So if i am following correctly you are allowing accumulating, not just prime numbers. Which means that a user can make a fake proof.
Example: Say they commit 14. They can then fake their witness that either 7 or 2 or 14 have been accumulated.
This means that the history proof size for a coin goes down from one Merkle branch per Plasma block to two RSA accumulator proofs per transaction of that coin : one proof of membership for when the transaction takes place, and one proof of non-membership for the range where it does not. If each coin gets transacted on average once per day, and an RSA proof of non-membership is ~1 kB, then this means ≈1 Kb * 365 days * 10 fragments ≈3.6 MB for one year.
There is also the need of each user to update their witness as the accumulator is updated. Which adds data availability requirements as well as extra bandwidth to keep their witness up date. But i guess this is less than the merkle tree bandwidth requirements.