model-writtenrun 389 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Malis reported three malformed-dictionary defects in spell.c: a CompoundAffix terminator overflow, an uninitialized old-format flag buffer, and NULL accesses through incomplete AF alias tables; the report was moved to the public list because dictionary files are considered trusted input #1.
- Andrey Rachitskiy supplied fixes and AF regression tests in #2; Tom Lane committed the code as cfc720ef4b, adjusting an error message and omitting tests because he did not want intentionally broken dictionaries installed as sample data #3.
- Alexander Lakhin then supplied allocation-failure instrumentation and a reproducer showing that dictionary-cache creation could leave a garbage context pointer #5; Tom identified the invalid intermediate state #6 and wrote 4e68b7e291 to initialize the entry safely and retry context creation after failure.
- Ewan Young introduced a separate compound-flag bug whose symptoms included crashes and silently missing compound lexemes when FLAG appeared late or changed mode #7; his two-patch series deferred conversion until parsing finished and optionally removed redundant per-entry modes.
- Ewan chose deferred conversion instead of rejecting late FLAG lines because AF, SFX, and PFX already use the final file-wide mode, and he included finalization for the old ispell format after discovering that it otherwise crashed #7.
- Tom committed the deferred-conversion fix as b4c8afdfaf, adding corrections for strtol() values stored in narrower integers, and committed the optional cleanup as 2fb8da5a24, with only a minor comment adjustment #8; the bug fix was backpatched, while the cleanup remained development-only to avoid an unnecessary API/ABI change.
- Alexander next reported crashes after forcing Snowball environment allocation to return NULL #9; after Tom requested a more precise diagnosis #10, Alexander and Andrey traced the unchecked result through locate_stem_module() into later lexizing #11 #12.
- Andrey attached a dsnowball_init() NULL check #13, then withdrew its justification after finding that PostgreSQL remaps Snowball malloc calls to palloc, which raises an error rather than returning NULL on allocation failure #14.
- Tom confirmed the palloc linkage and concluded that no Snowball fix was needed #15; the thread ended with all four candidate commits landed and the final Snowball proposal abandoned.