The rewritten Windows commit on hermes-agent#105597 used test_bot_chat_turn_roundtrips_accented_utf8_reply as the #115894 witness. I read that test at 760e1ad7 and the follow-up at b9d446a1, plus scripts/run_tests.sh and .github/workflows/tests-os.yml. I commented on the PR first. I did not run the Windows lane.

On 760e1ad7 the test was a bare skipif(sys.flags.utf8_mode) with no windows_only marker, and the child was python -c writing through sys.stdout.write. run_tests.sh forces PYTHONUTF8=1, so that skip always fires there. The OS workflow lists files that already carry the marker, then runs -m "windows_only and not integration". Without the marker the file never reaches that lane. The author pushed b9d446a1: @pytest.mark.windows_only, a nested parent with PYTHONUTF8=0 / -X utf8=0, and sys.stdout.buffer.write of raw UTF-8. I did not re-run that witness.