Revert base-frame rotation in observation.state: restore raw simulator values.
Eval feeds the policy extract_state_from_proprio(obs['robot_r1::proprio']), which
slices base_qvel straight from live proprioception, so training data must carry the
same quantity. The 2026-08-10 rotation of dims 84,85/152,153/155,156/253,254 by
-base_qpos[2] created a train/inference mismatch and is undone here.
Verified: nav episodes match the source recovery HDF5 to 5.96e-08 over all 256 dims;
place episodes byte-identical to their 2025-challenge parquets. Per-episode stats
recomputed. Videos and actions unchanged.
