You have much more control with arrays
How about:
1) you are dealing with colossal data sets where the data has to be mapped files and not allocated with malloc
or new
because of its size. Under this scenario worrying about what to do if you didn't reserve enough address space at the beginning may be moot, though I suppose you could unmap - extend - remap the file, unless prohibited by address fragmentation or my second point.
2) Code that uses lockless multiprocessing. Performance hits of stopping the threads for re-allocation (or any other "STL goodie") may be unacceptable, hence use arrays, you have got much more control, you may need to call a lot of functions to pause other threads before you resize anything.
BTW, I'm usually dealing with 1 and 2 at the same time. Arrays of structures + pointers work wonderfully, Compiling with C++ because you you can still use some C++ features elsewhere in the code. I'm sure I could think of many more examples if I tried hard enough