I don't know what the current state of pathing is. It might require a supplemental A*-tree generated that ignores/assumes-path-throughabilty for destroyable buildings, to then deal with the buildings much as non-locked doors are dealt with only when encountered in following their insecure-door-ignoring route.
(How) do climbing enemies their scaling of obstacles in advance? (How) do flying hostiles account for their freedom to exploit open-air access to bypass bridges?
Demoting some 'bait' to merely a task along the way (perhaps stack the path leading to its destruction above the original target on the list, then reassess the primary target once the tick-consuming obstacle is removed, however so moved or further obscured by the time the original - or retarget entirely and go from scratch if the original choice is no longer possible or has become less important) would be the first change to look for.
Then start giving priorites to 'missions' they might want to path for. Freeing a fellow creature (if and only if they are sentient enough to want to, and especially aren't territorial, or is your plan break them out to fight them?) might be lower on the list than getting 'food' or some other need, one such higher need naturally being that of the caged creature being of the opposite sex.
General hostility to dwarfs and dwarfish things (or across the whole set of civilised races, barring particularly noted affiliations) would give the destroyer some need to cause destruction at the furniture level if they could not sense a better purpose. And sacrificial barriers might feature in that list of alternatives if there's nothing more definitively tempting at all...