It is troubling in the same way, but more subtly, and it has to do with the interpretation of the logic rather than the logic itself. The problem with (a -> a) -> a is that you can prove any a. Now, this is indeed a problem if you're trying to use types to prove mathematical theorems (one interpretation). But what if you're using types to prove program correctness (second interpretation, this one computational)? Why is it troubling? Well, it's troubling because you may believe you've constructed a program that produces some result of type x, but really you haven't, because somewhere along the way, you've used a (a->a)->a function (or forall a b. a->b). But the thing is that from one interpretation you really have succeeded. Your type
is populated, but it is populated with a nonterminating function. Why is that a problem? It's a problem because it may cause me to believe that I have a program that does something, while in reality that program is useless.
Now back to my issue. Suppose that somewhere along the way you rely not on a non-terminating function but on a high-complexity function (e.g. a function that factors integers). You may then believe you've constructed a program, but your program is not only just as useless as the non-terminating one, but useless in the same way. A program that takes 10000 years is much more equivalent to a non-terminating program than to one that completes in one second. Your types are still populated with "false" elements, and so your logic, while now useful for proving mathematical theorems, may still prove "false" programs, in the sense of useless programs.
HOWEVER, what I said has a practical flaw, which still keeps excluding non-termination but allowing high-complexity useful. And that is that it's much easier for human beings to accidentally create programs with infinite complexity, rather than accidentally create programs with a finite, but large complexity. I don't know if we have an answer as to why exactly that is so. It seems that there are many cases of "favored" complexity classes, and why that is so is an open problem. Scott Aaronson lists the following as an open question[1]:
The polynomial/exponential distinction is open to obvious objections: an algorithm that took 1.00000001^n steps would be much faster in practice than an algorithm that took n^10000 steps! But empirically, polynomial-time turned out to correspond to “efficient in practice,” and exponential-time to “inefficient in practice,” so often that complexity theorists became comfortable making the identification... How can we explain the empirical facts on which complexity theory relies: for example, that we rarely see n^10000 or 1.0000001^n algorithms, or that the computational problems humans care about tend to organize themselves into a relatively-small number of equivalence classes?
Nevertheless, it is important to notice that what makes non-termination-exclusion useful in practice is an empirical rather than a mathematical property (at least as far as we know). Which is my main (and constant) point that computation and software are not quite mathematical, but in many ways resemble physics, and so relying on empirical (even cognitive) evidence can be just as useful than relying on math. The two should work in tandem. It is impossible to reason about computation (more precisely, software), with math alone; there are just too many empirical phenomena in computation (and software in particular) for that to make sense. I feel (and that may be a very biased, wrong observation) that the software verification people do just that, while the PLT people (and by that I don't mean someone like Mattias Felleisen, but mostly PFP and type theory people) do not.
How can that look in practice? Well, observing (empirically) that the complexity spectrum is only sparsely populated with programs humans write (and that's true not only for instruction counts but also of IO operations, cache-misses etc.), perhaps we can create an inferrable type system that keeps track of complexity? I know that integer systems with addition only are inferrable, but I'm not sure about multiplication (I don't think so, and I know division certainly isn't). Perhaps we can have a "complexity arithmetics" that is inferrable, and allows "useful rough multiplication" even if not exact multiplication? A Google search came up with some work in that direction: http://cristal.inria.fr/~fpottier/slides/fpottier-2010-05-en... (I only skimmed it).
[1]: http://www.scottaaronson.com/papers/philos.pdf