It seems you want something similar to what we implemented for ISPN-619, which was basically implementing this concept for CacheLoaders (as they are assumed to be much slower than Infinispan).
Currently commands are set right away, and the ReplicationQueue works at the messag layer.
The short answer is yes, you should implement your own buffer.
Still I like your idea, and it sounds something reasonable to implement on HotRod.. could you please open a feature request on JIRA?
If you would you like to try implementing it and contribute it you would benefit from our directions, reviews and some other hundreds of testers
This might make sense not only on HotRod level, bet even when any comunication happens across nodes: