0.7.3

[0.7.1] - 2026-07-28

Fixed

• Glossary terms now reliably reach the model. The glossary is sent inside
  the user-turn task statement as well as the system prompt (translation-tuned
  models weigh the former), reworded per task (translation vs. summary), and
  emitted as one terse line instead of an instruction paragraph. The term list
  is capped relative to the text length so a large glossary can no longer
  drown a short paragraph and derail the translation.
• Glossary matching misses fewer occurrences. Terms whose edges are CJK
  characters match inside unspaced Japanese/Chinese prose (the word-boundary
  guard no longer applies there), and English entries also match simple
  plural/possessive forms ("Caesars", "Caesar's").
• Blocks that are exactly one glossary term skip the model. Headings such
  as a bare person's name are rendered directly with the glossary target — in
  page translation, selection translation, and EPUB export — instead of
  risking a model round-trip mangling the mandated form.