I did not only test one network. I understand there are probably "worse" choices in terms of network, however QuakeNet, being the largest network, has the widest random sample population. You need only take a stats 101 course to realize that 5% of a large enough random sample size means it's likely to be 5% in other places too (except for specialty networks or hyper-regional ones).
For comparison, I just tested RusNet, which I guessed would have a much higher percentage, but the numbers are closer to 10-12%, ie. high, but not as high as I thought. This is, of course, a biased sample, and probably one of the worst-case scenarios. I should also point out that RusNet has (experimental) server-side utf-8 support, so it's not as big a deal there either. Finally, putting this all in context, RusNet only has 5,000 users and 6,000 channels. 10% of that number is small change.
But again, even if there are more than 10%, the vast majority of this 5% (or 10%, or 15%?) is still mIRC clients. Of all the channels I polled, almost 95% were mIRC. Other clients I noticed were: kVIRC (once), x-chat (once), Miranda IM (once, and ironic because it was a "Unicode build"), and irssi 0.8.15 (latest version, maybe 15 times), and a handful of eggdrops. irssi was the second largest number, but still paled in comparison to the amount of mIRC clients that I counted.
The rejoin function Khaled proposed is to avoid rewriting servers. mIRC is unable to join these codepage encoded channels, so they already have to stick to 6.35 anyway. Migrating the users that *do* upgrade seems like the sane solution, rather than doing nothing.
On a sidenote, if you think 6.35 is outdated, you should see the clients these users use. One some channels, the average version was 6.2 (and went as low as 5.9). I don't think they will be complaining about having to continue using their version. Clearly, these aren't the users who will be upgrading to 7.1 any time soon.