One problem with the above is that you'd need 16GB of storage to check block validity, even for SPV clients on things like cell phones.  So that's one problem that would need to be fixed.

Maybe if the r-value (number of rounds) were really low and only a small fraction of the memory cells wound up being read from the block header could include some kind of map showing which memory cells got hit during the hashing operation.  Then with the map a low-memory device could check a hash.

Even better, SPV clients could ask full-chain clients to calculate the map for them so it wouldn't have to be stored in the chain.  The point is that giving an SPV client a bogus map won't fool them so there's no trust there.