feat(hpt): configured HPT graph for the ABC yam baseline

[AnikethCheluva]

Wires the stems, trunk and flow head into a runnable experiment, so this stack
now delivers the actual HPT architecture rather than a DP model matched to it:
per-modality stems with cross-attention latent pooling, a 16-block shared
trunk, and flow matching.

The stage list is the dataflow, which is the payoff of decomposing:

observations --HPTStemStage--> hpt/tokens --HPTTrunkStage--> condition
actions --ActionTargetBuilder--> target --FlowNoisingStage--> flow/*
flow/* + condition --FlowDenoiserStage--> flow/predicted_velocity
--FlowVelocityLossStage--> loss/*

Four modalities get stems: the 14-D pose through an MLP, and the front plus two
wrist cameras through ResNet encoders. Each pools to 16 latents, so the trunk
sees 64 tokens plus its action token.

hpt.embed_dim is a single shared bus width read by every stem output, every
cross_attn modality_embed_dim, the trunk and the head's cond_dim. A test walks
all of those rather than trusting the interpolation, because a mismatch is the
most likely way this config rots and the stem stage only catches it at the
first batch.

The head uses CrossTransformer, as upstream does, which needs the token axis
the trunk's pooled vector lacks -- hence condition_as_tokens.

Also adds experiment/abc/yam_fstshirt_hpt as the architecture counterpart to
yam_fstshirt_dp: same data, same action space, same dataloader params, so a
delta between them is attributable to the model. A test asserts the two stay
distinct and keep sharing their data config.

Verified: composes, instantiates at 55.3M params, lints clean in train and
inference, and a real batch trains with gradient reaching all 564 parameter
tensors, then samples a (B, 100, 14) chunk at inference. Full suite 581 passed.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com

[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 👈 (View in Graphite)
• feat(abc): ABC-130k mcap->zarr pipeline + yam_bimanual embodiment #10
• 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.