[AnikethCheluva]

Stack context

Adds a configuration-only follow-up above #268 (which follows #267 and #263).
No tokenizer/detokenizer Python implementation changes, cluster deployment,
training launch, or changes to existing checkpoint bindings.

What and why

Replace the scattered robot/cotrain/YAM task launch menu with:

experiment/
  organize_stationary/{baseline,arc_vel,arc_dur}.yaml
  fold_and_stack_towels/{baseline,arc_vel,arc_dur}.yaml
  wireless_earbuds/{baseline,arc_vel,arc_dur}.yaml
  multitask/{baseline,arc_vel,arc_dur}.yaml

• Share clock settings and training bases rather than copying them.
• Retain distinct HPT180, ABC-DiT, robot-only, ABC/ABC+RL2 and alternate frozen
  data-pool presets under task-local variants/; unrelated experiments remain.
• Replace duplicate four-/six-clock organize leaves with
  abc.arc_streams=translation_rotation|full_separation. Width follows
  automatically through data, model, evaluator and rollout (18 or 20).
• Fix shared visual data/evaluator configs that hard-coded velocity instead of
  following the selected clock mode.
• Add canonical-codec velocity/duration twins for frozen towels/earbuds, using
  their baseline graph, optimizer and episode lists, and shared open-loop eval.
  These are fresh-run recipes, not compatibility aliases for removed E1 codecs.
• Include a task/defaults guide, old-to-new path map, updated config auditing
  and regression checks. Remove unused retired E1 M/D modifiers.

Defaults and migration

Primary models are HPT300. Organize uses the existing Elmo robot+human pool;
multitask uses the existing four-task ABC pool, not the seven-task proposal.
Towels uses frozen since0924plus74; earbuds uses frozen wireless_earbuds_updated.
All three variants within each task preserve the same episode selection.

ARC budgets are explicit: organize M100/D0.42/R24°, multitask M100/D0.81/R24°,
frozen towels/earbuds M200/D0.33/R100°. This is not a new calibration.
Duration means log interval seconds plus first-row delay, not absolute seconds.

Old CLI recipe paths are intentionally removed. Recorded runs/checkpoints
retain their original source, configuration and normalizer. New ARC recipes
require fresh normalization and training.

Verification

• All 380 shipped Hydra YAMLs compose and resolve offline.
• 196 focused CPU tests passed; four full-size pretrained-model instantiation
  cases deselected. Includes all 16 task/clock/layout combinations, data
  contracts, inference/decoder construction, matching train/validation clocks,
  preserved task membership and a tiny ABC-DiT training/sampling check.
• Pre/post resolved snapshots confirm frozen towel/earbud baseline configs
  unchanged, including their training/evaluation settings and episode lists.
• Ruff and whitespace validation pass.
• No GPU throughput, full pretrained checkpoint or hardware test claimed.

[AnikethCheluva]

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

• arc: one clock per arm, absolute-duration clocks with speed-bounded isotonic decode #272 : 2 dependent PRs (#273 , #274 )
• Organize real and sim experiments and remove retired E1 surfaces #270
• configs: organize Cartesian experiments by task #269 👈 (View in Graphite)
• arc: consolidate velocity/duration clocks and M×18/M×20 streams #268
• arc: encode first-row start delays in non-E1 clocks #267
• Bring Aniketh's ARC work onto main with the YAM task recipes and the station rollout runtime #263 : 1 other dependent PR (#266 )
• main

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