-gm-pm1: exact P-1 Stage 1 and low-memory product-exponent Stage 2 forG_p = 2^p - (2/p)2^((p+1)/2) + 1.- Structural
4pmultiplier in the Stage 1 exponent. -gm-ecm: Suyama Montgomery ECM with GPU projective x/z ladders.- CPU projection, GCD and inversion strictly modulo
G_pat stage/chunk boundaries. -gm-factor-chunk-bitsfor P-1 and ECM Stage 2 batching.- Atomic CRC32 Stage 1 and Stage 2 checkpoints.
-gm-safeindependent P-1 block/chunk replay and ECM Stage 1/chunk replay.- Deterministic mathematical tests and four small GPU validation cases.
The new implementation is contained in
src/modes/RunGaussianMersenneFactor.cpp plus CLI/dispatch declarations. No
file under third_party/aevum, kernels/marin.cl or kernels/prmers.cl was
modified relative to the working v99.87 Gaussian-Mersenne bundle.
Stage 2 is a safe product-exponent continuation rather than the ordinary
Mersenne V-trace/BSGS implementation. Those implementations assume arithmetic
and inversions directly modulo the tested Mersenne number; applying them
unchanged to 2^(4p)-1 could be obstructed by the unrelated lift cofactor.