again, I think it's fine to assume that library loads don't happen concurrently. Even if they did - having a top-level mutex is a much easier solution that also doesn't blow up OperatorEntry size.

btw, since you already have kernelsMutex_ above - why not just use that (even though we don't really need mutex)

The mutex costs us even less, the complexity added by just the mutex it is negligible, but then we get parallel and race-free library loading for free, don't see a reason for not doing it.

The default pattern for mutexes is to bind them directly to the data they protect and trying to keep these data units small. OperatorEntry size doesn't matter, this is not an entry in the dispatch table.