Wiser monks than I may well offer suggestions in the vein you suggested, but ++.

Update: However, unless I misread this, your randomization does NOT account for the differing quantities of available/extant "numbers (10)," "symbols (depends on character set)," "UC_letters (ditto)," and "lc_letters (ditto again)." If I'm correct, that may induce a pattern of output that would be detectable over a relatively small sample, and thus might reduce the work involved in breaking the p/w generated. OTOH, that might actually make breaking the p/w harder. Comments, anyone?   </update>

Also, the comment at line 81 doesn't seem quite right, in the overall context since to achieve exclusions, such as -N, the invocation must have multiple arguments -- one for length, and one for the exclusion.

One other trivial and highly personal observation (YMMV), I find the use multiple dashes as leading- and trailing-sub/function/etc delimiters (eg, lines like 9, 18, 21 & 33) "intrusive" (for lack of a better word) and somewhat un-lazy. FW(little)IW, a pair of newlines generally achieves - for me - the kind of clarity you're seeking (for which, another ++).