Stop cleared the done channel, so a waiter that raced a teardown got nil
from Done() and selected on a nil channel: an Up racing a Down hung in
waitForUp for its full 50s timeout where it previously failed fast with
"client gave up to connect". The same clearing made Alive() report a
timed-out run dead while its goroutine was still tearing down, letting
Up start an overlapping second run.
Keep the channel in place instead: the exiting run closes it, so late
waiters observe the real exit and Alive() stays true until the run is
actually gone — the lifecycle clientGiveUpChan used to have. Stop still
drops the current client and invalidates the generation.
The stale-running-chan test begins a fresh run before waiting, as Up
does, so the kept exit signal does not race the stale clientRunningChan.
Zoltán Papp
3615c01fe3