Close #4793

mteb/leaderboard/app.py:594-596 initialized default_languages from the polars language column (full codes like "eng-Latn"), while every consumer (update_scores_on_lang_change, update_tables, and _cache_on_benchmark_select) treats lang_select values as 3-letter codes
("eng").

Consequences in update_tables:

1. set(languages).issuperset(_benchmark_full_languages(...)) always returned False (full vs 3-letter), so the "skip when all selected" optimization never kicked in.
2. The mask pl.element().str.split("-").list.first().is_in(lang_set) split each row's "eng-Latn" to "eng" and looked it up against the full-code lang_set — matching zero rows.

Initial render dodged it (no update_tables); any filter click triggered the broken language filter → empty table.