Actually, I have a poolmon script running that should drop vhost count for unresponsive backends; the strage thing is, the backends are NOT unresponsive, they are working as ususal.
you are running and old version of it as the latest version is more compatible with recent dovecot releases.
current version in git correctly uses HOST-DOWN and HOST-FLUSH instead of modifying vhost count.
But even then I still consider poolmon a bit too aggressive in marking hosts down. It does HOST-DOWN and HOST-FLUSH already after first scan failure.
Maybe there should be some grace period and wait for few failed polls in certain period before doing it in case there is some temporary overload or networking failure.
Sami