The memory pool is just talk. There is no expectation that the memory pool has to satisfy some standard as to what will eventually exist in the block chain, and there are any number of ways that people could communicate transactions to one another without putting them in the memory pool. The memory pool can't be treated like a contract because there is no cryptography to enforce it--there is no contract until the transactions appear in the block chain, inherently.
Mike Hearn's proposal is nonsense because it requires miners to develop a concensus on which blocks in the block chain are dishonest. There is no way to prove cryptographically that a block is dishonest because the block chain itself is the concensus system--before there is concensus, there is no concept of dishonesty, at least as far as double-spending goes. In order to decide that a block is dishonest and reallocate the coinbase, a prior consensus mechanism would be required. Of course, such a consensus mechanism would also be subject to attacks just like the block chain.
Maxwell's proposal is very good. One only trusts the oscar not to double spend, which is perfectly reasonable if oscar is a well-known service. Normal everyday wallets for immediate payments would simply require a little more infrastructure.