Add baked selector splits (dinomaly, adaclip) over the shared cu3s universe
Adds splits/ with two baked train/val/test splits in the Cuvis.AI selector format, plus a README.
Both files are views of one universe (the 1136 measurements of the 15 cu3s sessions under data/),
identity (source, index) with source the posix cu3s_path and index the local_image_id:
dinomaly.json: train 308 (normals only) / val 148 / test 180, for the unsupervised Dinomaly
detector (train-on-normals).adaclip.json: train 808 (308 normal + 500 anomalous) / val 148 / test 180, for the supervised
AdaCLIP baseline.
val (148) and test (180) are identical across both files, so the two models stay directly
comparable. Selectors are concrete file_indices (position-independent, reproducible); the same
file resolves against the raw cu3s sessions (cu3s_multi) and a converted per-frame NPZ
materialization (npz_multi). splits.csv and splits_dinomaly.csv are kept unchanged as the
human-readable provenance the JSON was generated from.
Merging: baked selector splits (dinomaly, adaclip) over the shared cu3s universe.
refs/pr/1 ref