In fact, singling out ZFC gives you another example of this kind of result: if ZF is consistent, then so are ZF+C and ZF-C.
Oh, I see. I read "A is consistent with B" as "A+B is consistent", but I can see the validity of the reading "A+B is 'as consistent as' A".
In a deterministic algorithm, this bound will always exist. In a non-deterministic algorithm (e.g. one that uses a random walk), this bound will not. The funny thing is that in the real world (and for big data) the latter algorithms generally run faster than the former.
If you're wondering why a non-deterministic algorithm can't have bounds, it's simple! Because it would range from a trivial best case of I get really lucky and randomly hit every "good" factor in constant time O(1) to the trivial worst cast of I get really unlucky and keep guessing the wrong numbers ad infinitum.
It's therefore not far fetched that some algorithms might in fact always work correctly, but at the same time a proof for their correctness doesn't exist. You can never be sure (literally!) that you're not just confusing it with an algorithm that's wrong but happens to work on all problem instances you've thrown at it so far.
IIUC, when the article says unprovable, what it actually means is unproven. To correctly describe an algorithm as unprovable, you would have to have a proof of its unprovability, which would be quite something.