They would still use transaction.begin() / commit(), as that's what we use for controlling the flush cycle. In that IRC chat linked above we discussed some ideas how that Transaction implementation could look like. Basically it would call some SPI in ORM for triggering the flush.

Probably it'd make sense to raise a meaningful exception when trying to invoke rollback as that's not possible with the non-transactional backends. The transactional backends we have (ISPN, Neo4j) work with JTA; I don't think there is any need atm. for making them work with RESOURCE_LOCAL. It might make sense for some theoretical store in the future which has some kind of "local" TX but does not work with JTA.