I wonder of consequences if var_int is used in its longer than necessary forms (e.g encoding 1 as 0xfd0100 instead of 0x01)

This is already of interest if applying size limit to a block, since transaction count is var_int but is not part of the hashed header or the merkle tree.

It could also be used to create variants of the same transaction message by altered representation of txIn and txout counts, that would remain valid provided signatures validate with the shortest form, as that is created while re-serializing for signature hashing. An implementation that holds mempool by raw message hashes could be tricked to believe that a modified encoded version of the same transaction is a real double spend. One could also mine a valid block with transactions that have a different hash if regularly parsed and re-serialized. An SPV client could be confused by such a transaction as it was present in the merkle tree proof with a different hash than it gets for the tx with its own serialization or from the raw message.