Add official-protocol RoboCasa365 GR00T evaluation

[MilkClouds]

Summary

• add an opt-in rc365 protocol for the official target50 registry, pretraining-kitchen split, per-task horizons, 16-step success checks, and lossless 12-D Panda-Omron actions
• preserve the original arbitrary-task, configurable-camera, 7-D adapter and keep eval.yaml byte-for-byte unchanged
• add full and smoke RC365 configs plus a seeded GR00T N1.5 server
• pin the exercised RoboCasa and robosuite patch revisions while checking their semantic package versions

Scope

This PR changes only the RoboCasa adapter/config/docs, the GR00T server/config, and the RoboCasa image.
It contains no simulator-bound test, generated report, generic result plumbing, or checked-in parity experiment.

Dependency policy

• the PEP 723 GR00T script uses exclude-newer = "2026-07-19T00:00:00Z"
• the pinned fork's bundled Eagle model imports flash_attn directly, so Transformers kernels cannot replace it without changing upstream model code
• the exercised runtime therefore pins torch==2.7.0, transformers==4.51.3, flash-attn==2.7.4.post1, and ninja==1.13.0, with no build isolation only for FlashAttention
• RoboCasa 1.0.1 at b4684e6e and robosuite 1.5.2 at 5ce6643f are checked after installing the tested patch revisions

Validation

• current-head H100 direct-wrapper parity passes for KettleBoiling (pretrain, seed 0) and ArrangeBreadBasket (pretrain, seed 3)
• reset state, one-step state, all three RGB observations, five state observations, array metadata, and action order match exactly; every measured difference is 0
• current-head success-protocol parity against the pinned GR00T MultiStepWrapper passes transient-mid-chunk, chunk-endpoint, and final-partial-chunk cases
• legacy split=None and split="all" remain accepted, while rc365 remains restricted to pretrain and target
• local validation passes Ruff, formatting, ty, and pytest (441 passed, 2 skipped)
• all six GitHub checks pass

[MilkClouds]

@codex review

[MilkClouds]

Pushed the RC365 S2B integration onto this branch (a0ef7e3), ported to the refactored benchmark API:

• CPU rendering: VLA_EVAL_RENDER=cpu routes MuJoCo through OSMesa (Mesa llvmpipe cannot serve MuJoCo's EGL device-platform requirement). Verified in the charliecloud-converted container with no GPU: all three RC365 cameras render. Motivation: CUDA compute and EGL readback on the same GPU abort nondeterministically in the NVIDIA driver (controlled A/B: 6/6 stable split vs 4/4 abort co-tenant), so CPU rendering restores one-GPU rollouts.
• Hierarchical policy server (model_servers/rc365_s2b.py): the S2B System-2-plus-System-1 stack as one stateful policy; System 2 re-queried every 16-step chunk (gold sequence, global-only, random-valid, or MLLM planner).
• Adapter parity extended from Kettle to the six qualification tasks (recorded 32-step action comparison, committed report, pass).
• Qualification runner + Slurm template producing scorer-compatible dumps.

Full suite: 450 passed, 2 skipped. Follow-up work (GPU policy inference through the runner at scale) tracked on our side.

Pushed by Claude Fable 5 on behalf of @MilkClouds.