Another cool way (and unrelated way) to save space is for multiple Plasma chain operators to agree to submit roots together. This requires off-chain agreement, but it means that we can represent multiple Plasma chains in a single 32 byte commitment. Unavailability on one Plasma chain doesn’t impact any of the other chains. Some set of operators could refuse to participate, but this just means that the other operators will commit separately. We could even use a shared DBML/MMR contract.
One important note here is that the multi-root would need to be signed off on by the entire set of operators. What I think you would need is an aggregate BLS signature; the state can then keep track of the most recent root for each Plasma chain in some kind of tree structure, so that if there are groups of nodes that are persistently online then they can just keep updating the Merkle root of their subset as an O(1) operation.