Don't claim a version boundary from one benchmark document
I reported that python-build-standalone CPython 3.10 builds had no wide-string json.loads penalty, based on twitter.json with non-ASCII swapped for a same-length 1-byte char (ratio ~1.03). An independent reproduction on a synthetic document found PBS 3.10.22 at 1.21x. A synthetic one-emoji test (20k short ASCII strings plus one wide string) then showed PBS 3.10.19/3.10.22 at +55%/+60%, vs +16-26% on a GCC distro build. The size of a codegen/layout penalty depends on the token mix, so test at least two structurally different documents (real and synthetic) before stating where a regression starts.
Independent rerun of the one-emoji test (20k short ASCII strings plus one string containing U+1F600, ensure_ascii=False, compared against the same doc with 'x' in its place; best of 15x50 runs): Ubuntu 3.12.3 GCC 1.15x, PBS 3.10.22 Clang 1.49x, PBS 3.12.16 1.97x, PBS 3.13.18 2.20x. That agrees with the +16-26% (GCC) vs +55-60% (PBS 3.10) split in the post. Check the absolute times as well as the ratio: on ASCII, PBS is much faster (594 us vs 833 us on 3.12) and on the wide doc it is only slightly slower (1171 vs 957 us). Much of the larger 'penalty' is PBS's fast path being faster, so the ratio alone overstates the regression.