So you're suggesting "front-running" the construction of the block itself by securing consensus in the tx mempool? How would you achieve that consensus, given the Byzantine-general problem?
If you know who the block producers are a priori, you could achieve consensus by having them each submit a POW for their view of the current mempool. Disallowing duplicate results for the POW solution, you then chose the most common block hash submitted as the mempool consensus and all the corresponding transactions then get put in the block.
You need to know who they are in advance because otherwise you could have islands of nodes producing their own isolated consensus, which would indeed be a fork.