skills(query): optimization flight — seven twice-stated rules merged into rules/, removed-key history collapsed to one table, search/flow-filter/compareTo pointed at their anchors, context and the calling convention taught (net −2,139 tokens)

[os-litant]

Part of #14302 — the eval items (QRY-D-05 / H-02 / H-03) are deferred to #14296, so that card stays open.

Member of the skills catalog optimization program #14292. Governed surface (skills/**) ⇒ this PR stays DRAFT; needs:contract-review is on this PR and on #14302.

Audit record: posted verbatim as a comment on #14302 (read back byte-identical). It is the finding-by-finding spec this PR implements.

Head sha 49de20f (review round 1 + the role-word baseline ratchet, then merged with current main).

Files outside the declared surface

One file beyond the four authored skill files and the one generated doc:

file	why	prescribed by
scripts/role-word-baseline.json	check:role-word reds on improvement as well as regression. Round 2 deleted two Common-Mistakes blocks in rules/filters.md, removing one baselined occurrence, and the gate refused: role-word count improved 9 → 8 — ratchet DOWN: run node scripts/check-role-word.mjs --update and commit the baseline.	the gate's own message. --update moved exactly one row (skills/objectstack-query/rules/filters.md: 9 → 8) and no other, verified by git diff before committing; it is its own commit, 8208cf30. The data and api flights took the same path.

No ceiling was changed, in this file or any other.

Review round 1 — both FAIL spans, measured and fixed

Round 1 passed everything except two spans. Both were re-measured at source before being touched; one was a real defect, one turned out to be correct as written and is now stated more precisely.

1. ENGINE_FIND_OPTION_KEYS was under-reported — REAL DEFECT, fixed

The calling-convention table listed nine keys for engine find/findOne and the prose called the set closed. packages/objectql/src/engine.ts:396-400 spreads ENGINE_DRIVER_PASSTHROUGH_KEYS into it, and that constant (:368-370) is transaction, tenantId, tenantIds, timezone, bypassTenantAudit, preserveAudit. As written the table therefore told a reader that an explicit tenantId is refused by name, which is false.

Fixed: the find/findOne row now names all fifteen keys, and the prose carries the reason the source itself gives at :360-366 — on find/findOne (and update/delete) the option bag IS the base of the driver options, which is how an explicit tenantId reaches the driver; count and aggregate never forward the bag, so on those two the same keys are deliberately ILLEGAL. The count (context, where) and aggregate (six keys, no passthrough) rows are unchanged — both were already exact, and the source comment says the omission there is deliberate.

2. The $like/$ilike face list — VERIFIED AT SOURCE, kept and sharpened

Review offered two routes: verify each named face with file:line and the exact message, or reduce the sentence. Taking the first route: all three refusals hold. Each is an allowlist miss, which is why grepping the drivers for the literal $like does not find them.

face	allowlist that omits them	refusal site	code / status	message
driver-mongodb	— (explicit default: arm)	packages/drivers/driver-mongodb/src/mongodb-filter.ts:1071-1074	unsupportedFilterError at :366-371 sets INVALID_FILTER / 400	Unsupported filter operator "$like" on field … It is refused rather than passed through to MongoDB
objectql having	packages/objectql/src/having-filter.ts:136-141 — CONDITION_OPERATORS, sixteen operators, no $like/$ilike	unknownOperator at :174-205 → invalidFilterError	INVALID_FILTER / 400 (the doc table at :157-158 enumerates the faces sharing that envelope)	Unsupported operator '$like' in having. … An unknown operator is refused rather than ignored
service-analytics	packages/services/service-analytics/src/strategies/filter-normalizer.ts:418-435 — MONGO_TO_CUBE_OP, no $like/$ilike	miss branch at :1032-1046	invalidFilterError at :403-406 sets INVALID_FILTER / 400	[analytics] Unsupported filter operator "$like" on … Dropping it would silently widen the query

packages/spec/src/data/filter.zod.ts:697 states exactly this, so the spec .describe() is accurate and no spec-side twin card is owed.

