The "bug" in the counting used here is that the overall Kolmogorov(?) complexity is not being accounted for because it is shunted elsewhere.

This is true for the other "tinyforth" implementations as well - such as sectorforth and milliforth, because the actual code ends up in the "input part of the Turing tape" if you want to think of it that way.

The right way to count it might be to measure the closure of all your dependencies (in this case, all of the input bytes that are needed plus the bytes in the BIOS that are needed).