Training uses overlapping windows while evaluation uses full proteins

[jeffreyHoelzel]

Training defaults to windowed proteins with overlap, while evaluation defaults to full proteins.

Evidence:

• src/pepseqpred/core/data/proteindataset.py
  • _iter_windows supports window_size and stride.
• scripts/hpc/trainffnn.sh
  • WINDOW_SIZE=1000
  • STRIDE=900
• src/pepseqpred/apps/evaluate_ffnn_cli.py
  • Evaluation constructs datasets with window_size=None and pad_last_window=False.

Why this can hurt:

• Overlapped residues are duplicated during training and validation.
• Validation threshold selection is based on windowed validation arrays, but final evaluation is on full proteins.
• Overlap duplicates can overweight boundary regions or long proteins.
• If there are very long multi-pathogen proteins, length and overlap can distort training and metrics.

Planning direction:

• Run an ablation with --window-size 0 where feasible.
• Run an ablation with non-overlapping windows, for example stride == window_size.
• Add metrics that count unique proteins/residues separately from yielded training residues.