I hit a failure in TestSearcherManager (NOTE: doesn't always fail):
It was tripping the assert inside SearcherLifetimeManager.record,
because two different IndexSearcher instances had different IR
instances sharing the same version. This was happening because
IW.getReader always returns a new reader even when there are no
changes. I think we should fix that...
Separately I found a deadlock in
TestSearcherManager.testIntermediateClose, if the test gets
SerialMergeScheduler and needs to merge during the second commit.