With the L flag, floating-point values use the locale's decimal point (as documented in syntax.md: "inserts the locale's digit grouping characters and — for floating-point values — its decimal point"), but the hexfloat presentation ignores it:

// numpunct with decimal_point '?' and thousands_sep '~'
fmt::format(loc, "{:L}",  1234.5);  // "1~234?5"
fmt::format(loc, "{:Le}", 1234.5);  // "1?234500e+03"
fmt::format(loc, "{:La}", 1234.5);  // "0x1.34ap+10", expected "0x1?34ap+10"

For comparison, std::format(loc, "{:La}", 1234.5) with libstdc++ gives 1?34ap+10 (without the 0x prefix, which fmt always emits).

The hexfloat branch of write wrote the buffer produced by format_hexfloat directly. It now writes it through write_padded and replaces the . with decimal_point<Char>(loc) when the value is localized, using the same FMT_USE_LOCALE && specs.localized() condition as do_write_float. No digit grouping is applied to hex digits, the 0x prefix and the sign-aware zero padding from #4972 are unchanged, and non-localized output is byte-for-byte the same.

Tests: added {:La} and {:+012La} cases to locale_test.localized_double and a wide {:La} case to locale_test.wformat in test/xchar-test.cc, next to the existing tests that use the same numpunct facet. They fail on main ("0x1.34ap+10", "+0x0001.8p+0", L"0x1.8p+0") and pass with the change. The full ctest suite passes with g++ 13 in C++17 (FMT_PEDANTIC=ON, FMT_WERROR=ON), format-test, xchar-test and nolocale-test pass in C++11, and clang-format 21 reports no changes.