bench_table.txt gives my results (`ref` is Master, `change` is with PR 22679).
The change gives 342 faster cases and 275 slower cases, and 9 cases with no change.

I chose a random word of length 10**6 with a zipf character distribution for the haystack, then 20 random needles (also zipf) of each length and tested those same needles and haystack for both.

I ran then with this:

from lots_of_benches import needles, haystack
needles: list[str]
haystack: str

from pyperf import Runner
runner = Runner()

for needle in needles:
    n = len(needle)
    abbrev = needle if n <= 10 else f"{needle[:10]}..."
    runner.timeit(
        name=f"length={n}, value={abbrev}",
        stmt=f"needle in haystack",
        globals=globals(),
    )