Doesn't the consistent versioning approach make for eventual consistency which might never arrive?
Suppose you have a node where tombstones never work...<p>The missing part is monitoring error conditions when it times out...
Which also should be resistant to byzantine failure, so more than one monitor voting on it.<p>And when that happens you <i>have</i> to force stop the node or multiple thereof, diagnose which direction the consistency failed etc.<p>You need to have a well defined requirement set on the consistency and timeliness in the first place.<p>As far as turtles, assumption of faultlessness (of storage, of networking, of whole distributed machines, of the power supply) is one of these.
The moment your CRDT starts to involve backups things get funny.