The set of addresses that contains just letters is the second biggest set.

You've limited base 58 to be around base 49.

I'm not certain of the range produced from running the hashing algorithm to get addresses but I think the range is quite big still. As I reread this it feels similar to the argument for n-1 character addresses being less secure than n character long addresses - which is false but seems like it might be logical to someone new.

I'd be interested if you could change it to numbers and release the times it takes to generate those addresses though (I think it should take about 49/9=5.44x longer - but I'm not that sure).