Prepare EgoScale Stage 1 DLC full training

[Apexcmgd]

Summary

• publish the finalized self-collected Stage 2 data-loading contract
• add the six-dataset egoscale_stage1_ego_cartesian_clean_rl2 normalization and unified-action metadata
• add a non-destructive Stage 1 DLC production wrapper with explicit 100k-step training defaults
• add preflight-only support and update the staged-training runbook

Why

A fresh DLC checkout previously lacked the Stage 1 RL2 norm assets. The generic stage launcher also defaults to smoke settings, including 20 training steps and save/eval every 10 steps, which is unsafe for a full run unless every production value is overridden manually.

Validation

• bash -n scripts/run_egoscale_stage.sh
• bash -n scripts/run_egoscale_stage1_dlc_full.sh
• 85 passed for tests/cotrain/test_action_space.py and tests/cotrain/test_rlds_dataset.py
• clean DSW worktree at 26a13587a3a055183f6ce71883492513e2f68e2a
• Stage 1 preflight passed for all six EgoVerse/EgoVerse-RL2 builders, all norm/mapping metadata, and /data/junhe/models/pi05_base/params
• all 13 imported asset files match the verified DSW copies by SHA256; norm values are finite

Operational notes

No DLC GPU job was started by this PR. The runbook requires a separate 2-node x 8-GPU, 100-step checkpoint-and-resume validation before the 100k run. W&B credentials remain external and must be injected through DLC secrets.