Because if the compiler hangs, you don't get any executable code. And in the specific case of Haskell, we are talking about type constraints that are perfectly cromulent, are perfectly correctly expressed and mean something in some abstract mathematical world in which your have infinite compiler time, but in the real world will simply never finish compiling. It's not because you've failed and you can fix it by twiddling something, it's because Haskell is not actually that sort of language, and if you try to actually use the type system as a Turing-complete constraints satisfaction language, you'll get exponential performance... or doubly-exponential performance, or worse, though it stops mattering pretty quickly.

In both theory and practice, this means you can't really use Haskell that way. It's a typed language with the ability to occasionally skirt the limits, ride the line, and dip into a bit of undecidability when useful (and with a real, in-practice risk of it blowing up in your face even so), not a language in which one routinely uses the type system in a "Turing complete manner", if I may be allowed to gloss over precisely what that means. Richer type systems is a topic of ongoing research, but is not a solved problem.

While we're at it, what would be really cool is if we could design a program to automatically look at the Haskell program, decide if it's undecidable, and fix it.