Call the input a natural number \$n\$. I don't see a qualitative difference between:
- Returning one number, in a challenge where it so happens that the reciprocal of the correct output is always infinity or a natural number \$≤n\$
- Returning two numbers \$(a,b)\$, to represent a fraction \$a\over b\$ where the numerator and denominator are both guaranteed to be \$≤n\$
- Returning two numbers \$(a,b)\$, encoding \$an+b\$, to represent a natural number that is guaranteed to be \$≤n^2+n\$
- Returning three numbers \$(a,b,c)\$, encoding \$an^2+bn+c\$, to represent a natural number that is guaranteed to be \$≤n^3+n^2+n\$
- Returning four numbers \$(a,b,c,d)\$, encoding \$(an+b)\over(cn+d)\$, where the numerator and denominator are guaranteed \$≤n^2+n\$
- Returning two numbers \$(a,b)\$, encoding \$a×10^{b-\lfloor {n/2} \rfloor}\$ (floating point)
- Returning four numbers \$(a,b,c,d)\$, encoding \${a\over b}+{c\over d}i\$ (complex rational number)
- Returning four numbers \$(a,b,c,d)\$, encoding \${a\over b}+{c\over d}\pi\$
- Various combinations thereof, and beyond
So the options are:
- Consider each of the above to be a distinct language, following the established PPCG convention regarding command-line parameters. The N-mover answer would then be unchanged, except perhaps to call the language something like "Regex (ECMAScript, reciprocal output)".
- Represent this metadata in some standard format (with the number of each backreference, and what it represents –
\0could indicate the return match itself), and add it to the byte cost of the regex. Then this would be a new language, focusing on pure regexes and confining the metadata format in a way that honors that restriction.