Currently all greedy reducers pick the first sequence as representative (deterministic by sorted id). Configurable policy options:

• first (current default)
• longest_sequence
• best_quality (requires a quality column)
• earliest_time (requires a time column)
• max_label_value

Outer split (leakage-aware) + inner CV folds. Useful for proper model selection on biological data.

Split based on precomputed structural distances (id1, id2, dist edge list):

1. Graph clustering: threshold → clusters → group split
2. Connected components: threshold → components → split by component

• homology_then_distance — cluster by homology first (no leakage), then distance-aware selection within train/val
• homology_and_stratified_numeric — homology clusters → bin by numeric label (cluster median) → greedy cluster assignment preserving bin distribution
• time_and_leakage_constraints — temporal split (train earlier / test later) + guardrail: if a cluster appears in test, no cluster member leaks into train

• Sylphy embedding export format: validate embeddings.npy + embedding_ids.csv + meta.json layout in biosieve validate
• Eris-compatible output naming: splits/train.csv, splits/test.csv, splits/val.csv, folds/fold_*/