[aidang3019]

Stack context

Above #272 (aidan/arc-clock-single-stream-20261020), on Aniketh's #267 → #268 → #269 → #270 → #263. Fixes the must / should findings of a verified audit of #272; each was reproduced on real data first.

Must

• No min-speed floor on absolute clocks. Under streams: translation the translation clock also times rotation and the gripper. The 0.01 m/s floor (inherited from E1, carried in Train earbuds ARC absolute duration and native standard DP #266's decode) rewrote the decode of 720 of 2,113 real slow-pace ground-truth chunks: rotation replayed up to 10× faster than demonstrated, and the replan boundary shortened by up to 27 ticks. It was also discontinuous at zero distance. MAX_SPEED (5 m/s) stays as the motion bound; min_speed is removed. With the floor gone, the duration decode matches the velocity decode on 2,113 / 2,113 chunks.
• first_stream=0 replans by replan_every again. FirstStreamArcDecoder reported replan_steps = action_horizon in mode 0, so the robot executed the whole 100-tick chunk open-loop. The pre-arc: consolidate velocity/duration clocks and M×18/M×20 streams #268 behaviour (None → policy replan_every) is restored.

Should

• stationery_sp366e51_bc_time pins its episode count and name digest (417 train / 24 val), like the earbuds and towels pools, so a missing or deleted episode fails closed.
• frozen_yam_arc W&B ids are ${name}_s${seed}; a second seed no longer reuses the _s42 id, which resume: never rejects after the norm-stats pass.
• New pins: velocity / duration parity at 0.005 m/s; eval / robot parity on a noisy predicted token with garbage beyond the executed prefix (bounded by MAX_SPEED per tick); mode-0 replan; the time recipe's retimer speed controls (their tests were deleted in arc: consolidate velocity/duration clocks and M×18/M×20 streams #268); pool pins; the seeded id. The slow-parity and mode-0 pins fail on the arc: restore one-clock-per-arm streams; absolute duration with speed-bounded isotonic decode #272 code.

Validation

• Contract / recipe / retimer tests: 125 passed. 92-file CPU sweep: 1,374 passed, same 11 pre-existing failures + 1 collection error as arc: restore one-clock-per-arm streams; absolute duration with speed-bounded isotonic decode #272.
• Shipped-YAML audit 349 / 349. ruff check / format --check clean.
• No training launched; no cluster or robot process.

🤖 Generated with Claude Code

[aidang3019]

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: ARC Projection as a single-clock stream option #275
• robot: 20 N rollout gripper force limit with a live per-arm dashboard control #274
• arc: audit fixes — no min-speed floor on absolute clocks, mode-0 replan, pool pins, seeded W&B ids #273 👈 (View in Graphite)
• arc: restore one-clock-per-arm streams; absolute duration with speed-bounded isotonic decode #272
• Organize real and sim experiments and remove retired E1 surfaces #270 : 1 other dependent PR (#271 )
• configs: organize Cartesian experiments by task #269
• 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.