I've been aware of scaling artifacts in $rand(), but these samples are pretty wild. I wouldn't have expected 24 bits (2^24-1) to be so profound with this effect.
I do wonder if there might be a better, faster even, pseudo-random generator mIRC might use. AutoHotkey is using Mersenne Twister MT19937 by (C) 1997 - 2002, Makoto Matsumoto and Takuji Nishimura (which only needs a copyright notice in the help file; steal it from AutoHotkey's help file.) http://www.math.sci.hiroshima-u.ac.jp/~m-mat/MT/emt.html
((Oh, looks like they don't have any license requirements anymore.))