Two readings in the review did not survive measurement, so the sentence is corrected rather than weakened:

• driver-memory ANSWERS $like/$ilike. packages/drivers/driver-memory/src/memory-driver.ts:58 declares its vocabulary as [...FILTER_OPERATORS, '$like', '$ilike'], and the arms are at memory-driver.ts:1522 and memory-matcher.ts:458.
• filter-refusal.ts is not an operator refusal here. The hasDanglingLikeEscape import at :36 gates a MALFORMED PATTERN — a lone unpaired trailing backslash — at :720-726, using the spec's shared test so every face refuses the same patterns. It does not refuse the operator.

The prose now states which faces answer, which refuse, and the dangling-escape rule that binds all of them. Paid in-file: two Common-Mistakes blocks in rules/filters.md that restated the Logical Operators and $null sections directly above them — the deletion that then required the baseline ratchet above.

Per-item ledger

id	落点 (file:span at base 96b627d)	before	after
QRY-A-01	SKILL.md:3-9 frontmatter description	"…picking pagination strategy, or designing a list view's filter spec" — but the real view filter spec is [{ field, operator, value }] over the 20-operator VIEW_FILTER_OPERATORS enum, which the package never names	the list-view clause moves to the "Do not use for" arm and names objectstack-ui; a 3-row "which filter dialect" table routes ObjectQL $ / view rule / dataset measure filter
QRY-B-01	SKILL.md:214-231, 274-295, 303-310, 316-338, 361-378, 380-399, 548-566	seven rules stated twice, with cosmetically different examples, forcing the reader to pick an authority	deleted from SKILL.md; rules/* are the anchor and keep one-line pointers. The rules files did NOT grow to absorb them — they already carried every one
QRY-B-02	SKILL.md:43-51	"When to Use This Skill", six bullets restating the description that already routed the load	deleted
QRY-B-03	SKILL.md:32-40	"Skill Boundaries", a 4-row table restating the description and skills/README.md's cross-skill routing	deleted
QRY-B-04	SKILL.md:20-52	308 tokens of preamble before the first rule; the single most decision-changing fact — object is the first argument — appeared nowhere	the file opens with the calling-convention construct
QRY-B-05	references/_index.md:14	generator-owned; the fix is upstream in packages/spec/src/data/query.zod.ts	filed as #14441, untouched here
QRY-C-01	SKILL.md:496-539	the search-mirror prescription in full, duplicating objectstack-data/SKILL.md — including the heading, the "one prescription" sentence and the formula-field paragraph — while :535-537 already named data as the anchor	the rule plus the pointer: "mirror the related record's title into a stored field … the field, the write hooks and the lint wording are objectstack-data → Search Fields (searchableFields)". Anchor re-verified after the data flight landed on main: objectstack-data/SKILL.md:100
QRY-C-02	rules/filters.md:221-258, :282-286	—	KEPT per the flight ruling: this package is the anchor for date macros / filter tokens. Confirmed at merge: the delivered ui flight (54a1125) now reads "The vocabulary, both accepted spellings, the two resolvers and the near-miss list are owned by objectstack-query → rules/filters.md"
QRY-C-03	SKILL.md:617-632	"CRM Analytics Query Blueprint", five rows of dataset/widget surface objectstack-ui owns, whose own closing line conceded the routing	one routing line: dashboards and reports are metadata, model them in objectstack-ui
QRY-C-04	rules/filters.md:260-280	21 lines restating a flow-node rule objectstack-automation already states	one pointer line; anchor re-verified on the merged tree. See "follow-up for objectstack-automation" below for the two things the anchor does not carry
QRY-C-05	SKILL.md:624, rules/aggregation.md:199-207	compareTo documented twice here and a third time in full in objectstack-ui	both copies deleted, pointer kept. Established at source first: compareTo is LIVE, not retired — packages/spec/src/ui/dashboard.zod.ts:438 declares compareTo: strictObject({…}); what was removed is the BARE-STRING form and compareTo.offset (dashboard.zod.ts:272, :290). Live ⇒ delete + pointer. Target re-verified on the merged tree: objectstack-ui/SKILL.md:1410
QRY-D-01	SKILL.md:274-279, 327-337, 440-445, 550-558; rules/aggregation.md:16-27, 138-144, 156-166; rules/pagination.md:12-18, 218-222, 160-177	roughly half of ~1,173 tokens of tombstone callout was "why it was removed" narration for keys retiredKey() types never	one 6-row "removed key → live replacement" table in SKILL.md; the per-key history deleted in all three files
