But the cumulative weight of that tx is not so big, so why the merchant should accept it?
NP, the merchant waits of course for normal amount of confirmations.
On your picture it has 1 confirmation only?..
I can draw more pictures. But I don't think it's necessary. Imagine that the attacker started preparing for the attack a month ago. He spent the whole month to accumulate PoW on top of the second doublespend. He published no transactions during the month. Then he publishes the first doublespending transaction, provides the first confirmation, thus attaching it to recent part of the tangle, waits for the merchant to send him his puchase. Then publishes his secret subtangle and attaches in to the legit subtangle. The first doublespending transaction now is rejected by the network, the second doublespend has more weight.
He would have to attach it "below" the merchant's tx, but yes, you're right, it's a possible attack vector. Anyhow, the referencing algorithm is not yet finished, so we are discussing it with CfB right now.