If a 51% attacker stopped including all broadcast transactions in blocks "we" would quickly figure out a rule or rules to reject their blocks.
Something like "ignore a longer chain orphaning the current best chain if the sum(priorities of transactions included in new chain) is much less than sum(priorities of transactions in the part of the current best chain that would be orphaned)" would mean a 51% attacker would have to have both lots of hashing power AND lots of old, high-priority bitcoins to keep up a transaction-denial-of-service attack. And they'd pretty quickly run out of old, high-priority bitcoins and would be forced to either include other people's transactions or have their chain rejected.
I'm tempted to code that up and run some tests on a testnet-in-a-box, but there are much higher priority things on my TODO list; I don't think a 51% attack is likely. You'd spend a lot of time and money on an attack that "we" would neuter within a day or two.