Oh, I guess didn’t describe it properly.
This is important for the reason you described. There would need to be a new transaction type which includes a hash of the witness. Miners could then provide the witness matching that hash to prove the transaction tried to use an address it didn’t provide.
I had thought this would let you make stateless clients without assuming EIP 648 but just realized it only pushes the problem one level down; miners can still DOS other miners by sending out blocks with witnesses which don’t prove all addresses which the block accesses.
So I suppose something like EIP 648 is necessary, and it’s much simpler too.