In fact, Monte Carlo tests are done *specifically* using pseudorandom
sequences so that they can be repeated and validated against different
models -- nothing as spectrally poor as linear-congruential, but still
repeatable.
/dev/random is for the times when you want *real* physical entropy...
and since the whole *point* of a kernel is to hide you from physical
variation, the only practical way to get it is with cooperation from
the kernel. That's why I'd argue that /dev/random is *special* and
should at best be a strongly worded default=yes option.
As for portability: I've heard of at least one named-pipe based
user-space portable implementation of /dev/random. Also, Ted's kernel
code has been ported (or has inspired related implementations) to
several other platforms...