The original Rogue level algorithm picks a width for each room between 4 and 25, and a height between 4 and 7.
Ignoring "gone" rooms, that gives 22^9 * 4^9 possibilities with only a 32-bit seed. So some combinations of rooms must be impossible! How does the subspace of rooms produced by the limited seed space compare with the theoretical distribution if rnd() was really random?
Weekend project, maybe? I have been thinking ever since #RoguelikeCelebration about how small RNG seeds (or poor PRNG algorithms) limit the expressive range of PCG in ways that might be unexpected.