Ah, you're still misunderstanding my point: You can get atomicity in the

worst-case where the communications medium fails *and* stealth payments
that use up no extra space in the blockchain. This gives you the best of
both worlds.

Sounds great! How does a lightweight client identify such transactions without any markers?

Regardless, there are lots of other useful features that require BIP70 to work well person to person, like messages, refund addresses, etc. So extending it with ECDH makes sense in the end anyway no matter what.