Reconcile the sample catalog with deployed truth; add a **kwargs escape hatch to every setup hook

[OwenPriceSkelly]

Upstreams everything the deployed clusters are running that the repo catalog didn't have yet (sources: the staging/ snapshots regenerated from the admin-dashboard manifests, 2026-08-04/06), plus one convention change.

Catalog reconciliation

• Prune dead configs — the seven configs dropped from every deployed set on 2026-07-30: the fairchem-core 1.x OC20 baselines beyond escn/equiformer (dimenet, gemnet, painn, schnet, scn) and the BYO-weights-only envs (allegro, nequip — <family>:custom is the supported path for user weights now). README tree and modal_app probes updated to match.
• fairchem v2 indexed-device fix — _fairchem_device() routes cuda:N through torch.cuda.set_device() because MLIPPredictUnit._setup_device asserts device in ["cpu", "cuda"] and resolves the GPU from torch state. Indexed devices are exactly what multi-checkpoint verification hands workers on a multi-GPU node; this killed all 8 fairchem-v2 checkpoints on the 2026-08-06 Polaris sync (4x A100, VERIFY_JOBS=4). Hotfixed on the ALCF install since 2026-08-06; applied here to nvidia + amd allscaip/esen/uma (ROCm torch exposes the same torch.cuda API).
• esen header — drop the fairchem-1.x pyg find-links dance (fairchem-core>=2.20 doesn't need it, and the find-links block universal lockfile resolution).
• uma checkpoint table — document the uma-s-1p2 bug and the not-yet-released uma-s-1p2p1 replacement (re-add when the next fairchem-core ships).
• orb v2/v3 — orb.py (both vendors) becomes the v2 remnant it is on every cluster: trimmed to orb-d3-v2 (no v3 dispersion model exists), orb-v2:custom, and orb-models>=0.5,<0.6 (0.4.x dies at import on fresh builds — git-only undeclared pynanoflann; 0.6 needs Python 3.12 and belongs to orb_v3). orb_v3.py picks up the deployed <3.13 cap (dm-tree pin has no cp313 wheel), drops the omol ids (failing verify everywhere since 2026-05), and gains orb-v3:custom + setup_from_path(arch=, precision=).
• New: OrbMol env (ROCm) — amd_configs/orbmol.py on orb-models>=0.7,<0.8: orbmol-v1-conservative (upstream alias for orb-v3-conservative-omol), orbmol-v2, orbmol:custom (arch= kwarg). Wraps ORBCalculator to default atoms.info charge=0/spin=1 (the models raise without them), defaults edge_method="knn_scipy" on ROCm (0.7's knn_alchemi default launches CUDA Warp kernels and ROCm torch self-reports as cuda), torch pinned to the rocm6.4 index. Not yet deployed — ships to Frontier on the next deploy.

**kwargs escape hatch

Every setup()/setup_from_path() now forwards **kwargs to the calculator constructor, so setup_kwargs= / --kwarg can override constructor knobs the config doesn't name. Config-chosen defaults (default_dtype etc.) move into kwargs.setdefault() so user overrides don't collide with named arguments. A new contract test in tests/sample_configs/test_setup_from_path_hooks.py pins the convention; the rootstock create skeleton and docs/environments.md document it.

Deliberately not included: Delta's tace cu128 torch-index pin (cluster-specific — its driver caps at CUDA 12.8) and the polaris-only ani/pet variants (cluster-specific by construction).

Tests: ruff check, ruff format --check, ty check, and the full pytest suite (875 passed) are green.

🤖 Generated with Claude Code