v16 mix: 4,208,639 rows — twin gates, short Russian re-sourced, referees held out

v14 mix: v13c recipe with short-floor 0.5 (train/val short-lane recomposition; test split unchanged)

v13c mix: v12 sources + vpermilp/lid-* crawl datasets, 120k cap per language

upload mix: 1,567,513 rows

recompress: zstd level 3 -> 19 (111 -> 89 MB, content byte-identical, read speed unchanged)

docs: fix card language counts (26 langs, tags from data)

v11c mix: 1,535,170 rows (767,227/415,473/352,470), 30 classes

initial commit