Looks like probably a problem that has come in user land before as well (I think Grant pinged me about it once).
Yup, I found
SOLR-4044 which describes the same problem.
May be we can add a check at this line to wait for # of znodes under /solr/live_nodes to match up the number of servers being started?
We should do that to solve this particular problem because this is a test-framework class and it should work regardless of whether you use CloudSolrClient or HttpSolrClient but IMO, we should also find a way to solve it inside CloudSolrServer.