Following up on this thread.

It seems the data model (to deal with the tricky business of translations) indeed set the practice to place translations on a Sense level.

• A list of Sense Statements further describing the Sense and its relations to Senses and Items (e.g. translation, synonym, antonym, connotation, register, denotes, evokes).

This is also reflected in the SVG image of the data model where translations are at a Sense level, and not the Lexeme level.