Register the pretrained 2D models

[aymuos15]

Registers the pretrained 2D classification models and resolves the default model from the processing mode. 3D behaviour is unchanged: the 3D defaults are exactly as before.

Stack: fourth of four. Builds on ssk/2d-napari-selection.

Progresses #298.

Changes

• Adds resnet50_2d and resnet50_2d_1ch to the download registry, hosted on Hugging Face at brainglobe/cellfinder_2d_default.

• Records each model's dimensionality and channel count, so the default model resolves from the processing mode and whether a background channel is present. This generalises the existing single-channel fallback:

  dimensions	background	model
  3	yes	resnet50_tv
  3	no	resnet50_1ch
  2	yes	resnet50_2d
  2	no	resnet50_2d_1ch

• Substitution only applies when the model is still the default, so an explicit choice is never overridden.

• A model whose dimensionality disagrees with the requested mode is now rejected up front, instead of failing later on a shape mismatch.

• The napari model dropdown lists only models matching the selected dimensionality.

• Tests: registry consistency, the mode/channel matrix, mismatch rejection, dropdown filtering.

CI cache key

Also rolls the brainglobe cache key, which was the literal string brainglobe. Since actions/cache never overwrites an existing key, the saved entry predated the Hugging Face hosted weights, so every job restored a cache without them, re-downloaded, and then declined to save, which was enough to hit Hugging Face rate limits across the matrix. The key now includes a hash of the model registry with a prefix restore-key, so adding a model only re-downloads the new file. This rides with this layer because it is this layer that changes the registry.

Model provenance

Trained on serial2p data with a held-out test set (n = 10,756), resnet50, 100 epochs. Both are the best-validation checkpoint of their run.

Model	Channels	Input shape	Accuracy	F1	ROC-AUC
resnet50_2d	2	(50, 50, 2)	0.974	0.973	0.995
resnet50_2d_1ch	1	(50, 50, 1)	0.964	0.963	0.992

Both are uploaded and public. Verified from a clean environment: download_models fetches each over the registered URL, pooch's hash check passes against the pinned sha256, and the loaded networks report the input shapes above with a 2-class output.

The rest of the stack

#	Branch	Scope
1	ssk/2d-detection	2D cell candidate detection
2	ssk/2d-training	2D training in the CLI
3	ssk/2d-napari-selection	2D/3D selector in the plugin
4	ssk/2d-model-registry	pretrained 2D models

Review bottom up. Branches live in this repository rather than a fork, since GitHub stacks cannot span forks.