The consensus is that SCStemmers "Stemmer #4" works well for Serbian Wikipedia and Wiktionary text after a Cyrillic-to-Latin mapping upgrade. It doesn't lose any tokens, and it handles non-Serbian text well.
So, the next step is to create an Elasticsearch plugin based on the stemmer and test that out as part of an analysis chain. (Other elements of the analysis chain include folding accents used primarily to mark pitch accent, which are used in dictionary and encyclopedia entries, but not normal text, and probably additional general ICU folding.)
If it's a success for Serbian, the follow up would be to test it for Croatian, Serbo-Croatian, and Bosnian wikis.