Reflector::reset has an (admittedly unlikely) race condition

[nightkr]

version and data use separate mutexes. In most cases, either only one is used, or the data mutex is held for the duration of the version mutex. However, reset releases data before grabbing version, which can lead version to be reverted if single_watch (via poll) applies an update in between the two locks.

The practical effect of this would be that poll() restarts can get duplicate events (though they would just be no-ops).

[clux]

Thanks! Yep, this looks wrong. And it also doesn't look like there are any good reasons anymore to need two mutexes, so we should just combine their use.