Well, the 1982 1k implementation he was trying to beat did not cover these rules either. In fact, this version implements pawn promotion to queen, which the ZX 1k version did not. So the part about beating a 33-year-old record stands.
An interesting challenge would be to see if you can add all the missing rules and still keep it under 1k. So the lovers of strict definitions would finally have their version!
Another interesting question: If you could design your own instruction set, what is the smallest you could make a complete chess program? Is there an ultimate limit, information-theory style?
I propose an instruction set which consists of a single instruction. I'll give it the mnemonic CHS, and when executed, it causes the CPU to play a game of chess with the user.
(That's a long way to say that the question is somewhat under-constrained.)
Instead consider a CPU architecture and accompanying notation designed especially for playing chess. The other idea is not very interesting.
Assuming some arbitrary architecture plus the necessary code to emulate it on a normal architecture (including your stupid CHS example) would normalize the "size" of the program, anyway -- unless there are hardware structures that are especially good for chess. Which was the essence of the question.
An interesting challenge would be to see if you can add all the missing rules and still keep it under 1k. So the lovers of strict definitions would finally have their version!