The idea is to steal only if request fails to return work in a reasonable time frame i.e. if a job is keeping the scheduler busy for more than a minimum amount of time.
I also have no number to back it up (we deal wit a relatively small numbers of fds at $JOB and we care of latency more than throughput), but I do not buy epoll_wait generating late batches of work: it seems to me that you should only be using epoll for fds that are unlikely to be ready for io (i.e those for which a speculative read or write has returned EWOULDBLOCK), which means you should not have large batches of ready fds. Even if you do, the only case you would need to steal after that is if another processor run out of work after this processor returned from epoll wait (if it did before, it should have blocked on the same epoll FD and got a chunk of the work directly or could have signaled a request for work), which might be less likely.
Anyway at the end of the day a singe mostly uncontested CAS is relatively cheap on x86 especially if the average job length is large enough, so maybe it is just not worth optimizing it further, especially if it requires more complexity.