Let the harness speak: drop the benchmark table from a preliminary artefact

Add the LAFA window benchmark result

State the gain as a difference, not a ratio

Correct the distinct-sequence count

Correct the distinct-sequence count: 487,237, not the embedding row total

Name the bank row count for what it is: protein accessions, not sequences

Frozen bundle: learned sparse encoder and reference bank, GOA 227 cutoff

initial commit