b39241fc

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

Record a four-mass D0 real-part defect with an ignored reproducer

Fifteen real-mass four-mass D0 rows from the fast-engine bench cases
(claude/engine-alt 82596cf) have a wrong real part and a correct imaginary
part in every route of this crate: the symbolic expression at 256 bits,
SymJIT, native f64 and native 256-bit Float. OneLOop 3.6 and COLLIER (COLI
and DD) agree on all of them; OneLOop 3.4 agrees except on one row, which
this crate gets right. All failing rows have s12 * s23 < 0, which is not
sufficient on its own. The previous main gives the same values, so the defect
is in the D0 expression rather than the binary64 primitives.

Add the rows with OneLOop 3.6 references as an ignored test, and note the
defect in the README status and the 2026-10-17 performance report.

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