QRY-D-02	rules/aggregation.md:88-96	the alias/push-down bug-history parenthesis, describing behaviour that no longer exists	deleted; the live rule ("Read the result under alias ?? field") stays
QRY-D-03	SKILL.md:20-28	opening blurb restating the description, plus a three-way callout legend whose warn arm is used exactly once	deleted; the legend's one live sentence folded into that single callout
QRY-D-04	SKILL.md:87-88	"inspired by Prisma, Strapi, and MongoDB" — positioning prose	deleted
QRY-D-05 / H-02 / H-03	evals/README.md	—	DEFERRED to #14296 item 2. evals/** untouched, 558 tok unchanged
QRY-E-01	whole package	58 ts fences, 0 check-skill-examples markers	NOT DONE — not in the card's IMPLEMENT list, and the corrected examples are engine.* call shapes, which are not resolvable from @objectstack/spec alone. See the gap ledger
QRY-E-02 (falsehood 1)	SKILL.md:407-425 + Rules list :427-434	the canonical expand example projected fields: ['title','status'] while expanding assignee and project — dropping both FK columns, so the query resolves NO relations and reports no error	the example keeps the FK columns, and the requirement is now the FIRST Rules bullet. Contract: packages/spec/src/data/query.zod.ts:509 ("The projection must RETAIN the foreign-key column"); pinned at packages/objectql/src/query-expression-conformance.test.ts:1575-1586; mechanism re-located at source — packages/objectql/src/engine.ts:8358-8359 (const val = record[fieldName]; if (val == null) continue;) then :8372 (if (uniqueIds.length === 0) continue;)
QRY-E-03	SKILL.md:60-69, 256-266, 280-294, 341-352, 407-425, 584-593 vs :371-377, 390-396, 604-612	two incompatible calling conventions, neither named; the most-copied shape ({ object: 'account', … } as an option bag) is refused by every engine method	a 5-row calling-convention table with the legal key set per method, quoted from packages/objectql/src/engine.ts:368-407 — find/findOne = the nine query keys PLUS the six driver passthrough keys (review round 1); ENGINE_AGGREGATE_OPTION_KEYS = context, where, groupBy, aggregations, having, timezone; ENGINE_COUNT_OPTION_KEYS = context, where, both without passthrough and deliberately so — plus where a bare { object, … } literal IS correct (findData's query, an expand value) and the top/filter alias fold. Rejection pinned at packages/objectql/src/engine-unknown-option.test.ts:124-135
QRY-E-04	SKILL.md:341-352, rules/aggregation.md:43-54	both group-by examples carried fields: (and orderBy:), neither of which is in ENGINE_AGGREGATE_OPTION_KEYS; rules/aggregation.md:57-59 called fields "a readability convention" without saying it is rejected	both shown as engine.aggregate(obj, {…}) with legal keys only, and the rejection stated
QRY-F-01	absent	context — 34 uses in packages/objectql/src/*.test.ts query literals, 40 in examples/** + packages/apps/** — had zero coverage in any package	a section: the RLS / system-read escape hatch, { isSystem: true } (the real spelling at examples/app-showcase/src/security/bind-position-sets.ts:50 and two siblings), and the query-bag vs trailing-argument rule with "the trailing one wins" (packages/objectql/src/engine.ts:1865-1877)
QRY-F-02 / G-03	SKILL.md:464-495, :479-504	only the object form search: { query, fields } was taught — 5 real usages against 74 for the bare string, which packages/spec/src/data/data-engine.zod.ts:120-127 calls "the canonical Tier-1 contract"; and three spellings of the narrowing knob appeared with none introduced	the bare string plus sibling searchFields is the canonical example (the shape packages/qa/dogfood/test/showcase-search.dogfood.test.ts:33,41 pins); one line names all three spellings and which layer takes which; the object form demoted to one line
QRY-F-03	SKILL.md:136-146, rules/filters.md:18-21	four string operators listed, three omitted ($icontains, $like, $ilike), and the case rule stated nowhere while the canonical examples were exactly the case-sensitive traps	three rows added to BOTH tables, with the rule quoted rather than paraphrased from packages/spec/src/data/filter.zod.ts:548-549, the ASCII-only folding domain, and — after review round 1 — the measured split of which faces answer $like/$ilike and which refuse them
QRY-F-04	SKILL.md:314-400, rules/aggregation.md	MED confidence, not in the card's IMPLEMENT list	NOT DONE. See the gap ledger
QRY-G-02	ten sites	four spellings of one version fact	one spelling survives the merges

