fix(vector): harden packaged migrations

This commit is contained in:
2026-07-12 01:32:33 +02:00
parent c0d50e9b08
commit 3588a7749b
14 changed files with 337 additions and 54 deletions
@@ -50,3 +50,33 @@ No unresolved Task 2 correctness concern found. One deliberate contract choice i
writer `INSERT` and `UPDATE` are table-level because the approved direct adapter health probe uses
`has_table_privilege` for those authorities. Least privilege is retained by withholding broad
`SELECT`, `DELETE`, DDL, ownership, and credentials.
## Review fix wave
The post-implementation review found four production-boundary gaps. They are fixed as follows:
- Migration SQL now ships inside the `tht` wheel (`tht/migrations/vector`) via explicit
setuptools package-data and is discovered through `importlib.resources`, rather than relying on
a source-checkout-relative directory.
- Both status and apply reject ledger versions absent from the installed manifest, including
nonnumeric future version labels. This treats a binary/database downgrade as drift instead of
silently reporting a healthy state.
- Migration files are ordered by parsed integer version; spellings such as `2` and `02` are
rejected as duplicate versions.
- Every migration transaction pins `search_path` locally to `pg_catalog, pg_temp`; catalog calls
and the ledger are schema-qualified. pgvector is installed into the locked `vectors` schema,
tables use `vectors.vector`, and `PgVectorStore` qualifies vector casts and the cosine operator.
A hostile admin default path with a writable shadow schema cannot redirect migration objects.
- The core image build asserts CLI discovery. Image verification now starts an ephemeral pgvector
database, runs the installed image's migration command, and compares pristine apply/status JSON.
Additional verification after the fix wave:
- Focused migration, adapter, hostile-path, and wheel suite: `27 passed`.
- Full harness: `477 passed, 5 deselected`.
- Production core image build: passed, including build-time CLI discovery.
- Core-image apply/status smoke against `pgvector/pgvector:pg16`: passed.
- Changed production and test files: Ruff clean; `git diff --check` clean.
- Full Ruff remains at the same 34 pre-existing unrelated test-file findings documented above.
No dependency changed, so the committed Python requirements lock did not require regeneration.