That is a problem with design of the simulation, or expectations of it, not the random number generator.
Real distributions of (almost) anything do permit a non-zero chance of unlikely sequences. If you're doing something to prevent particular sequences (such as runs) then you're introducing a bias - regardless of the quality of the random number generator.
As Mario said, it is easier to shrink the state space that the random number generator works with than to use a larger state space and then adjust the output to a smaller range (by lopping off zeros, modulo/divide, etc). The two are distinct, but you are treating one as the other. They are not interchangeable either (usually).