Janitor activation of vbucket post rebalance failure results in unnoticed dead DCP stream

We saw this recently in production. Sequence of activities was:

1. Rebalance failed while dcp takeover was in progress for some vBuckets. State of these vbuckets was dead on old master but not yet active on new master.
2. Janitor showed up at some point to cleanup the situation - change the state of dead vbuckets to active among other things.
3. Prior to setting the vbucket states the replications were cleaned up. As part of the clean up, new replications were established from the about-to-be-active-but-currenlty-dead vbucket to its replicas.
4. Next, janitor changed the state of the vbucket to active in KV engine.
5. Because of the vbucket state change, KV engine "marked the stream as dead" (i.e. sent a DCP_STREAM_END). This was not picked up by ns_server and no new replication stream was set up.