Muon optimizer (Nesterov momentum 0.95, update orthogonalized by 5 Newton-Schulz steps) for the 2-D weight matrices inside the blocks, peak learning rate MUON_LR on the same warmup+cosine schedule; embeddings, head and LayerNorms stay on AdamW. val_loss 1.532297 vs parent 1.560530 (2000 steps, seed 1337, eval as in train.py).
