0.11.5-M9

Reproduce and fix repeated watch-close failure on OS-X (#393)

Previously after ~100 iterations, the provided `test("openClose")` case
stops receiving events from OS-X. Presumably we must be leaking
something

Moving most of the relevant `CarbonApi()` calls into the `run` method so
they all run on the same thread seems to make the problem go away. We
still need one call running in constructor body for some reason, even if
it doesn't do anything, because otherwise the calls in the `run` method
seem to hang

The test case hangs on my macbook without this PR, passes with