Decidability isn't even that useful. If typechecking takes a million years, that's also bad. What you want is guarantees that correct programs typecheck quickly. Without this, you end up in swift land, where you can write correct code that can't be typechecked quickly and the compiler has to choose between being slow or rejecting your code

My 2¢ from an interactive theorem proving perspective: In so-called computational type theory, typechecking is indeed allowed to be undecidable, and you get a lot of cool expressive power (e.g. well-behaved quotients and subtypes) as a result. This was one of the big ideas behind NuPRL back in the day, and Istari[0] more recently.

[0] www.istarilogic.org