662d9069

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

cli: derive Float S-matrices exactly from the prepared momenta

Under --precision float the generic engine differenced rounded momentum
components to obtain master-integral invariants. Exactly light-like legs
with non-dyadic components (the four-photon pool's outgoing photons) then
reached the masters as tracked roundoff, e.g. p^2 = +2^-114 with 2 bits at
128 bits. Float applies no on-shell threshold, and oneloopmaster's
four-mass D0 evaluated that input orders of magnitude off while reporting
8 bits, so lane 0 of top-box-0-pool returned -4.91 instead of -8.26e-4 at
128 bits (and similarly at 104, 116, 140, ... 260 and 400 bits for lanes 0
and 1).

NumericalPool now derives every omitted S-matrix exactly from the prepared
momentum expressions for carriers without fixed precision, as it already
did for batches mixing supplied and omitted matrices. Fixed-precision
carriers keep Ninja's momentum differences and OneLOop threshold.

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