cagliostro-v4 Phase B checkpoints

Training checkpoints from cagliostro-v4 Phase B, a comparison of optimizers and architectures at 146M parameters and 2.95B tokens per run. Each arm lives on its own branch (B-c0, B-m1, B-m1L1) as one latest.pt with model and optimizer state. The file is overwritten as training goes and the branch history is squashed, so only the newest checkpoint is kept.

These are research checkpoints, not a release, and they are not in transformers format. Access is approved manually.