Funded additions from the delivered objectstack-ui flight

All five rows PR #14338 listed as absent here were verified at source before porting, and all five are TRUE. Paid inside rules/filters.md, whose named payer is QRY-C-04's deletion.

ui row	verified at	ported
both {token} and ${token} spellings are accepted	packages/spec/src/data/date-macros.zod.ts:203 — DATE_MACRO_WRAPPED_RE is /^\$?\{([a-zA-Z0-9_]+)\}$/; context-tokens.zod.ts:109 — CONTEXT_TOKEN_WRAPPED_RE = DATE_MACRO_WRAPPED_RE	yes
two more near-misses: {user_id}, {organization_id}	packages/spec/src/data/context-tokens.zod.ts — CONTEXT_TOKEN_SUGGESTIONS maps user_id → current_user_id and organization_id → current_org_id	yes
a token embedded in a larger string is not substituted	packages/core/src/utils/filter-tokens.ts:50 — "Values that merely CONTAIN braces are left untouched"	yes, with the 'user-{current_user_id}' specimen
isDateMacroToken(tok) as the author-time check	packages/spec/src/data/date-macros.zod.ts:166, re-exported by packages/spec/src/data/index.ts:82	yes, with its sibling isContextToken and the "pass the token WITHOUT braces" caveat
the build rule's name and why it exists	packages/lint/src/validate-filter-tokens.ts:65 — export const FILTER_TOKEN_UNKNOWN = 'filter-token-unknown'	yes, with the failure mode: an unresolved token reaches SQL as a literal, matches nothing, and the widget renders 0

None withheld.

Per-file token delta

countTokens is ceil(utf8 bytes / 4) (scripts/check-skills-token-ratchet.mjs:456). Figures measured at head 49de20f; round 3 changed no skill file, so they are round 2's.

file	before	round 1	head	Δ vs before	ceiling
skills/objectstack-query/SKILL.md	5,443	3,784	3,915	−1,528	5,552
skills/objectstack-query/rules/aggregation.md	2,226	1,846	1,846	−380	2,357
skills/objectstack-query/rules/filters.md	2,144	2,100	2,136	−8	2,149
skills/objectstack-query/rules/pagination.md	1,381	1,158	1,158	−223	1,382
edited skill files, total	11,194	8,888	9,055	−2,139 (−19%)
skills/objectstack-query/evals/README.md	558	—	558	0 (deferred)	567
skills/objectstack-query/references/_index.md	518	—	518	0 (generated)	—
whole package	12,270		10,131	−2,139 (−17%)

Every edited skill file shrank against its pre-flight size. No ceiling raised, no file added, no file deleted, scripts/check-skills-token-ratchet.mjs untouched.

Gap against the ≈ −3,530 target, itemised rather than smoothed

cause	tokens
QRY-C-02 (−380 in the audit) was OVERRIDDEN to KEEP by the flight ruling — this package is the date-macro anchor. The −3,530 figure is −4,088 − (−558), which still carries that −380	+380
QRY-F-04 (≈ −400, MED confidence — reframing the aggregation cluster around a "which aggregation spelling" routing table) is not in the card's IMPLEMENT list, so it was not done	+400
the five funded ui rows above post-date the audit and are not in its −4,088	≈ +250
review round 1: the fifteen-key row and the measured $like face split, net of what was paid in-file	+167
residual: every MERGE-INTO leaves a pointer line the audit's raw-span estimates did not price, and the E-03 / F-01 / A-01 constructs landed slightly above their +120 / +90 / +70 budgets	≈ +194

premise_false

None of the audit's findings had a false premise; every span was re-verified at BASE (git diff --stat a59f78d 96b627d1 -- skills/objectstack-query is empty, so the audit's line numbers were exact). Corrections to stated premises, none of which changed a decision:

