It’s definitely tough. Recently I spent some time taking some OSM data I had snap shotted and looked to reverify. I actually had a bug that stored every OSM ID, but it saved everything as node, when some were actually way to relation. So I had to ping OSM API + Overpass to check if the object I saved has a matching node, way, relation, in vicinity. It’s surprising how many evolve tags or get deleted from every a small CSV I had!

Type+ID+version is very good as an ID in my opinion. I remember other stuff like Placekey that attempted to put the category and location into the ID with hashes or something. Interesting but complicated for a changing dataset.

OSM evolves daily and I think that calls for very special way of making IDs not so much stable, but maybe we could say traceable in a change graph. Seems to work well.

Wikidata has a very interesting ID system also, I am not sure if I fully understand it but there is something about how IDs for things that are nearby semantically or numerically then also have some potentially relation. I see a lot of those IDs attached to OSM data and it gives another layer of stable ID for an entity, also for brands.