Yes, I thought about eval, that would need special rules indeed. The "nodes in the compiled tree" approach is a good proposal - I was trying to formulate the rules in such a way as to be unambiguous, and that would be a watertight way to get it right. Kudos to Juerd for a great suggestion on how to implement that with very little effort.
As to the point, it's true that these rules favour experienced programmers, but then these folks tend to be hard to beat in the pro league anyway. What usually leaves me unsatisfied at least about most simpler golf challenges is that concise in terms of expressiveness approaches tend to be too long when written out, so simplistic, brute force approaches, often really hackish ones, are favoured.
Basically, what I'd like to see is a challenge to use the language's expressive power to the maximum.
Makeshifts last the longest.