nx_bigfloat120_exp.nx

nx_bigfloat120_exp.nx -- SOVEREIGN exp oracle: binary64-in, binary64-out, all internals 120-bit bigfloat, EVERY constant self-computed (no mpmath, no python): ln2 = 2 * atanh(1/3) = 2 * (1/3 + (1/3)^3/3 + (1/3)^5/5 + ...) [39 odd terms] Range reduction k = nearest(x / ln2) is taken in bigfloat; r = |x - k*ln2| with the sign handled by the reciprocal trick (exp(-r) = 1/exp(r)) so the Taylor series sums positive terms only. 30 terms: r <= ln2/2 gives a 2^-122 tail. license_tier: ORIGINAL