Since OP is using Haskell, the actual code most likely won’t really touch the JSON type, but the actual domain type. This makes migrations super easy to write. Of course they could have written a fancy in-house way to do that, or just use the safe-copy library which solves this problem and it has been around for almost two decades. In particular it solves the “nested version control” problem with data structures containing other data structures but with varying versions.

Yes, that's what we do: Migrations with proper sum types and exhaustiveness checking.