OK, looks good. I tried your test with my earlier workaround and the wrapper got deallocated too early, proving that my workaround was indeed wrong and your test is useful. I am still concerned theoretically that the CoroWrapper.send() signature is different from a real generator's send() method, but I think that send() to a coroutine is an internal detail anyway, so I can live with that, and I don't see another work-around.

When you commit, can you do upstgream (Tulip) first?