feat(abc): ABC-130k mcap->zarr pipeline + yam_bimanual embodiment

[AnikethCheluva]

Port of GaTech-RL2#606 onto the graph fork.

Adds the YAM two-arm teleoperation station behind ABC-130k as a first-class
embodiment (id 7) plus the converter, subset builder, uploader and metadata
backfill under egomimic/scripts/abc_process, with their sbatch launchers.

YAM is structurally close to Eva -- bimanual, parallel-jaw, per-arm
obs/cmd_ee_pose plus gripper -- with three differences that drive the code:
extrinsics come from the published station rig rather than the data (so
eef_frame, which needs none, is the mode that works on ZED-X episodes);
intrinsics are per-episode because two station types with different
calibration ship in the dataset, hence INTRINSICS = None; and the world
frame is station-anchored rather than egocentric, so there is no head pose or
keypoint stream.

A batch cannot mix image sizes and these datasets ship 640x480, 640x360,
1280x720 and 848x480, so image_hw pins every episode to one size. The
intrinsics are rescaled by the SAME per-axis factors -- row 0 with width, row 1
with height. Scaling by a single shared factor is what produced the anamorphic
K bug, and every projected overlay inherits it. Both helpers read the channel
layout rather than assuming HWC, because the decoder emits channels-first and
assuming otherwise turns each sample into a decode failure or a garbage scale
factor.

Graph integration: ActionTargetBuilder now takes action_key instead of
hardcoding actions. Planar names its chunk actions, but the Eva/Human/Yam
cartesian transform lists concatenate into actions_cartesian, and the graph
should read the key that exists rather than force embodiments to rename their
output. egomimic/hydra_configs/model/abc/yam_bimanual_dp.yaml is the
graph-native DP stage list for this data, and the ABC data configs declare
batch_keys for the keys their transform lists add -- tools/config_graph.py
derives seed keys from the key_map alone, so without that declaration it
reports the concatenated pose and action keys as unprovided. Both the train and
inference graphs lint clean.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01DcPCNAW3GKERANrqvfSCY1

[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

• feat(arc): arc tokenize and detokenize as graph nodes #11 : 4 dependent PRs (#12 , #63 , #68 and 1 other)
• feat(hpt): configured HPT graph for the ABC yam baseline #46
• feat(abc): ABC-130k mcap->zarr pipeline + yam_bimanual embodiment #10 👈 (View in Graphite)
• feat(hpt): Add Qwen stem and annotation prompt stage #53
• fix(data): actually honour bounds_check #45
• feat(flow): flow-matching stages for the HPT head #44
• feat(hpt): stems and trunk as graph stages #43
• feat(hpt): trunk primitives and tensor helpers #42
• feat(val): nested validation dataset groups #9
• refactor(transforms): factor rotation representation out of transform modes #8
• main

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