model-writtenrun 9313 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Alexander Lakhin opened the thread with an ECPG failure after make clean and proposed using installed ECPG binaries and libraries for installcheck in #1.
- Alexander expanded the proposal into conditional USE_INSTALLED_ASSETS handling for ECPG, libraries, and regression drivers in #2, then corrected header selection and PGXS propagation in #3.
- Robert Haas argued that installcheck should use build-tree assets against the running server to avoid mixing installations in #6, while Tom Lane argued for testing installed client programs as well as the server in #7.
- Alexander explained the binary-package testing use case in #8, but Tom questioned running installcheck after cleaning the build tree because rebuilding could introduce version skew in #11.
- Peter Eisentraut reproduced a separate missing utils/errcodes.h failure in #13, and Tom fixed generated-header ordering in 7dc5a96aa2, announcing in #15 that the original ECPG problem remained.
- Tom traced the ECPG failure to a pattern rule whose build-tree ecpg prerequisite could not be built and requested a narrower patch in #16; he also identified stale test binaries when switching between check and installcheck and suggested a separate build target in #19.
- Alexander retained the broader installed-asset design in #20, extending dependency handling and proposing that only test artifacts without installed equivalents be built locally.
- Michael Paquier proposed rejecting the behavioral change in #21, but Alexander objected that this overlooked the later discussion in #22.
- Tom's detailed review in #23 requested documentation, simpler Makefile conditionals, correct invocation from the ECPG test directory, and justification for changes outside ECPG.
- Alexander responded with a substantial rework in #24 that used installed pg_regress, centralized library-build suppression, covered all procedural languages, and documented package-testing semantics; building regression shared libraries remained necessary because regress.so was not installed.
- Dmitry Dolgov reproduced a pg_upgrade test failure caused by locating pg_regress outside the temporary installation in #25, which Alexander fixed by passing DESTDIR in #26 before rebasing onto subsequent test infrastructure changes in #27.
- After Andres Freund questioned an erroneous ready-for-committer status in #31 and Michael corrected it in #32, David Steele warned that the patch still lacked reviewer or committer interest in #33.
- Alexander supplied a final rebase but requested withdrawal for lack of review interest in #36; Tom confirmed withdrawal in #37, leaving only the separate generated-header fix committed and the broader installcheck redesign set aside pending agreement on its need.