[snimu]

Turns on Nesterov momentum for Muon by default.

What changes

• MuonConfig gets nesterov: bool = True. It is passed to dion's Muon, whose own default is False.
• This changes the default for every SFT and RL run that uses Muon. Set optim.nesterov = false to get the old behavior.
• Resuming works: the momentum buffers are the same with and without Nesterov; only the update read from them differs.

Why

• In our pretraining runs (MoE, d=384, Muon), Nesterov lowered the final loss by 0.0128 ± 0.0017 nats, paired over 4 seeds and several learning rates. These are internal, unpublished experiments.
• Keller Jordan's Muon and Moonlight use Nesterov momentum.
• dion's own reference implementation (muon_reference) defaults to nesterov=True.

The effect on SFT and RL fine-tuning has not been tested.

Validation

• Ruff passes. The config and SFT unit tests pass (366).
• SFT smoke run with Muon on 1 GPU (configs/debug/fake/sft.toml, Qwen3-0.6B, lr 1e-3, 20 steps, H200): trains without errors, nesterov = true in the resolved config, loss 8.46 at step 1 and 6.61 at step 20 (noisy on the fake data: 4.80 at step 12, 8.37 at step 14).