor(sqrt v) via integer Newton (y=(x+v/x)/2 to a fixed point); fx_rsqrt = 1/sqrt via isqrt; LayerNorm = (x-mean)/sqrt(var+eps), all Q16 integer. Completes the trio every transformer layer needs: matmul (CAP-NOFLOAT-GEMM) + softmax (CAP-NOFLOAT-SOFTMAX) + norm. Float rsqrt/norm is non-deterministic; this is reproducible forever. Verified self-consistently: isqrt on perfect squares is EXACT; LayerNorm output has mean~0 + variance~1; fwd-sum == rev-sum (deterministic). No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL