By: ValentinHirschi <valentin.hirschi@gmail.com>
Record binary64 native and tensor timings against OneLOop Fortran Add performance/2026-10-17-claude: the gg -> HHH top-loop pentagon (graph26: 5 D0, 10 C0, 10 rank-two bubbles and 5 A0 per point) takes 126.8-128.0 us per point natively against 103.4-104.9 us for OneLOop (2024 snapshot at -O2/-O3, 3.6 at -O3), a 1.21-1.24 ratio; mixed scalar fixtures are 0.70-1.20x; B11/B111/An are 0.27-0.51x of olo_b11/olo_bn/ olo_an. The previous main took 1.7 ms per mixed C0 and 8.2 ms per mixed D0 because binary64 dilogarithms ran MPFR. The 2024 OneLOop snapshot returns a wrong real part for one graph26 box (OneLOop 3.4/3.6 and COLLIER agree with this crate). Raw TSV/JSON, row inputs and comparison tables are included. Document the binary64 reciprocal/norm fast paths and the report, and apply rustfmt to the files added or changed on this branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>