Converging to a single value (which is a liveness property) and what the values will contain (which are usually safety properties) are different stories... when the partition heals, if the nodes detect they have a different version for a given piece of data, they can just remove it, and the system is convergent, which means it is eventually consistent (a liveness property). What happens in the merge function instead provides additional safety properties. For example in the case of a CRDT that models a Set type, with the merge function that is "union of all the values in the nodes", you can state the safety property that if a Set element X was added into a node, at no time will happen that the nodes will be able to talk again, and the system will converge into a Set which excludes such value X.
