It isn't acceptable to just permit memory-leaks in case of cycles. There's no way we'd ever see something like that in a serious JVM.
Unless there's some serious machinery to provide assurances against reference cycles, of course. Which sounds like an interesting research project, come to think of it.
That's only in the new --gc:arc which is still somewhat experimental. The old gc, --gc:markandsweep IIRC, does detect cycles, and has a deadline schedule (i.e. you can tell it "I now have 5ms, collect as much as you can"); But it has per-thread heaps, which means passing data between threads often entails a copy. You also have --gc:boehm (no cycles, shared heap), and --gc:none (no gc, you should take care of freeing memory yourself).