If your goal is to allow varying size big nums so that you can do maths in the secp scalar field, you don’t want to enforce minimaldata – otherwise if you take the hash of a blockheader to generate a scalar, and naturally end up with leading zeroes, you have to strip those zeroes before you can do any maths on that hash. Stripping an arbitrary number of zeroes is then also awkward with minimaldata rules, though not impossible.
(I think a minimaldata variant that only applies to values pulled directly from the witness stack, but not from program pushes or the results of operations) would be interesting)