Another case where a long cycle length is useful would be the sort of scenario in my post, at much larger scale. If you have thousands of machines re-seeding at every service restart and then generating sequences of millions or billions of random numbers from a sequence, and you want a vanishingly small probability that those sequences will never overlap, could you use a CSPRNG?
Re: the generator properties, do you _really_ believe that? I mean, good non-CS PRNGs usually have fairly rigorous presentations and pass statistical tests. They aren't as sexy as crypto stuff, and maybe don't get the eyeballs, but they're relatively straightforward proven math whereas all of the CSPRNG results are based on conjecture. It seems at least as likely that some unexpected property of a CSPRNG will be discovered, a la RC4.
I do believe that. RC4 is a good example, actually; the unexpected properties that make RC4 be considered completely broken for cryptography are considered irrelevant for simulations (I won't swear by this, but I'll bet that RC4 passes all of TestU01's tests). The standards are just so much higher for cryptographic primitives. I'm sure unsexy generators can be suitable to this or that application. But due to the ad hoc nature of the statistical tests used, you can't be sure that your generator is not oddly correlated with a concrete application. I've written a comment a few days ago---on a thread very similar to this one---about a real-world case of this happening [1].
From a cryptanalysis standpoint I think that's consistent -- the probability of generating any particular single shuffle is astronomically low, so the difference between the actual probability and 0% is probably outside the realm of what's computationally feasible. But with something like MT19937 I can show that its at least possible for every outcome to have a non-zero probability. Perhaps this is just interesting academically, but maybe not if you're actually playing a card game for instance?
You have already covered the distinction between guaranteed equidistribution and not, so I don't think there is much to add there. I do think the concern is mostly academic, since the difference between a subsequence that doesn't exist and one that will never be reached before the universe collapses is irrelevant to the user. I will note that Richard Brent, someone who is much better qualified than I to discuss this, has argued against the usefulness of equidistribution: http://maths-people.anu.edu.au/~brent/pub/pub240.html
In particular, if there were some 0 <= y < 2^128 such that AES_k(x) != y for all 0 <= x < 2^128, then there must also be some z such that two different values x1 and x2 encrypt to the same value, which can't happen if AES is reversible.
As to why we insist that cryptographic primitives are uniformly better? Well, that is more subjective. I personally believe that they have gotten fast enough that the set of cases where they are unacceptably slow is vanishingly small, and the trend is clearly on the side of crypto getting faster. And I can sleep at night knowing that my results are not wrong due to some unknown, unexpected, property of my generator.