model-writtenrun 2400 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro opened the thread with #1, proposing obsolete langinfo-probe removal, nl_langinfo_l(), and a thread-safe localeconv wrapper using Windows thread-local state, glibc extensions, BSD/macOS localeconv_l(), or a locked POSIX fallback.
- Heikki Linnakangas approved the first two components and requested clearer localeconv code and encoding commentary in #2; Thomas landed the probe cleanup with an additional pg_upgrade include removal as 14c648ff00 and the nl_langinfo_l() component essentially unchanged as 35eeea6230.
- After reporting macOS build failures in #3, Thomas identified a misspelled xlocale header guard in #4 and corrected it without an attached patch in 2488058dc3.
- Optional MinGW-with-ICU testing exposed missing system collations after the nl_langinfo change, so Thomas proposed sharing the MSVC code-page lookup path in #5, which landed unchanged as bf3401fe81.
- Thomas revised the remaining localeconv implementation in #6 to copy numeric and monetary fields using their respective locales and preserve Windows locale names as wide strings; he also added strftime_l() so the obsolete warning about temporary global-locale changes could be removed.
- The expanded series in #7 added error handling and a check_locale() rewrite that initially preserved environment-derived empty names and Windows canonicalization; Heikki questioned the helper structure and suggested separating validation from canonicalization in #8.
- Drawing on Heikki's review and observations attributed to Peter Eisentraut and Tom Lane, Thomas changed the validation design in #9 to reject empty backend locale names and drop canonicalization, while leaving initdb behavior and CREATE COLLATION compatibility questions outside the patch.
- Thomas removed unnecessary refactoring and reported successful eight-thread Windows localeconv testing in #10; he subsequently reported a request for a better standardized localeconv interface in #11, without treating that as a prerequisite.
- Peter questioned abandoning canonicalization in #12, then reviewed the strftime patch in #13 and requested reuse of report_newlocale_failure() along with more accurate error-reporting commentary.
- After several months, Peter rebased Thomas's series without review corrections in #14, then supplied cosmetic localeconv fixes and the shared strftime error handler in #15, while leaving Windows error handling and per-category validation questions open.
- Peter committed the localeconv component and its fixup from #15 as b98be8a2a2, additionally repairing the BSD feature macro and its error path; his report in #16 explained that the typo had accidentally exercised the fallback implementation and exposed a NetBSD failure.
- Peter committed the strftime component and its error-handler fixup from #15 as 890fc826c9, then posted the remaining check_locale() patch and a new "646" ASCII codeset mapping in #17 after NetBSD testing.
- Andres Freund reported an old-style declaration warning in the newly committed localeconv file in #18 and fixed the declaration order in 116e851db5.
- Peter's further analysis in #19 suggested that NetBSD newlocale() accepted invalid names as the C locale, exposing the missing "646" mapping and potentially accepting names that later setlocale() calls would reject; Tom advocated finishing the removal of setlocale() in #20.
- Peter closed the commitfest entry as done in #21 because the thread-safety objective had landed, deferring the remaining locale-validation work and accompanying codeset patch until related locale and collation changes made further progress.