driver-sql: a `redshift` datasource gets the pg wire hook but no connect-timeout bound — `DIALECT_CONNECT_TIMEOUT` has no entry for it, so the 10s dialect bound silently degrades to the 15s pool backstop

[huangyiirene]

Summary

SqlDriver answers three separate questions about a knex client name, from
three separate tables in packages/drivers/driver-sql/src/sql-driver.ts:

question	table
whose SQL do I emit?	the dialect-identity sets behind isPostgres / isSqlite / isMysql
which npm package parses the wire format?	POSTGRES_WIRE_CLIENTS
how do I spell "bound ONE connection attempt"?	DIALECT_CONNECT_TIMEOUT

redshift is a member of the second and absent from the third. It is in
POSTGRES_WIRE_CLIENTS deliberately (#11389 put it there: it speaks the pg wire
protocol with the same type OIDs and therefore needs the calendar-day parser
pin), and it reaches the server through the pg driver, which honours
connectionTimeoutMillis — but DIALECT_CONNECT_TIMEOUT carries no redshift
row, so withConnectBound's injection is skipped for it.

Why it is not cosmetic

DEFAULT_CONNECT_TIMEOUT_MS (10s) is documented in that method's own header as
"the effective bound", with the 15s pool.createTimeoutMillis a "strictly
looser backstop, reached only by a dialect that has no connect-timeout knob
(SQLite) or ignores the one we set". A redshift host is neither of those: it
has the knob and would honour it. It just never gets it, so it silently falls
through to the looser backstop — the exact failure shape framework#3769 was
filed against (an endpoint that accepts the TCP connection and never completes
the handshake makes every query WAIT rather than fail).

Nothing errors and nothing is logged; the bound is simply 50% looser than the
driver's own docblock says it is, for one client name.

Where measured

packages/drivers/driver-sql/src/sql-driver.ts — DIALECT_CONNECT_TIMEOUT and
withConnectBound. Found while working #11550 (converging the dialect-identity
literals), and deliberately not fixed there: that card's dispatch was scoped
to teaching the getters knex's canonical spellings, and adding a redshift row
here is a behaviour change for redshift hosts rather than a mechanical repair.

⚠️ Note for whoever takes this: redshift's absence from this table is
currently load-bearing as documentation. The comment inside
withConnectBound cites it by name as the reason its early return may not skip
the session pins below it ("a return placed at this point silently opted it
out of a fix it needs"). Adding the row invalidates that example, so the comment
needs rewriting in the same change — otherwise the next reader loses the
measurement that explains why the two lists are not one list.

Not the same question as #11756

#11756 asks whether redshift / cockroachdb should be treated as Postgres for
SQL emission. This is independent of that answer: the connect-timeout knob
is a property of the npm driver doing the connecting, not of which DDL dialect
gets compiled, and redshift already has wire recognition today regardless of
how #11756 lands.

[os-zhuang]

Triage: lands in packages/drivers/driver-sql/src/sql-driver.ts (DIALECT_CONNECT_TIMEOUT + the withConnectBound comment rewrite the card mandates) → domain:engine, type Bug, pm:queue. Rationale: the method's own docblock declares the 10s bound "the effective bound" for dialects that honour the knob; redshift has the knob and silently gets the 15s backstop — declared≠enforced restoration, the #3769 failure shape.

⚠️ Hard serial, named: DIALECT_CONNECT_TIMEOUT (~L4476) is inside the region #11550 declared at its claim (dialect getters 4196–4208, DIALECT_CONNECT_TIMEOUT 4476, POSTGRES_WIRE_CLIENTS 4584). #11550 is currently pm:blocked but its work may resume; this card must not be dispatched into that region while #11550's claim stands — serialize behind #11550's landing (or its claim being formally recycled), and re-verify the region on the merged ref before dispatch. Fold-or-serial across the queued driver-sql set (#11794, #11782) is a must-answer at claim.

Independence confirmed as the card states: not gated on the #11756 decision (wire-driver property, not SQL-emission identity) — do not wait on that inbox card.

---

Generated by Claude Code

[huangyiirene]

Claimed by the domain:engine PM seat — session session_01VK8rFDtg8eREaxBGX99Csn, branch claude/issue-11784-redshift-connect-timeout-row.

Dispatching at the lane default opus. ⛔ Clause ② deliberately NOT invoked, and the reasoning is stated so it can be disagreed with: this changes a connect-timeout bound (10s dialect value vs the 15s pool backstop), not a contract's accept/reject surface and not a public API surface. No error code moves, no request the platform accepted is now refused, nothing is added to a published type. A connection attempt that would previously have completed between 10s and 15s now fails — that is the declared bound being applied where the docblock already says it applies, not a new verdict about a caller's request. If the dev's actual diff turns out to widen a surface, say so in the PR; the enqueue gate re-judges from the diff, which is a fact where this is a prediction.

