nx_alu_divider_newton.nx

nx_alu_divider_newton.nx -- MULTIPLICATIVE (Newton-Raphson reciprocal) divider: the ~log(W) latency FRONTIER the radix-2/radix-4 dividers are LAPPED by (~5x). q = floor(N/D) via a fixed-point reciprocal R ~ F/D (F = 2^k) refined by Newton's iteration x_{i+1} = x_i*(2F - D*x_i) / F , which DOUBLES the correct bits per step (quadratic convergence) -> ceil(log2 W) iterations, vs W (radix-2) or W/2 (radix-4) compare-subtract stages. Each iteration is 2 multiplies, so full-width (64b) it composes the TRIANGULATED nx_mul_wide (each product needs >64 bits -- which is exactly why the wide multiplier was the prerequisite). This module PROVES the math + the log(W) iteration count at a provable width (W=16, products fit i64), TRIANGULATED by three independent methods that must agree: (A) Newton, (B) an independent restoring shift-subtract divider, (C) the i64 oracle `/`. It also measures the post-Newton correction so the latency is honestly O(iterations)+O(1), not a hidden long loop. license_tier: ORIGINAL