nx_mathspec_pow.nx

nx_mathspec_pow.nx -- SOVEREIGN SPEC EMITTER for the POW kernel: every constant computed on the 120-bit bigfloat; written to _pm_pow_spec.nx for the MATH_KERNEL/POW emitter. No python, no minimax tables: the kernel's correction polynomials use EXACT-RATIONAL Taylor families L_j = 3/(2j+3) (log2 atanh-correction, |s| <= 0.101 -> 10 terms) P_k = 2*B_{2k}/(2k)! (exp R(z)=z(e^z+1)/(e^z-1) Bernoulli form, 8 terms) where B_2..B_16 = 1/6 -1/30 1/42 -1/30 5/66 -691/2730 7/6 -3617/510 (DLMF 24.2, mathematical definitions like 1/k!). (2k)! division runs as a loop of small divisors 2..2k (bf_div_small bound). The sqrt(3/2)/sqrt(3) anchor cuts come from 53-step bigfloat binary search on sig^2 vs 3*2^103 / 3*2^104. 32-bit hi/lo splits (dp, cp, lg2) mask AFTER rounding; remainder via bf_sub. Layout: c[0]=one c[1]=half c[2]=two c[3]=cut_sqrt15 c[4]=cut_sqrt3 c[5]=bp1 c[6]=dp_h c[7]=dp_l (log2(1.5) split) c[8]=cp_h c[9]=cp_l c[10]=cp(2/(3ln2)) c[11]=lg2 c[12]=lg2_h c[13]=lg2_l c[14..23]=L10..L1 (hi-first) c[24..31]=P8..P1 (hi-first, signed) c[32]=three c[33]=1025.0 c[34]=1076.0 license_tier: ORIGINAL