# Local pgvector Task 2 report ## Outcome Implemented ordered, idempotent production migrations and the `tht vector migrate` interface, including `tht vector migrate --status --json` with pristine JSON output. ## Implementation - `001_extensions.sql` installs pgvector. - `002_schema_tables.sql` creates `vectors.schema_records`, `vectors.evidence`, and `vectors.memory` with the `VectorWriteRecord` columns and `vector(768)` embeddings. - `003_roles.sql` creates passwordless `NOLOGIN` reader/writer roles. Deployments inject credentials (or grant these roles to separately-created login roles); no production secret is stored in the repository. - Reader authority is schema usage plus table `SELECT`. - Writer authority is schema usage, table `INSERT`/`UPDATE`, narrow hash-probe column `SELECT`, and sequence `USAGE`. It has no `DELETE`, broad row `SELECT`, DDL, or ownership authority. - The migration runner discovers ordered SQL files, records SHA-256 checksums in `public.tht_vector_migrations`, serializes runners with a transaction-scoped advisory lock, and applies the full pending batch in one transaction. - Status distinguishes applied, pending, and checksum-drifted migrations. Apply refuses drift. A failed migration rolls back both prior migrations in that batch and ledger writes. ## TDD evidence RED was observed with a real `pgvector/pgvector:pg16` testcontainer: 6 failures for the missing module, missing command, and missing schema. GREEN verification: - Focused migration + direct adapter integration: `23 passed`. - Full harness from the documented `harness/` cwd: `473 passed, 5 deselected`. - Targeted Ruff (`tht` plus the new L0 test): clean. - `git diff --check`: clean. The new L0 coverage exercises clean install, idempotent rerun, pristine JSON status, checksum drift, transaction rollback, exact tables/columns/dimensions, role isolation, sequence authority, and the real `PgVectorStore.health()` plus `VectorWriteRecord` upsert path. ## Existing repository lint baseline The requested full `ruff check .` was run. It reports 34 pre-existing violations in unrelated test files (unused imports and one-line semicolon statements). None are in Task 2 files; changing them would exceed this task's scope. The complete harness test gate is green. ## Self-review No unresolved Task 2 correctness concern found. One deliberate contract choice is worth noting: 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.