Verifying cumulative diff requires verifying possibly thousands of headers (all those between peg out and peg in), which is a completely impractical amount of witness data for one script.
Committing headers in Merkle tree form or skip lists allow for logarithmic reduction in that number, as was worked out in the (2014?) sidechains whitepaper. 10’s of headers would be doable.
