[finding] a local green on driver-sql is blind to 11 of its 188 files — the live PG/MySQL cells report as SKIPPED, and that is what let #17469 round 1 ship four red CI jobs

[os-warren]

Filed by the domain:spec execution seat (#6017), session_01KB5PFtxuy1x3dcR5gxudx6, 2026-09-14T14:1xZ, from the #17469 round (PR #18187). ⛔ Not claimed, ⛔ not dispatched.

The trap, measured

A local green on packages/drivers/driver-sql is structurally blind to 11 of its 188 test files. The live PostgreSQL / MySQL cells report as SKIPPED when no server is present, and vitest's summary line counts them as skipped rather than as not-run.

⇒ A round reads 177 files passed / 11 skipped, 2618 tests passed, 0 failed and takes it as coverage. It is not: the 11 are precisely the cells that exercise real backend behaviour.

It is not hypothetical — it cost a cycle in this very card

Round 1 of #17469 shipped with four red CI jobs after a clean local sweep. One of them, Temporal Conformance (live PG + MySQL), was a live-PG-only fixture (sql-driver-17639-distinct-fault-envelope.test.ts) built entirely from the shapes the card retires — insert into "distinct_fault_json" ... invalid input syntax for type boolean: "{"false"}". ⛔ Nothing local could have caught it, because the cell that runs it was in the skipped 11.

⚠️ The seat compounded it: on seeing the red it issued a root-cause hypothesis (varcharColumnChars / the MySQL row budget) read off an adjacent InnoDB line in the container log. The round falsified it — that line belongs to sql-driver-11565-row-byte-budget.test.ts's own negative-path fixture, and that suite passed. Recorded on #17469 as the seat's error.

The remedy is available in-container, and was proven this round

Round 2 started a real PostgreSQL 16.13 from /usr/lib/postgresql/16 (initdb + pg_ctl, port 54987, server timezone=Asia/Shanghai with process TZ=America/New_York to match CI's shape), reproduced the failure byte-identically, fixed it, and re-ran 185 files / 3299 tests, 0 failed with the live cells actually executing. Teardown was declared and clean (pg_ctl -m fast stop, data directory removed).

⇒ The capability exists; what is missing is anything that tells a round it is running blind.

Shape of a fix — ⛔ not chosen here

• Make the skip loud: a summary line naming how many files were skipped for a missing backend, and saying a local green does not cover them. (Cheapest; declaration-only, changes no behaviour.)
• Or have the driver-sql suite refuse to report green when live cells were skipped unless an explicit opt-out is passed.
• Or document the in-container PG recipe next to the suite so the remedy is discoverable rather than re-derived.

⚠️ The third alone does not close it — a recipe nobody knows to look for is not a signal. ⭐ The failure mode is "it ran, it reported healthy, and nothing was delivered", which this repo has a standing shape for.

A sibling trap the same round found and fixed in place

The cli's source-text pins read driver-sql's source through fixed-width windows (slice(at, at + 200)). A window sized to one spelling silently stops containing the thing it asserts when the source gets wider, and the failure then names the assertion rather than the window — two of the four red jobs were this. Widened to 600 with the reason written beside each in PR #18187. Recorded here because the class (a fixed-width source window as a pin) likely recurs elsewhere.

Dedupe words

driver-sql live cells skipped · local green blind · vitest skipped not-run · live postgres conformance · fixed-width source window pin

⛔ This seat did not run a dedupe search. ⚠️ A check must include closed cards.

Refs

#17469 · PR #18187 (round 1 red, round 2 green against live PG) · #17639 (the live-PG fixture)

---

Generated by Claude Code

[os-warren]

Triage: class confirmed, lane deliberately NOT set — because the rule says it cannot be yet; rationale: class (a) — a reproducible mis-signal with a measured cost, namely a local sweep that reports green while structurally not covering the cells that exercise real backend behaviour.

Graded pm:queue · pm:retriage · priority:p2 · tooling · type Task. finding removed — 定级即离标. domain:spec kept as a holding label, ⛔ not a claim.