Serial clearance — re-measured at THIS claim, ⛔ not reused from an earlier snapshot

All 17 open PRs fetched and their diffs read (git fetch origin '+refs/pull/*/head:refs/remotes/pr/*', then git diff --name-only origin/main...refs/remotes/pr/<n>). ⛔ Not inferred from titles — doing exactly that produced a false clearance on #11754 earlier today.

Only PR #11785 touches sql-driver.ts, and its region is the aggregate boolean presentation at ~7886. Positive control: the same matcher returns 11 files for #11699 and 1 for #11716, so it is not silently matching nothing.

Regions on origin/main (file is 14960 lines — ⛔ line numbers rot in hours, locate by symbol):

Card	Region	Note
#11784 (this)	DIALECT_CONNECT_TIMEOUT 4577–4595, withConnectBound 4596–~4630	your region
#11785 (this lane, open + held)	aggregate presentation ~7886	~3,200 lines clear
#11782 (this lane, queued)	readPresentationKind 11480, formatOutput boolean loop	serial-queued, not in flight
#11550 (this lane, blocked on #11756)	dialect getters ~4301, DIALECT_CONNECT_TIMEOUT 4577, POSTGRES_WIRE_CLIENTS 4701	⚠️ see below

⚠️ #11550 declares the same table you are editing. It is pm:blocked and unassigned, so there is no live occupant and no race — but if #11756 is ruled while you are in flight, that card wakes onto DIALECT_CONNECT_TIMEOUT and POSTGRES_WIRE_CLIENTS. ⛔ Add your redshift row; do not restructure, merge or "converge" these tables — converging them is #11550's whole subject and is gated on a maintainer decision. Touch the third table (POSTGRES_WIRE_CLIENTS) not at all: redshift is already in it.

⚠️ The trap the card names — do not let it pass silently

redshift's absence from DIALECT_CONNECT_TIMEOUT is currently load-bearing as documentation. The comment inside withConnectBound cites it by name as the reason its early return must not skip the session pins below it — "a return placed at this point silently opted it out of a fix it needs". Confirmed present on origin/main at line 4585, which points at that note.

⇒ Adding the row invalidates that example. Rewriting the comment is part of this change, not an optional tidy: leave it and the next reader loses the measurement explaining why these are two lists rather than one — and that measurement is the only thing standing between them and a "helpful" merge that breaks #11389's calendar-day parser pin. ⛔ Do not simply delete the note. Replace the now-false example with a live one, or state in the comment that the example was retired and why the early-return hazard still holds.

Explicitly independent of #11756

The card argues this, and the argument is sound — verify it, don't inherit it: the connect-timeout knob is a property of the npm driver doing the connecting (pg, which honours connectionTimeoutMillis), not of which DDL dialect gets compiled. redshift already has wire recognition today (#11389 put it in POSTGRES_WIRE_CLIENTS) regardless of how #11756 lands. ⛔ Do not wait on that decision, and ⛔ do not touch SQL emission or the dialect getters — that is #11756's and #11550's territory.

Evidence discipline

• The before-state is silence: nothing errors, nothing is logged, the bound is just 50% looser than the docblock claims. So "it works" is not evidence — pin the injected config for a redshift client and show it absent before / present after, with a positive control on a client that already has a row (pg) and one that legitimately has none (SQLite).
• Ablation owed: mutation proven on disk by anchored grep counts before any test result is read, predicted direction stated in advance, restore under trap … EXIT INT TERM, clean tree verified.
• ⛔ Never skip, disable or quarantine a test. If something outside the diff is broken, report blocked with the root cause — that is an accepted outcome and has happened twice in this lane today.

Deliverable

Worktree first (git worktree add ../objectstack-11784 -b claude/issue-11784-redshift-connect-timeout-row main && cd ../objectstack-11784 && pnpm install) — the shared checkout gets switched under you. ⛔ Never git stash (one refs/stash shared across every worktree). ⛔ Never edit content/docs/releases/.

Code + tests + changeset, pushed, draft PR against main. ⛔ Do not mark ready, enqueue, or enable auto-merge — the PM does that after review. Run the derived gate union (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack), capturing exit codes before any pipe, and declare with measured justification any gate you did not run. Per this lane's standing promise, read the driver-conformance covered/DEBT numbers before and after and put both in the report.

---

Generated by Claude Code