Is it obvious that the following holds?
By moving a subset of transactions, whose combined feerate is S, to the front of a linearization whose first chunk has feerate <= S, while keeping the relative order within the moved and within the non-moved transactions the same, the feerate diagram will be >= the old one (obviously subject to that being topologically valid).
If so, I may be able to prove prefix-intersection merging always results in something >= both original linearizations.