Calibrated checkpoint: tau via percentile_95 over full ID val set

Refresh model card with accurate shipped-checkpoint metrics and expanded usage

Upgrade to 150k-step longrun best: val_loss 0.0639 @ step 128250 (was 0.0726 @ step 21850)

Refresh model card with accurate shipped-checkpoint metrics and expanded usage

Replace with v2 run: 8L, Beta(1.5,1), transparent plots, val_loss 0.0726 @ step 21850

Upload onebox_8L.ckpt with huggingface_hub

Upload README.md with huggingface_hub

initial commit