Replace the 100-row task with a 5-row task and admit 11 more authors

Remove annotation-scheme leakage from model-facing columns

Add exact 50-sentence benchmark task

Rebuild Sphragis from the complete treebank sentence pool

Chunk training splits at fixed task sizes

Unify source text across task sizes

Freeze exact evaluation chunks

Add chunk-size dataset variants