[aperson30]

Why

When a scan is kept in UserData, an idle-time process (scripts/user_dataset_second_opinions.py, running on bdmap1 since Oct 16) runs it through the other models whenever the site is idle and a GPU worker is free, and saves each result beside the original. It only did this for ePAI, Atlas-Net and LesionSegmenter. The site now offers three more models.

What

The default list is now every model in the picker: ePAI, Atlas-Net, LesionSegmenter, SuPreM, R-Super-Pancreas-Kidney, R-Super-Merlin. Everything else about the process is unchanged (only while the site is idle, only on a GPU worker, cancelled the moment a user job appears, daily cap, back-off, one atomic folder per version).

• All three new models already write combined_labels.nii.gz in the viewer's label scheme (checked by running each on a test scan through the production run path), so the saver needs no change.
• Their weights are single files, not nnU-Net folders. services/model_provenance.py gets a single-file rule: the checkpoint and the class list are hashed into version.json, with the settings the runners fix (SuPreM: U-Net backbone; R-Super: whether lesions are masked by the model's own organ masks).
• The R-Super releases have no entry in MODEL_TO_VIEWER, so label_support says known: false for them instead of guessing which labels they can produce.
• Docs, .env.example and the script's help text updated.

Rollout

The running process holds the old list in memory, so after deploying it has to be restarted once (stop file, then start; documented in deploy/USER_DATASET_SECOND_OPINIONS.md). The backlog is 387 kept scans x up to 3 new models, paced by the existing cap of 100 runs per rolling 24 hours (about 12 days); raise SECOND_OPINIONS_DAILY_CAP if the workers are known to be idle.

Tests

• test_user_dataset_models.py: the default list equals the models in UploadPage.tsx's picker (fails if a model is added to one and not the other, and with the list set back to three); every default model has a checkpoint rule and settings; a kept scan is owed a version from every other model. The loop tests now use an explicit three-model list so they test the loop, not the default.
• test_model_provenance.py: the single-file rules equal the runners' variables, defaults and fixed settings (drift guard); hashing the weights and class list; missing file and missing class list; label support for the R-Super releases. Fails when the SuPreM rule is removed.
• 150 tests pass in the user-dataset set (137 on main). test_user_mri_dataset.py::test_daily_caps_apply_per_user fails in that combined run on main too (it passes alone); not touched here.