Add fair two-GPU Action Flow training

[ElmoPA]

No description provided.

[ElmoPA]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• Add decoded-noise scale ablation to UNITE Action Flow #62
• Add benchmarked no-checkpointing Action Flow variant #61
• Add fair two-GPU Action Flow training #60 👈 (View in Graphite)
• Prepare UNITE-backbone Action Flow recipe #54 : 1 other dependent PR (#64 )
• Add pure AdamW 200M Action Flow ablation #51 : 2 other dependent PRs (#58 , #59 )
• Scale Action Flow to 200M with Muon #48 : 2 other dependent PRs (#49 , #57 )
• Add 98k Action Flow reconstruction codec profile #47
• Port three torus candidate objectives to conditional U-Socket graphs #39 : 1 other dependent PR (#40 )
• Report checkpoint validation failures and support finite archive passes #34
• Add reconstruction-only Action Flow warmup #26
• codex/usocket-action-flow-bc-20260905

This stack of pull requests is managed by Graphite. Learn more about stacking.