Apparently on some devices the weak reference table is limited (numbers around 52000). Even though we don't use that many weak references, when GC is not called for a while they can stack up and a crash will occur before the GC has had the time to clear the references. The C++ peer now holds on to a global ref (strong) which can be obtained to queue the java peer and then release automatically so that the GC can take over after the runnable has been executed.
