Thanks for the suggestion.  In the case of  gmaxwell's function, when the network checks the result it's still going to run a huge number of SHA256 calculations that the attacker doesn't need to compute, so as far as the network is concerned any contribution from the hacker is going to look bona fide.  Of course  for gmaxwell's function the optimization is so easy there are probably compilers which would find it, but it's also easy to make much more cryptic (even cryptographically unidentifiable) optimizations.

It is possible to force the attacker to do the calculation as advertised by incorporating oblivious hashing into the hash function, but that  would slow things down even more than devoting every second block to a default hash function.