Rebuild Sphragis from the complete treebank sentence pool

Chunk training splits at fixed task sizes

Enforce exact verse boundaries and nested CoNLL-U

Represent text chunks as JSON lists

Unify source text across task sizes

Remove CoNLL-U provenance leakage

Validate all CoNLL-U with official loader

Freeze exact evaluation chunks

Add chunk-size dataset variants