Sparse or zero-valid windows still participate in training

[jeffreyHoelzel]

The dataset yields all windows, including windows with no valid labeled residues. The trainer handles zero-valid masks by producing zero loss, but the optimizer still performs a step.

Evidence:

• src/pepseqpred/core/data/proteindataset.py
  • ProteinDataset.__iter__ yields every window from _iter_windows.
  • The final mask may be all zero when a window contains only uncertain or padded residues.
• src/pepseqpred/core/train/trainer.py
  • _batch_step creates zero loss when mask.sum() == 0.
  • Training still calls zero_grad, loss.backward(), and optimizer.step().

Why this can hurt:

• Adam step counters still advance on no-information batches.
• Learning-rate dynamics and optimizer moments can be affected without a meaningful gradient.
• If some pathogens or proteins have many uncertain/unlabeled regions, they can consume training steps without learning signal.

Planning direction:

• Add logging for zero-valid windows and zero-valid batches by split, fold, rank, and source.
• Skip optimizer steps when a batch has zero valid residues.
• Optionally filter zero-valid windows in ProteinDataset for training.