Reductions refuse what they used to accept: GPU row-max non-finite guard, averageGrads length/NaN refusal, int32 accumulator bound, capacity-weighted cluster loss, finite reduced-gradient check, empty CUDA_VISIBLE_DEVICES

Refuse NaN/Inf at every int8 quantize; requant rounds half-up (retrained); bounded units; README for September

verified: bound GEMM memory, certify the multiply LUT, test the shipped paths

Release: SpikeWhale slider panel (HF dataset picker, stop/back), DaisyChain-Web (P2P WebRTC training, DaisyAdam, checkpoints, room host approval, verified-units-only)

Old-hardware training through emulated GPU logic