model-writtenrun 7026 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Joseph Koshakow proposed making date_trunc() reject invalid units even for infinite inputs, matching finite-input behavior and date_part(), and supplied implementation and tests in #1.
- Michael Paquier agreed with validating supported units in #2 and reaffirmed approval in #3, deciding against backpatching because previously successful queries would begin failing.
- Michael Paquier committed #1 unchanged as d85ce012f9 and announced completion in #4.
- Peter Eisentraut reopened the issue in #5 with missing regression tests and a fix for timestamptz_trunc_internal(), whose Datum-return macro could return a memory address interpreted as a timestamp on 32-bit systems.
- Michael Paquier reproduced the 32-bit failure, checked the missing test coverage, and took responsibility for the PostgreSQL 18 open item in #6.
- Michael Paquier applied the return-value fix from #5 unchanged as 2242b26ce4, then committed its tests with an additional explicit-timezone infinity case as 572c0f1b0e; both were backpatched to PostgreSQL 18.
- In #7, Michael Paquier explained that he applied the fix before the tests to avoid intermediate failures during 32-bit bisects, and Joseph Koshakow acknowledged the completed fixes in #8, leaving no outstanding work identified in the thread.