benchmarks: the PostgreSQL-family and QuestDB arms use the vendor's documented client paths (autocommit, binary COPY, binary cursors) - #251

Closes #250. Changes only benchmarks/experiments: l1_tpc.py, l3d_dense.py, e2_hybrid.py, l4_tsbs.py, CAMPAIGN.md (section 7 row 76), and PROTOCOL.md (section 7, three rows).

PostgresTPC (and postgres_tuned) runs the timed calls in autocommit, set after the final commit of build(), which stays one transaction. The four CRUD calls and the analytics reads send one message each. New-order and payment keep BEGIN, three statements, COMMIT through with self.cx.transaction():. psycopg pipeline mode is not used (measured slower, 0.82x and 0.88x).

PgVector.connect and PgAgeE2.__init__ call register_vector(self.cx), and the dense and product loads use binary COPY with cp.set_types([...]) and numpy rows. The pgvector import is lazy, inside those two methods, so a stage that never touches pgvector never needs the package. The sparse pgvector load is untouched (1.09x, a null). QuestTS reads go through a _q helper that passes binary=True.

Effects, from the G3 laptop A/B (12 interleaved repetitions, E cores 8-11): CRUD calls 1.83x to 1.96x faster, pgvector ingest column 9.1x to 13.2x faster, QuestDB q_high 1.34x and the other five reads null. None of it flatters ArcadeDB. The 1.9x CRUD figure still needs a mini confirmation before it is printed.

Safety for the live campaign: tier 1 (ArcadeDB only) stages do import these four modules, so the check before merging is real, but nothing changed at module level (no new top-level import, constant, or call): only the bodies of the PostgresTPC, PgVector, PgAgeE2, and QuestTS methods changed, and no ArcadeDB arm runs them. build_images.sh is not touched. The pgvector==0.5.0 and neo4j-rust-ext==6.3.1.0 pins are tracked as a pending TODO row (76) in CAMPAIGN.md section 7 and land with the tier 1 to tier 2 image rebuild. Until then a pgvector_dense or pg_age_e2 cell on the current client image fails at connect with an ImportError, loudly and by design.

Gates run: compile() on all four touched modules; import of l1_tpc, l3d_dense, e2_hybrid, l4_tsbs in a venv with psycopg 3.3.6 and pgvector 0.5.0; the full benchmarks/experiments pytest suite, whose failure and error set (1 failed, 45 errors, 580 passed, 20 skipped) is identical on origin/main f6190b2 because those tests need results/ and .notes data that a worktree lacks; pre-commit hooks; comparator_pins_check, structure_check, and version_pin_check identical to origin/main. Laptop smoke through the lane classes against real PostgreSQL 18.6 (digest 7341002d), pgvector 0.8.6 (1d50c689), dbbench:pg-age, and QuestDB (931af415), origin/main against this branch: documents (crud, orders, stock, payments, q6) digests identical; dense 20,000 SIFT vectors stored rows and search digests identical; e2 5,000 products stored rows and view counts identical (the retrieval top-k differs between two HNSW builds of the same table, and origin/main against itself differs the same way); QuestDB 300,000 points, six answers identical.

Merge only after checking that no stage on mini is importing the touched modules at that moment.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ubz3m6QXHESerCkehC1qgi