You’re confusing transactions (the abstract object) and their serialization (the bytes on the wire in the P2P protocol or on disk).
Sure, SegWit introduces an extension to the P2P protocol (BIP144), which relays witnesses along with transactions, and old clients wouldn’t understand such messages.
But old clients don’t see them. Witnesses are only included when the requester asks for them, which old clients don’t do. For them, new clients remove the witnesses before relay. Exactly because the witness of a transaction does not contribute to its txid, it is in fact possible to remove them before relay, without invalidating them. That is what makes it a soft fork.
