fix(vector): harden direct pgvector parity

This commit is contained in:
2026-07-12 01:09:24 +02:00
parent b09341f07e
commit e7e948c77c
3 changed files with 291 additions and 41 deletions
@@ -48,3 +48,28 @@ deprecation warnings.
- The legacy single `connection` form stays read-only through the public port, matching its
previous adapter behavior, while remaining available to the explicitly documented bulk-loader
transition.
## Review fix wave
The Task 1 review findings were addressed in a follow-up TDD cycle:
- Search now validates requested kinds against the global known-kind set, intersects valid kinds
with each collection, and skips unrelated collections. A direct-versus-HTTP parity test covers
the multi-collection case.
- Health requires all three allowlisted tables, an `embedding vector(N)` column on every table,
the expected dimension on every table, and the appropriate read or write table privileges for
each configured side. Empty and partial schemas return deterministic, credential-free details;
unexpected database failures expose only their exception class.
- The Docker L0 fixture now provisions separate least-privilege reader and writer roles. Tests
prove the reader cannot insert, the writer cannot execute the cosine-search SELECT, and the
adapter still routes search to the reader and upsert/hash operations to the writer. Direct
upsert uses an atomic `INSERT ... ON CONFLICT DO NOTHING` followed by `UPDATE` for an existing
key, avoiding broad SELECT authority while retaining conflict-safe hash/upsert semantics.
Fresh verification after the fix wave:
- Docker L0 + HTTP port/search parity: `42 passed` (earlier checkpoint); the final L0 file has
`16 passed` including the stricter raw-role search denial.
- Expanded focused adapter/config suite: `56 passed`.
- Full harness: `466 passed, 5 deselected`.
- Changed-file Ruff lint/format and `git diff --check`: clean.