Totally agree with timings. Those remember/caching goes god know where. Anyway, int itself will not be faster, of course, but message is, that you can directly compile other non-int part. And timings then should be average ones for every procedure with respect to different orders (and restart of course within every order change). But i thought it's unesrtood.
About lexical scoping: the idea is to minimize lexical scoping itsef (i didn't say that it becomes 0 at all). For relatively long H it will be essential. You understand that say string x86 instructions (movs, and so on) are muuch slower than even fpu instructions. Besides no operators grabbing will be present (while operands will be present in both cases).