[aidang3019]

On main (a525d738). Aniketh's PR #259 work reconciled onto main as one co-authored commit (5666d7a8), then Aidan's stationery / earbuds / towels recipes and the station rollout runtime, so the three YAM task runs train and deploy from one checkout. Head 97b48673: 272 files, +40,511 / −5,524.

Pruned on 2026-10-18 at Aidan's request: the recipe tree keeps only the _base/ recipes and one leaf set per task at M 200 waypoints / D 0.33 m (plus the M/D-free time leaves), and the checkpoint-era tokenizer snapshots (#230's robot/arc_codecs/, the station's pinned_arc_bbb813a0/) are not carried.

Review in four groups

1. Aniketh's #259 onto main (5666d7a8, one commit, co-authored). His 60 commits were on the 2026-09-24 main, 26 merged PRs behind; this commit is his content adapted to main's data-context architecture: data/abc_arc recipes move to ZarrDataModule + prepare_context, the ABC ARC export guard becomes model_contract/abc_visual.yaml + inference / data_requirements declarations on his models (stage_ids.sampler 4 for HPT, 3 for DiT), open_loop_sim gains main's evaluation hooks and the history-stack rule, MultiDataset.to_state keeps contract-free digests, ABC-DiT skips external weights under restoring_parameters(). Multistream / six-clock / log-clock evaluator and tokenizer, ABC-DiT, ResNet-MLP stems and his experiment/abc_arc recipes are otherwise his. @aniketh: #259 can take this commit as its rebase onto main.

2. Recipes. _base/stattempo (FT recipe on the frozen slow-pace 366 pool), _base/stattempo_openloop, _base/arc_projection, _base/m200_d033, _base/towels_hpt300 (HPT-300M, tempo evaluator), _base/earbuds_updated_hpt300, _base/m200_d033_hpt300; the shared loader data/abc_arc/stationery_yam_loader and the three frozen pools (slow-pace 366, earbuds_updated, towels since0924plus74 / new290plus74) with their manifests. Leaves:

• stationery: bc_rl2_stattempo_slowpace366_{arcdur,arcvel,arcdur_abstime,arcvel_nosmooth,arc_projection_dur,arc_projection_vel}_m200_d033
• earbuds: earbuds_updated_time_hpt300, earbuds_updated_{arcdur,arcvel,arcdur_abstime,arcvel_nosmooth}_m200_d033_hpt300
• towels: towels_{since0924plus74,new290plus74}_time_hpt300, towels_{since0924plus74,new290plus74}_{arcdur,arcvel}_m200_d033_hpt300

Every task launches through scripts/e1/stationery_ft.sbatch with EXPERIMENT=yam_arc_grid/<leaf>; a relaunch changes only logger.wandb.id (docs/STATTEMPO.md). The six stationery leaves resolve byte-for-byte to the configs authored on 10-17 (tests/fixtures/stattempo_contracts.json pins them at 8e80f090). The hybrid / tri / event / reference, r72 / r360, tempo-bucket, Elmo+Aidan, Aria-cotrain and M 100 recipes are gone from the tree; their codecs, decoders and evaluator modes stay, and scripts/e1/build_slowpace366.py regenerates them from the manifest if a run needs them.

3. Rollout runtime (da1429cd … fa072050):

• Elmo's Consolidate YAM rollout recording, active timer, and ARC controls #230, cherry-picked with his authorship: declared decoder_attribute controls and transactional apply_control_bindings, execution_prefix, replay / hold tempo, fastest-stream chunk termination with its waypoint cap, HDF5 episode recording, the rollout timer, station yaml without inline inference_graph. Its vendored codec copies are removed again in group 4.
• The camera alignment monitor (plus the station's direct-D405 --source realsense), the HDF5 tempo checker, the stalled-episode / max_steps fixes.
• The 10-12 IK branch: opt-in pose relaxation, bounded differential IK recovery, TRAC-IK endpoint selector (yam-tracik extra), --check-model CPU preflight, towels / earbuds HPT300 rollout profiles.

4. Decoding the stack's own tokens + station extras:

• BimanualArcDecoder decodes every layout these recipes train with the live egomimic.rldb.zarr tokenizers (absolute timing, hybrid R, event / reference / projection, speed floors) with the same tempo and termination controls; the live E1 codec gains decode-only helpers (detokenize(times=), as_duration_token, stream_clocks, stream_ends), tokenize untouched. FirstStreamArcDecoder executes the live multistream interval tokens to their first moving stream. The E1 contracts declare first_stream: 1, execute_percent: 50 and the four controls (model.arc_decoder_controls); the time contract wraps its chunk in a TimeChunkRetimer for the naive speed-up. No checkpoint-era tokenizer copies: a bundle trained by another tokenizer revision needs that revision's runtime.
• Station-only extra: view-only reset guides on both dashboards (reset_overlay_templates.yaml, egomimic.scripts.build_reset_guides).
• docs/YAM_RUNTIME.md documents the decoder, the declared controls, the IK options, the monitors and the reset guides.

Decisions on the open questions from the #229 handoff

1. Legacy unbound checkpoints stay unsupported (main's loader); the shared profile defaults to the bound towels HPT300 180k bundle, YAM_HPTFLOW_* env variables select another.
2. No vendored codecs: tokens decode with the live tokenizer; older bundles keep their own runtime.
3. No encode-side timing change is ported.
4. Decoder controls are declared decoder_attribute targets; fastest-stream is on by default in the E1 contracts (what the station ran), the dashboard turns it off.
5. Prompt follows the default bundle (fold towels); the per-task profiles pin their own.

Not carried on purpose: the overlapped-inference loop (tried 10-08, reverted on the arms), the station's hash-bound pinned_arc_bbb813a0 routing and the live-tree BimanualArcClockDecoder (both depend on byte-exact tokenizer snapshots), and the station's uncommitted hpt_stems / teleop edits (main already has them).

Verification (WSL CPU venv)

• ruff 0.8.6 check + format clean (main's gate).
• scripts/audit_hydra_configs.py: 415 / 415 YAMLs compose and resolve; all 17 task leaves and 7 bases resolve to their variant, M, D, W&B id and declared decoder.
• scripts/e1/validate_stattempo_configs.py: PASS, the 6 stationery recipes match 8e80f090.
• Recipe / decoder / rollout / contract suites: 606 passed, 15 skipped; the 6 remaining failures (MuJoCo / mink missing, no torchvision write_video, W&B-resume and ABC layout tests) are identical on the pre-rollout tag.
• Not done: no training smoke, no checkpoint reload, nothing on the arms. First station use: fetch into a new worktree, --check-config then --check-model on the towels profile, then a rollout as the validation of the decoder path.

🤖 Generated with Claude Code