Strictly speaking, by “lazy” I simply meant that it does not hurry-up to reclaim memory until it is truly pressured to do so.

“It is,” as a friend of mine would once diplomatically put it, “suboptimal ...”
(I think he was secretly a Vulcan.)

Nevertheless, and (of course) knowing nothing at all about the particulars of this application, I do think that you are compelled to find a practical way to “throttle” the maximum number of threads that exist, and to use some kind of a queue to absorb the moment-to-moment fluctuations in the incoming workload.   Not an ideal situation, perhaps, but I see little available choice in the matter.