So a type-system that does the right thing in 99.999999% of the cases is completely useless according to you?

Don't throw out a number like that. We're talking about a Turing complete language, which means we're talking about protections that have to range over all possible programs and abstractions. It may be that your problem domain is extremely biased toward some small region of that space, in which case a type system that doesn't work there is actually useless.

Either way, it is hard to reason about factors like that numerically when we've not qualified which kinds of programs we're expecting the type system to actually work for.

I think the question is rhetorical.

Not really, a language in which some type errors are impossible (rather than merely very unlikely) makes it drastically easier to prove something useful about programs (e.g. to optimize away runtime checks that would otherwise be needed as a safety net and impossible cases).