My best guess is something like this, where loc is the formatting locale ([time.format] p2):
if (narrow literal encoding is UTF-8)
if (locale_t cloc = ::newlocale(
loc.name()))
if (const char* enc = ::nl_langinfo_l(CODESET, cloc))
if (/*enc is not UTF-8 */) {
iconv_t ic = ::iconv_open("UTF-8", enc);
if (ic != (iconv_t)-1) {
// use ::iconv to convert from locale's encoding to UTF-8
But that seems pretty involved ... and {fmt} doesn't do any of that.
I tried testing the example from the paper on Linux, and fmt::format fails with an exception. Debugging it shows that it tries to use the formatting locale's std::codecvt<char32_t, char, mbstate_t> facet to convert the string. But that's not right, because that codecvt specialization is defined by the standard to convert between UTF-8 and UTF-32 only. So it can only work if the input is ASCII, or the locale uses UTF-8, in which case there's nothing that needs converting anyway.
AFAIK the standard doesn't provide a way to convert from an arbitrary locale's encoding to the execution charset, or even to get the name of an arbitrary locale's encoding (C++23 provides a way to get the name of the execution environment's encoding, but not an arbitrary std::locale's encoding).