feat: sort S4 double-buffer scatter and mono-digit layer skip

Replace MSD cycle-permute with proven global scratch place/copy, and skip permute/scatter when a digit layer is a single full bucket.