model-writtenrun 1094 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Robert Haas traced a buildfarm failure to whole-row and expression indexes accessing before the beginning of a TupleDesc, and proposed three patches in #1: guards suitable for backpatching to PostgreSQL 18, HEAD-only bounds assertions with caller cleanups, and a whole-row index regression test.
- Tom Lane endorsed all three patches in #2 and explained in #4 that his own assertions confirmed access through attribute number zero; he accepted that the HEAD-only assertion might expose additional missed callers.
- After Tom reported another occurrence during pg_upgrade regression testing in #6, Robert confirmed the index fix was pushed in #7; cceb9c18a5 preserves the first patch from #1 unchanged apart from branch context.
- The remaining patches from #1 landed unchanged on March 24 as the bounds assertion and caller cleanups in c98ad086ad and the whole-row index test in 4647ee2da3.
- Alexander Lakhin opened a follow-up bugfix round in #8 by showing that COPY FROM with a WHERE condition referencing tableoid triggered the newly added bounds assertion.
- Tom initially suspected whole-row COPY references in #9, withdrew that suspicion after checking the existing handling in #10, and concluded that system columns should be rejected because they are not ready when filtering occurs.
- Tom supplied a new HEAD patch in #11 rejecting all system columns, including tableoid because its current value identifies the COPY target rather than the destination partition; he proposed preserving stable-branch behavior with a simple access guard.
- Peter Eisentraut supported prohibition and pointed to earlier unresolved work in #12; at the thread's final message, #13, Tom deferred changing tableoid semantics and favored an intervening period of rejection before any future reintroduction.
- On April 6, 21c69dc73f landed the HEAD patch from #11 exactly, while 40fa04e7c8 implemented the separately discussed access guard for PostgreSQL 14–18; all attached patches were then landed, with richer tableoid semantics left for future work.