Add deepcell types 2026-06-23 weights to registry

[rossbar]

Okay here's the replacement for #5 !

@xuefei-wang can you confirm that the hash for this set of weights is the right one (i.e. the listed hash matches the md5sum of the weights that should be uploaded)

[xuefei-wang]

Heads-up @rossbar — I pushed two commits directly onto this PR's branch (add-dct-2026-06-23), to keep this in sync with the deepcell-types package release. Summary of what changed and why:

1. Serve the vocab-bundled 2026 checkpoints; default to 2026-06-15

• 2026-06-15 → deepcell-types_2026-06-15_resmlp.pt (b819a7e0…) — paper headline
• 2026-06-23 → deepcell-types_2026-06-23_resmlp_ptft.pt (402e94c1…) — SSL/pretrain-finetune arm

The original deepcell-types_2026-06-23.pt (067f558b…) is the un-bundled file: it lacks the bundled ct2idx/canonical_channels, so the deepcell-types package's predict() vocab guard rejects it. The bundled repackages carry that vocab (weights byte-identical, verified) and pass the guard. These now match the deepcell-types package registry (xuefei-wang/deepcell-types#70).

2. Add comparison baselines (deepcell-types-baselines + download_deepcell_types_baseline()) for cellsighter / maps(+_stats.npz) / xgboost(+.remap.json), hashes matching the 2026-06-30 retrains (xuefei-wang/deepcell-types#74). Each baseline maps to a list of assets so companion files travel together. Nimbus is intentionally omitted — its weights are distributed upstream (angelolab/Nimbus-Inference) and shouldn't be re-hosted here.

3. Dropped the legacy 2025-06-09 (×2) entries — legacy CLIP checkpoints, incompatible with current deepcell-types code, and with no deepcell-types consumers (both the old and new packages fetch via their own _auth, not deepcell_auth). If you'd rather keep them archived in the manifest, easy to restore.

Merge/conflict: I also merged current main in to resolve the conflict from the torch-mesmer addition (#7) — kept all of your torch-mesmer changes and added download_deepcell_types_baseline to __all__ (this branch predated main introducing __all__, so the new function wouldn't otherwise have been re-exported).

⚠️ Before merge: the new md5s require uploading the exact files to users.deepcell.org at the manifest keys — bundled models b819a7e0/402e94c1, and baselines cellsighter c5a105fb, maps 1ad5e30c + stats 1f462d0c, xgboost 9e51cf1a + remap fba1b0e7.

[rossbar]

We do not want to remove previous versions of the model - these are important for reproducibility/showing progress.

[xuefei-wang]

But the old checkpoint requires the users to checkout an old snapshot of the code base in order to use, right?

[rossbar]

But the old checkpoint requires the users to checkout an old snapshot of the code base in order to use, right?

Yup! But that's no problem as it's all in the repo - all we have to do is point to a commit hash for users who want to do this (e.g. hubmap)

[xuefei-wang]

But the old checkpoint requires the users to checkout an old snapshot of the code base in order to use, right?

Yup! But that's no problem as it's all in the repo - all we have to do is point to a commit hash for users who want to do this (e.g. hubmap)

Gotcha, reverted! The PR should be good.

[xuefei-wang]

Confirmed — every hash in this PR matches the md5sum of the weights that will be uploaded. Verified against the release checkpoints on our side:

version / baseline	asset_key	manifest hash	local md5sum
2026-06-15	models/deepcell-types_2026-06-15_resmlp.pt	b819a7e0b177ad5330394eab3c6c7ad8	✅ match
2026-06-23-ptft	models/deepcell-types_2026-06-23_resmlp_ptft.pt	402e94c103c5e489a57433cd107009d3	✅ match
cellsighter	models/deepcell-types_baseline-cellsighter.pth	c5a105fb044ad82c82817a32aabbae7c	✅ match
maps	models/deepcell-types_baseline-maps.pth	1ad5e30ceaccfdb91050b663099258fb	✅ match
maps	models/deepcell-types_baseline-maps_stats.npz	1f462d0c8bf531af73026d415d52728d	✅ match
xgboost	models/deepcell-types_baseline-xgboost.json	9e51cf1af8c6c43b00871a821fde0f57	✅ match
xgboost	models/deepcell-types_baseline-xgboost.remap.json	fba1b0e705e5f7747eb2f9cae30815ba	✅ match

Two notes on the two main-model assets:

• Both are the vocab-bundled repackages — they carry ct2idx (51 cell types) + canonical_channels (278 markers) so deepcell-types' checkpoint-vocabulary guard can verify class/marker ordering at load. The un-bundled originals predate that guard and have different md5s (704616a1…, 067f558b…); those must not be served.
• The bundled files are uploaded under the asset filenames above, i.e. without any _bundled suffix.

The four legacy/other entries (2025-06-09*, cellsam, torch-mesmer) are unchanged by this PR and were not re-verified here.

Merging — the deepcell-types side (vanvalenlab/deepcell-types#47) will pin vanvalenlab/deepcell-auth@main once this lands.