59b96e59

By: ValentinHirschi <valentin.hirschi@gmail.com>

Record the final candidate's Float runs on the streams' and two-mass rows

STATUS.md, final candidate: Float at 128 and 256 bits on the streams' sweep
(37,642 rows) and the two-mass corpus (128,462 rows), native and
interpreted, against 09b2c74 and f2adce4. TM: no row worse. SW: none worse
at 256 bits, where 1,184 (native) and 1,180 (interpreted) values that
09b2c74 returned non-finite are now right (the soft-box relabelling); at
128 bits four relabellings of a box with a mass one ulp above two
invariants lose 3.3e-12 (native) and 1.2e-12 (interpreted) where 09b2c74
kept 7.8e-17, within ten times their one-ulp conditioning (18.6), as at
f2adce4: they join the J rows. Also recorded: the rows still wrong as at
09b2c74 (the largest errors of the sets, among them a streams' box whose
finite part is exactly 0 on every route for COLLIER's 496 + 370i), v1
14736's generated route back at 09b2c74's 0.65 (f2adce4 had it right),
the C0 SymJIT asset that PR #9's sheet change leaves stale (binary64 C0
values of the NinjaRS master calls unchanged on every route) and the
debugging switches ZZ_GEN_LOG and ZZ_GEN_DD in src/sheet_exact.rs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>