Not to mention his thinking out of the box and tool around the issue, trading space vs speed in some scenario.
Conceptually it is closed to dock as it provide an api call for some usage without a full blown vm.
It's just a shortcut addressing mode.
The 6502 instruction set takes advantage of this by having additional instructions to load/store values to zero page as well as ones to load/store to generic 16 bit addresses. An instruction accessing zero page memory will be two bytes (op-code + one byte address), while an instruction accessing a generic address will be three bytes (op-code + two-byte address). The zero page access therefore takes less code space, and will run faster (even though the memory access speed is the same) since there are only 2 vs 3 instruction bytes to fetch.
Given the 10x speed overhead of the SWEET16 interpreter, it'd really not make much difference if it chose to use non-zero page locations for its 16 registers. It makes more difference in hand written code where you are optimizing for every byte of code space and cycle of instruction timing.
For the “not much difference” would it be intended because it is using space to trade of low speed. Given this page use less space and less speed, would it still be important.
As far as speed, given that the interpreter overhead of SWEET16 means it's about 10x slower than 6502 code, adding an extra clock cycle for non-ZP access would at worst make that 11x vs 10x. Would that really make a difference to the cases where it was useful? The trade off is that NOT using ZP for SWEET16 would give those 32 bytes of ZP (16 registered x 2 bytes) back to the developer.
However, It seems that SWEET16 was not much used, and I doubt that any change like this - where to place it's registers - would have made any difference to that. It's maybe interesting to consider would a different interpreted code (with different feature set from SWEET16) have been any more widely used?
Yes, you can consider ZP as a page of registers, but it's not like there was any choice - seeing as the 6502 basically had no registers (just A, X, Y) you had to have your program variables in memory, and ZP was faster/smaller that the alternative!