[RyanPCo]

Makes the Aria MANO fit, the dominant cost of the raw Aria migration, 6.5x faster without
changing its output. On one 4,587-frame episode (2,400 iters, A40) MANO goes from 338 s to
52 s for both hands, and the keypoints match the previous fit within 1e-3 mm.

• Batching (aria_to_lance --mano-batch, default 8192): the fit loss was normalized by the
  whole batch's valid-point count, so the batch size changed each frame's gradient scale
  and, through Adam's eps, the result. The loss is now normalized per 512-frame chunk
  (norm_group), and up to --mano-batch frames are optimized together. Same results as
  fitting each 512-frame chunk alone (~2e-4 mm); ~2.1x faster on its own.
• Fingertip-only forward (aria_utils.mano_joints_fn): the loss uses the 16 kinematic joints
  and 5 fingertip vertices, but the stock MANO forward skins all 778 vertices every step.
  mano_joints_fn regresses the joints from the shaped template and skins only the 5 tip
  vertices, giving the same 21 points.
• docs/zarr_to_lance_migration.md: the raw-path cost section uses the new numbers (~40 A40
  GPU-h for MANO over the 2,569 Aria episodes, down from ~235).

Testing: on a compute node, tests/unit/test_mano_fit.py (new: mano_joints_fn equals the full
forward's joints for both hands; the batched fit with per-chunk normalization equals
fitting each chunk alone) passes with the MANO weights linked in (3 passed); without them it skips. Full unit suite at the top of the stack: 1,565 passed, 5 skipped (3 of the skips are test_mano_fit without weights in the checkout; it passes with them linked in).

🤖 Generated with Claude Code

https://claude.ai/code/session_01HjCQBqQCr2igFEMR7yUovw

[RyanPCo]

Automated adversarial review (read-only; findings were checked by a second skeptical pass before posting): 1 nit.

[RyanPCo]

nit (naming): Reusing zeros as the base for four index_add calls is correct (out-of-place), but readers have to check that. Use a small group_sum = lambda x: torch.zeros(n_groups, device=device).index_add(0, gid, x) for group_denom, group_count, pos_loss and reg_loss.

[RyanPCo]

Review follow-up (force-pushed). The 1 finding is fixed:

• fit_mano_to_aria_batched uses a group_sum(x) helper (fresh zeros + index_add) for denom, count, pos_loss and reg_loss instead of reusing a shared zeros base.

https://claude.ai/code/session_01HjCQBqQCr2igFEMR7yUovw