66bebf00

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

Record the final candidate's validation and its accepted exceptions

STATUS.md gains a section on the final candidate: what it adds to f2adce4
(PR #9, PR #10, the contour chart tests at a Gram determinant at the
rounding level, the one- and two-mass sheet repairs, the contour line
products and reflection, the soft-box relabelling and the test model), the
results against 09b2c74 and f2adce4 on the verifier's sets, the streams'
sweep and the tests' rows (every binary64 route and DoubleFloat), the
NinjaRS master calls in binary64 and at 256 bits, the gates, and the
regressions against 09b2c74 that were accepted on 2026-10-20, row by row
with their inputs, errors (now, 09b2c74, f2adce4) and causes: the route
rounding of the contour chart's arguments (v1 19098, v1 19320, CT 2168),
the CT 31-70 scan and g1 9239, NaN where 09b2c74 was already wrong by more
than 1e-6 (9 entries up to 0.77, 78 of O(1)), the rows within ten times
their one-ulp input conditioning, and the 128-bit Float rows.

The verifier's corrections of f2adce4's documentation: the row-by-row
sentence holds on the oracle-sweep rows only (the fresh sweep's counts are
given); the vanishing-Gram class with the fresh sweep's counts and maxima
after the fix; the 128-bit Float losses next to a vanishing Gram
determinant and the 0RRC collinear box from the pin; the on-shell contour
box CT 337 (NaN in 16 relabellings on every route, DoubleFloat and Float
included, right in 8); the Float 256 moves of the master calls ("almost
all calls, at most about 1e-70", with the measured counts); line 73 of the
near-threshold scan (1.3e-11 to 3.2e-10, now 1.35e-11); the nine
rounding-level Gram boxes (1.6e-11, one at 2.2e-9) and the scan's 199 rows
(3.9e-12) in PRECISION.md and README.md. README.md also describes the
one- and two-mass refinements and the soft-box relabelling, including that
symbolic uses of the D0 expressions must call regular_d0_labelling.

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