The only thing I can really think of that doesn’t end up adding some sort of VM into Celestia is TIA addresses controlled by a key that is generated by Distributed Key Generation, and creating the rollup such that, as part of the output of a valid (rollup) block, a signature is also generated over a valid Celestia transaction with the TIA DKG key, which when submitted to Celestia would alter the TIA balance according to what happened in the rollup.
However in this case validator set collusion has more devastating effects, since the validators together could just sign any arbitrary TIA transfer transaction, so even in this case it would be necessary that byzantine behavior would cause a loss on the rollup greater than the balance of the TIA-backed-DKG (or even cause a loss elsewhere, if proof of collusion can be submitted to some other contract/chain/whatever, in case the rollup can’t recover from total validator set collusion).