The arc implementation doubles as the specification. Piggy-backing on mzscheme's gensyms seems like an attractive prospect at first, until you realise that we now have a part of arc which isn't defined in arc. The only way that could work is if gensym becomes considered as a language axiom, and I'm not sure if it's really a candidate for that.
In practice, I wouldn't believe to be that much of a problem. Who would want to mimic the abominable symbols generated by arc with the current methods anyway, intentionally or otherwise?
You have a good point there. Fortunately it can be fixed trivially, by returning the definition of 'ar-gensym to its original form and using 'string->uninterned-symbol instead of 'string->symbol. (I suspect this is what mzscheme 'gensym does anyways.)