By: ValentinHirschi <valentin.hirschi@gmail.com>
stability test: count cut-constructible and rational parts separately cancellation_stable compared the result with the sum of whole-cut finite parts, so a cut whose cut-constructible and rational parts cancel passed (rank n+2 final review, finding 4: fr-light-zero-n6-r8 lane 2, generic f64 24% wrong, reported Success). The finite part's magnitude per cut is now |cut-constructible| + |rational|; the test only gets stricter. It also flags rn2-phys-real-n6-r8 lane 1 (generic f64 off by 4.4e-6). Results are unchanged; the test stays opt-in. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>