Feature Reconstruction
This source-first release excludes precomputed embeddings, graph tensors, structural features, and packed feature caches. Reconstruct all derived features from data/proteins.parquet using the published PhaseFlow preprocessing workflow and documented pretrained components. Keep generated caches outside this package so the release inventory stays immutable.
Bind regenerated features to both protein_id and sequence_sha256, and verify sequence length before cache reuse. The fixed files in data/training_plan/ are training-protocol metadata, not feature caches. Validate their identity against data/training_units.parquet before creating multi-device batch inputs.
benchmark/ provides evaluation inputs only. Inference may require freshly generated features for benchmark sequences, but benchmark records must not enter the training or model-selection workflow.