It could be useful to have a more fine-grained distinction of states for cluster members, e.g. joining, up, leaving, down and so on.

For example, when a member tries to join it is in a joining state until everyone else in the cluster agrees that this node is available now. Then it can be promoted to to the "up" state. Other states can be handled in a similar way.

The reason for this proposal is to be able to see when a node should start or stop performing user-defined tasks as a cluster member. E.g. the node that is joining, but is not up yet should not be allowed to serve application level requests (assuming that cluster nodes, when they are ready, serve some sort of requests or perform other application specific tasks).

Or, when node is about to leave, it could do it gracefully, so that others are aware and can prepare for that, e.g. by migrating tasks/data from this node to other cluster nodes.

Very nice and small gossip framework!

But it seems that you use absolute time which is not very safe in distributed environments.
Probably it is a better idea to use vector clocks instead? Some well-known projects like Akka Cluster, Riak, Cassandra and Voldemort seem to use this approach.