Represent text chunks as JSON lists

Remove identifying metadata from metrical lines

Fix metrical chunk aggregation

Unify source text across task sizes

Freeze exact evaluation chunks

Add chunk-size dataset variants

Test aggregate genre split ratios

Rebuild work-stratified splits with seed 776