The Tx that Merchant A receives may or may not be heard about by Merchant B. It is unwise IMO to make a security feature rely on the assumption that all nodes will know about all unconfirmed Tx. Then, then the buyer mines his own block sending the coins back to himself. Merchant B will accept zero-conf Tx building off that block if he doesn't know about the Tx from Merchant A. The buyer could peer directly with Merchant B so Merchant B is likely to see the new block faster than Merchant A. This gives the buyer more time for his zero-conf Tx to be accepted by Merchant B. Merchant A must "be on the ball" to quickly realize the dbl-spend and make sure all other nodes know this. Naturally or purposefully occurring 1-2 block blockchain forks may make this proposal unreliable as well.

Also, what if there are buyers who always dbl-spend coins, so that merchants will be discouraged from accepting zero-conf Tx using this method? There is no additional cost to buyers who dbl-spend in this scheme anyway.