Grouped splits prevent family leakage but are not label-stratified

[jeffreyHoelzel]

The default split type keeps families/groups intact, which is good for leakage control. However, folds are assigned primarily by group size, not by positive/negative support or source balance.

Evidence:

• src/pepseqpred/core/train/split.py
  • split_ids_grouped and build_grouped_kfold_splits keep groups intact.
  • Grouped k-fold assigns larger groups first to balance fold size.

Why this can hurt:

• Folds can have very different positive rates.
• Some validation folds may contain source/pathogen groups not meaningfully represented in training.
• Metrics can look much worse than expected if the split is actually a difficult cross-family generalization test.

Planning direction:

• Add split reports with per-fold:
  • protein count
  • valid residue count
  • positive residue count
  • negative residue count
  • positive rate
  • source/pathogen/family counts
• Consider grouped stratified splitting where possible.
• Preserve leakage safety, but balance label support more deliberately.