I'd suggest my project, Lady Deirdre, which aims to address exactly the problem you described: providing a common framework for custom language semantic modelling. The linked page includes a comprehensive book on this topic. My other project, Ad Astra, demonstrates how to use Lady Deirdre in practice (symbol resolution in a lazy fashion).
More generally, what you're looking for is usually called "incremental computation". Beyond my work, there are several solutions in the Rust ecosystem: Salsa, Adapton, and Anchors.
I also recommend my article explaining how these algorithms work in practice, as well as an article by the author of Anchors.