I've also found that ripgrep performs surprisingly well with folders containing large numbers of text files - this makes raw json on disk more convenient for many simple usecases which don't require complex queries.
Imagining a moderately large JSON object, a lot of them actually, not hard to read or write exactly but they do carry some common references.
SQLite can: only rewrite the changed fields with a query, and transact the update across the objects. The filesystem can do atomic swaps of each file, usually, if nothing goes wrong. SQLite would also be faster in this case, but I'm more motivated by the ACID guarantees.