Previous implementation averaged token count across all paradigm cells,
giving disproportionate weight to lexemes with more cells (e.g. through
syncretism). Now groups by lexeme first, then averages the per-lexeme
means, matching the "aggregated per lexeme" definition in the docstring
and README.
Numeric impact is small (±0.02-0.03 across all four tokenizers), no
conclusions in README change.