Hi Nicolas!
Yes I was talking about transliteration, as described with Lemon. Yes I also realize that it could be automatically inferred, but only if things in the Lexeme namespace and it's data model in Wikidata are understood for a programming model (or later Wikifunctions or Sparql queries) to make those inferences. Isn't that correct? If you already have working knowledge on a practical example in Sparql or some other language that can query Lexemes and derive those transliteration inferences across the Wikidata Lexeme namespace then I would really appreciate it and have that documented. Ideally documented on a page in Wikidata, and that could be linked to from the Lexeme documentation and tutorial pages for advanced users.
As I hinted at before, I think a good first step would be to document and see how transliterations can be automatically inferred currently in Wikidata and its tools ecosystem if need be, by using those transliteration examples in Lemon as found within its own cookbook and docs: https://lemon-model.net/