f2adce45

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

Record the combined candidate's validation and its open test failure

STATUS.md gains a section on the combined candidate (the contour-box
regression fixes, the two-mass complements and the Symbolica pin on
9a898ee, regenerated once), checked against 09b2c74 and the three branch
heads (7f3193e, 19fdecd, 9c07ded):
- No interaction: on the D0 and A0-C0 sweep rows (seeds 1-4), the 843
  NinjaRS master and 7,644 suite calls, the contour branch's 3,387 and the
  two-mass branch's 128,462 rows, every binary64 route and DoubleFloat
  give the contour head's bits where the two-mass head leaves a row
  unchanged and the two-mass head's where the contour head does; no row is
  changed by both.
- Real 0-, 1- and 2-mass boxes are bitwise those of 09b2c74 on every
  binary64 route and in DoubleFloat; their 256-bit values move only with
  the pin.
- The 72 oracle-sweep failure lists (D0 and A0-C0, seeds 1-4, three
  routes, 128/256/512 bits) equal those of 09b2c74 and the branches.
- Instructions per D0 call on the 22 NinjaRS streams: geometric mean
  -0.12%, at most +0.29%, at least -0.63% against 09b2c74.
- NinjaRS master D0 calls changing bits against 09b2c74: 107/206/217/215
  of 309 (hand-written/generated/interpreted/SymJIT), worst error 1.9e-12
  to 5.2e-12; C0 unchanged in binary64 and DoubleFloat.

Two corrections from re-measuring on this build: the r14 = 1 box is right
at 128, 256 and 512 bits in all 24 relabellings on both Float routes, and
the three-mass box of the remaining Float rows is wrong at 128 and 192
bits in all 24 relabellings (12 of them where OneLOop is unstable), right
from 256 bits.

Open: with the pin, six relabellings of the near-shell two-mass sweep
regression track 216 of 256 bits but keep 207.7, 8.3 bits below what they
track, so the 256-bit check of
complex_two_mass_boxes_next_to_their_shells_match_the_references_at_128_and_256_bits
(8 bits of slack) fails in 12 checks; with the previous pin they kept
208.3. The slack is empirical and was not widened.

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