nx_bigfloat120.nx

nx_bigfloat120.nx -- ME3 rung 1: SOVEREIGN fixed-120-bit binary float, pure Nishi. The in-house high-precision oracle that retires the last python/mpmath debt in the math-engine arc: 120 significand bits = 67 guard bits beyond binary64, far past what correctly-rounded f64 transcendental reference values need. Rung 2 (arbitrary limbs on bigint.nx) extends this; the API is chosen so that swap is additive. Representation (struct-free, mmap blocks of 3 i64): b[0] = e -- unbiased exponent; value = (hi*2^60 + lo) * 2^(e - 119) b[1] = hi -- significand bits 119..60, normalized: hi in [2^59, 2^60) b[2] = lo -- significand bits 59..0 Values are POSITIVE-ONLY (plus a zero sentinel e = BF_ZE); callers track signs at the boundaries (series with all-positive terms by construction: exp uses the reciprocal trick for negative r, atanh is odd so |s| drives the series). This keeps the core branch-light and audit-friendly. Internal sub-ulp bits are truncated (no sticky): each op is exact to 2^-119 relative; the transcendental pipelines run < 100 ops, so accumulated error stays below 2^-110 -- 57 bits of margin over the half-ulp-of-f64 decision boundary. Division lives in nx_bigfloat120_div.nx (same-file mul+div codegen quirk law). license_tier: ORIGINAL