Add compositional Muon CE-track leaders

[gabrielnan]

Summary

Adds two CE-track compositional Muon submissions:

• nanogpt_comp_muon_mp1: OV-only isotropic Compositional Muon, 2150 steps.
• nanogpt_comp_muon_mp4: OV isotropic Compositional Muon plus row-uniform Muon on the tall MLP expansion matrix, 1600 steps.

nanogpt_comp_muon_mp4 is now the lower-energy CE-track leader under the benchmark objective: pass native CE and accuracy gates, then rank by training energy.

Results

Submission	Total energy	Training energy	Val acc	Native CE	Status	GPU
nanogpt_comp_muon_mp4	42,059 J	34,235 J	0.7286	1.3182	pass	A100-SXM4-80GB
nanogpt_comp_muon_mp1	58,224 J	48,309 J	0.7344	1.2856	pass	A100-SXM4-80GB

Changes

• Adds nanogpt_comp_muon_mp1 submission, README, result.json, nvml.json, and run.log.
• Adds nanogpt_comp_muon_mp4 submission, README, result.json, nvml.json, and run.log.
• Updates the CE-track leaderboard wording to rank passing rows by training energy.
• Adds both rows to the CE-track leaderboard, with nanogpt_comp_muon_mp4 first by energy.

Energy Variance Check

We repeated the nanogpt_comp_muon_mp4 1600-step config with SEED=42 to sanity-check energy variance. Two full SXM4 repeats passed with total energies 48,013 J and 43,710 J; population CV was ~4.7% for total energy and ~5.1% for GPU energy. The original 42,059 J run is low but within observed host/power variance, and the improvement over the previous 58,224 J leader is much larger than the repeat variance.

Verification

• python3 -m py_compile submissions/nanogpt_comp_muon_mp1/submission.py
• python3 -m py_compile submissions/nanogpt_comp_muon_mp4/submission.py
• submit.precheck_submission(...) returns @gabrielnan for both submissions
• Modal benchmark runs completed successfully and wrote artifacts.