[thekevinbot]

Refs #1562

Root cause

A real fixed per-change overhead, plus a noisy 15-pair measurement. It is not a recent regression: commits back to 2026-10-14 fail the same bar on this host.

• On main, 300-pair raw runs: native median 388us, dirsql 435-439us, ratio 1.12-1.14 across 3 runs. That is about 48us of fixed cost per change.
• Tick breakdown of one change on main, in us after the write: hook spawn 36, cat EOF 225, child reaped 247, reconcile done 313, consumer has the event 337.
  • The reconcile took ~56us live. row_key alone took ~17us, because it ran three pragma queries on every rerun.
  • The range refit (DELETE, min/max query, INSERT) took ~14us.

Change (packages/rust/src/db.rs)

• A table's key candidates are cached per PRAGMA schema_version, so any schema change invalidates them.
• get_table_columns uses a cached pragma_table_info(?1) select.
• The range refit is skipped when every change is an in-place update.
• Otherwise the range is fitted from the rowids the reconcile placed, with no min/max query.

After the change, reconcile is ~18-21us live (row_key is ~0.5us).

Measurements (same host, interleaved runs, bar = 1.1x, lower is better)

	main	this branch
e2e passes, round 1	1/8	5/8
e2e passes, round 2	3/10	7/10
e2e passes, round 3	6/10	7/10
typical bar median	~1.13	~1.07
side-by-side A/B, both orders, 400 pairs x 3		branch/main ~0.95

The test is still not reliably green on a loaded host: the same binary ranges 0.95-1.34 from run to run. The remaining ~20-25us gap is structural:

• waiting for the hook child to exit after its stdout EOF: ~21us, which is cat's own exit cost
• two thread/task hops before the SSE write: ~9us

Things tried with no measurable gain, reverted: prepare_cached for the transaction statements, translating watcher events on the reading thread, reaping the child on the stderr thread (worse), and spinning on try_wait after EOF.

Changelog / Migrations

• Changelog fragment added under <root>/<pkg>/changelog.d/ for each changed package (or: skip-changelog trailer on a commit with reason)
• Migration fragment with the same slug added under <root>/<pkg>/migrations.d/ for each changed package (every subsection but Summary _None._ when nothing breaks)

E2E Verification

• Ran e2e suites locally for every affected SDK
• Python SDK e2e: pass
• TypeScript SDK e2e: pass
• Rust core e2e (if applicable): N/A (cargo test --lib: 1299 passed)
• packages/python/e2e-attestations/<branch>.json receipt written if packages/python changed (just e2e-attest-python)
• packages/ts/e2e-attestations/<branch>.json receipt written if packages/ts changed (just e2e-attest-ts)
• Command(s) run:
  • testing-conventions e2e attest 'uv run python -m pytest tests/e2e/on_file_query_test.py tests/e2e/query_contract_test.py tests/e2e/declared_table_name_test.py tests/e2e/fanout_test.py tests/e2e/use_case_dossier_test.py tests/e2e/use_case_join_test.py tests/e2e/use_case_notebook_test.py tests/e2e/use_case_transcripts_test.py tests/e2e/watch_speed_test.py -q'
  • just e2e-attest-ts
• Result summary:
  • Python: 22 passed. watch_speed_test passed on that run, but it passes only ~70% of runs on this host (see above).
  • TypeScript: 18 passed.