The cut miss on --output-delimiter -1 is the clap vs GNU getopt split. getopt_long takes the next argv as a required option-argument even when it starts with -. Clap treats -1 as a new flag unless allow_hyphen_values(true) is set. The portable spelling is --output-delimiter=-1. A probe that only passes the delimiter as a separate argv will keep reporting that, which is fair, but it should also emit the = form so you can see which implementations accept both.

Two cases a --help probe usually never generates, and where rewrites still drift after the flag-alone matrix is green: attached short options (-d, vs -d ,) and a -- end of options before a filename that starts with -.

On base64 -d, the partial write matters more than the byte mismatch. GNU emits the bytes it managed to decode and then exits non-zero, so base64 -d <bad | downstream without pipefail keeps going on truncated output. If the diff only compares stdout when the exit status matches, that case disappears.

Sort exiting 1 on a usage error instead of 2 is the one I would block a swap on. GNU sort is 0, 1 for --check disorder, and 2 for a real error. sort ... || [ $? -eq 2 ] then treats a bad invocation as "out of order" once the binary is uutils.