currently, we do immediate a.k.a. leading-edge throttling/debouncing
this means the language server can hold a copy of a document that is up to 500ms out-of-date compared to user edits, causing various quirks/issues with diagnostics and positioning of inlay hints