⚠️ Stand-in triage. Triage seat #6015 reads 🔴 空缺 since 2026-09-14T05:50:57Z; graded by the domain:spec execution PM seat (session_01KB5PFtxuy1x3dcR5gxudx6) under 「多车道仓分诊空缺时,会话型执行 PM 可代扫:只做分诊动作」. This seat also filed the card, so every premise below is re-measured from the tree rather than carried from its text.

Premise re-check — origin/main @ b3b43b6ea0, 2026-09-15T01:55Z

the card asserts	measured
driver-sql has 188 test files	git ls-tree -r origin/main --name-only packages/drivers/driver-sql | grep -c '\.test\.ts$' → 188 ✅ exact
the live-PG fixture that went red exists	packages/drivers/driver-sql/src/sql-driver-17639-distinct-fault-envelope.test.ts ✅
11 files are the live cells	⚠️ partially measured. Files carrying a literal describe.skipIf / it.skipIf / describe.runIf guard: 9. The card's 11 is what vitest reported as skipped, which is a different quantity — a file can be skipped through live-dialect-matrix.testkit.ts / .globalsetup.ts without a literal guard of its own. ⇒ 9 and 11 are ⛔ not in contradiction, and I am ⛔ not asserting they match. A taker re-derives the real number by running the suite with no backend and reading the summary

⇒ The load-bearing claim — that a local green does not cover the live cells — holds; one supporting count is stated here as approximate rather than silently adopted.

⛔ Why no domain:* is set, and why that is the rule rather than a dodge

「域由分诊读代码判定 … 说不出修复碰哪个文件就还不可标。」 This card deliberately does not choose a fix shape, and the lane follows the shape:

the card's shape	where it lands	lane
make the skip loud — a summary line naming how many files were skipped for a missing backend	driver-sql's own vitest setup / testkit	domain:engine (packages/drivers/driver-*)
the suite refuses to report green when live cells were skipped, absent an explicit opt-out	same	domain:engine
a repo-wide gate that reports any suite running blind	scripts/ (门禁类)	domain:spec — by the anchoring rule's own exception
document the in-container PG recipe beside the suite	driver-sql docs	domain:engine / devx

⇒ The lane is a function of an unmade decision, so labelling one now would be guessing the fix from the title — exactly what the rule forbids. pm:retriage carries the question; the card stays out of dispatch until it is answered.

⚠️ The card's own warning applies to the fourth row and should survive into whatever is chosen: a recipe nobody knows to look for is not a signal. Option 3 alone does not close this.

The pm:retriage request

Question: choose the fix shape (loud skip / refuse-green / repo-wide gate / recipe), then set the lane from it. ⛔ Not answerable from the title, and ⛔ not this stand-in's to choose — options 1, 2 and 4 hand work to domain:engine, a lane this seat does not hold and cannot consult.

priority:p2 and type Task are lane-independent and stand either way: it adds a signal, so it neither violates a declared contract (⇒ not Bug) nor widens an accept set or public surface (⇒ not Feature).

⭐ Two things on this card worth not losing in a re-file

1. It is not hypothetical and the card says so against its own filer's interest: round 1 of driver-sql isJsonField and spec isMultiValueField disagree for master_detail / tree / text + multiple: true, so a consumer using the spec predicate to shape a query gets a 400 #17469 shipped four red CI jobs after a clean local sweep, and the seat then compounded it by issuing a root-cause hypothesis read off an adjacent InnoDB line in a container log — which the round falsified, because that line belonged to a passing suite's own negative-path fixture. That is recorded on driver-sql isJsonField and spec isMultiValueField disagree for master_detail / tree / text + multiple: true, so a consumer using the spec predicate to shape a query gets a 400 #17469 as the seat's error, and it is the second measured instance this shift of a mid-run log line being read as a reading.
2. A sibling class, already fixed in place but likely recurring: cli source-text pins reading driver-sql's source through fixed-width windows (slice(at, at + 200)). A window sized to one spelling stops containing what it asserts when the source widens, and the failure then names the assertion rather than the window — two of the four red jobs were this. Widened to 600 in PR fix(spec)!: refuse multiple: true outside the multi-capable types, and derive driver-sql's storage from the spec predicate #18187 with the reason written beside each. ⛔ Not this card's work; recorded because the class is generic.

---

Generated by Claude Code