Usefulness of `rand()` - or who should call `srand()`?
Use the new <random> header instead. It allows for multiple engine instances, using different algorithms and more importantly for you, independent seeds.
[edit]To answer the "useful" part, rand generates random numbers. That's what it's good for. If you need fine-grained control, including reproducibility, you should not only have a known seed but a known algorithm. srand at best gives you a fixed seed, so that's not a complete solution anyway.
Well, the obvious thing has been stated a few times by others, use the new C++11 generators. I'm restating it for a different reason, though.
You use the output for scientific calculations, and rand usually implements a rather poor generator (in the mean time, many mainstream implementations use MT19937 which apart from bad state recovery isn't so bad, but you have no guarantee for a particular algorithm, and at least one mainstream compiler still uses a really poor LCG).
Don't do scientific calculations with a poor generator. It doesn't really matter if you have things like hyperplanes in your random numbers if you do some silly game shooting little birds on your mobile phone, but it matters big time for scientific simulations. Don't ever use a bad generator. Don't.
Important note: std::random_shuffle (the version with two parameters) may actually call rand, which is a pitfall to be aware of if you're using that one, even if you otherwise use the new C++11 generators found in <random>.
About the actual issue, calling srand twice (or even more often) is no problem. You can in principle call it as often as you want, all it does is change the seed, and consequentially the pseudorandom sequence that follows. I'm wondering why an XML library would want to call it at all, but they're right in their response, it is not illegitimate for them to do it. But it also doesn't matter.
The only important thing to make sure is that either you don't care about getting any particular pseudorandom sequence (that is, any sequence will do, you're not interested in reproducing an exact sequence), or you are the last one to call srand, which will override any prior calls.
That said, implementing your own generator with good statistical properties and a sufficiently long period in 3-5 lines of code isn't all that hard either, with a little care. The main advantage (apart from speed) is that you control exactly where your state is and who modifies it.
It is unlikely that you will ever need periods much longer than 2128 because of the sheer forbidding time to actually consume that many numbers. A 3GHz computer consuming one number every cycle will run for 1021 years on a 2128 period, so there's not much of an issue for humans with average lifespans. Even assuming that the supercomputer you run your simulation on is a trillion times faster, your grand-grand-grand children won't live to see the end of the period.
Insofar, periods like 219937 which current "state of the art" generators deliver are really ridiculous, that's trying to improve the generator at the wrong end if you ask me (it's better to make sure they're statistically firm and that they recover quickly from a worst-case state, etc.). But of course, opinions may differ here.
This site lists a couple of fast generators with implementations. They're xorshift generators combined with an addition or multiplication step and a small (from 2 to 64 machine words) lag, which results in both fast and high quality generators (there's a test suite as well, and the site's author wrote a couple of papers on the subject, too). I'm using a modification of one of these (the 2-word 128-bit version ported to 64-bits, with shift triples modified accordingly) myself.
This problem is being tackled in C++11's random number generation, i.e. you can create an instance of a class:
std::default_random_engine e1which allows you to fully control only random numbers generated from object e1 (as opposed to whatever would be used in libxml). The general rule of thumb would then be to use new construct, as you can generate your random numbers independently.
To address your concerns - I also think that it would be a bad practice to call srand() in a library like libxml. However, it's more that srand() and rand() are not designed to be used in the context you are trying to use them - they are enough when you just need some random numbers, as libxml does. However, when you need reproducibility and be sure that you are independent on others, the new <random> header is the way to go for you. So, to sum up, I don't think it's a good practice on library's side, but it's hard to blame them for doing that. Also, I could not imagine them changing that, as billion other pieces of software probably depend on it.