• premise_false: PM mechanism assumption 2 (half) — the QRY-A-01 description edit regenerates content/docs/ai/skills-reference.mdx ONLY. skills/README.md is byte-unchanged: it renders just the description's FIRST sentence, which A-01 did not touch. The generator writes both files and check:skill-docs verifies both; one of the two simply had no delta.
• premise_false: PM mechanism assumption 3 — no skills/objectstack-query heading is registered exhaustive. The Leg 2 BINDINGS table holds no row in this package. Run before the first deletion and after the last: green every time. No deletion here could red Leg 2.
• premise_false: the audit's line cite for falsehood 1 — the FK skip is at packages/objectql/src/engine.ts:8358-8359, not :8345-8347 (which is a comment block). The mechanism is exactly as described and is confirmed, with the second half at :8372.
• premise_false: review round 1's two readings of the $like faces — driver-memory answers $like/$ilike rather than refusing them, and filter-refusal.ts's hasDanglingLikeEscape gate is about a malformed PATTERN, not the operator; the three refusing faces do carry an INVALID_FILTER / 400 envelope. Evidence in the Review round 1 section above.

Also measured while executing the C-05 ruling: compareTo is a LIVE surface, not a retired one, so the live branch applied (delete + pointer, not delete with no pointer).

follow-up for objectstack-automation

QRY-C-04 deleted 21 lines here on the ruling that the flow-filter rule's anchor is objectstack-automation/SKILL.md. That anchor carries the core rule (a dropped filter condition WIDENS the query, so routing it would let a delete_record empty the object while the run reported success). Two things it does NOT carry, both deleted here:

1. Precedence on config.filter. The flow template engine runs first, but hands a recognised filter placeholder through untouched for the engine to expand; a flow variable named after a placeholder shadows it.
2. The static-checkability split. objectstack validate FAILS on a {record.…} filter token naming an unknown field, or hopping through a relation the start node does not expand; the same reference OUTSIDE a filter (a message body, an http url, a write payload) stays a warning, because there it renders a blank rather than widening a query. An unresolved flow variable is not statically checkable at all.

Neither is a query-DSL rule, so neither belongs here. Suggested for that package's own flight.

No follow-up for objectstack-data (the anchor is complete, re-verified after that flight landed) or objectstack-ui: the audit's incidental falsehood 2 (the ui claim that unknown placeholders "silently pass through as literal strings", contradicted by packages/core/src/utils/filter-tokens.ts:378) is already gone from main — the ui flight landed as 54a1125 and now points at rules/filters.md, which states it correctly.

Gates

Head sha 49de20f, a merge of current origin/main (a98b61b) into the branch, so the derivation is not from a stale tree. Exit codes captured before any pipe.

node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands re-derived over the now 6-path change set: 42 commands — six more than rounds 1-2, pulled in by scripts/role-word-baseline.json: check:bash32-floor, check:cli-command-ids, check:entry-guard, check:parse-guard, check:pnpm-filter-targets, check:watch-hint-literal. All 42 were run at 49de20f:

• 41 green.
• 1 NOT MEASURED, by that gate's own design: node scripts/check-test-completeness.mjs exits 3 with "PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named … record this gate as NOT MEASURED. ⛔ It is not a red." Recorded as NOT MEASURED, not as green.

The gate CI caught, now green, plus the six newly-pulled families:

✓ check-role-word — Ledger: 43 baselined file(s) still carrying it (126 occurrence(s)) in scripts/role-word-baseline.json.
✓ check-bash32-floor: 26 tracked shell file(s) … name no bash 4+ construct outside a comment, a guarded ${VAR:-} read, or a non-command position.
✓ check-cli-command-ids: 315 command-id literal(s) across 112 file(s) outside packages/cli all resolve to a real command path
✓ check:entry-guard: 198 scripts/ file(s) — every entry guard goes through invoked-as.mjs; 149 export bindings, 147 of them inert on import
✓ check:parse-guard (exit 0)
✓ check:pnpm-filter-targets: 142/181 `--filter` occurrence(s) across 32 file(s) resolve against 78 workspace package(s)
✓ check-watch-hint-literal: 43 declaration(s) across 4 rostered name(s) … no unrostered spelling of the idiom in the tree.

The five card-named gates, at the same head:

✓ check-skills-token-ratchet: 37 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted.
✅ 265 prose examples type-check across 3 surface(s) — every marked block parsed, so tsc ran the SEMANTIC pass on all of them
check-skill-identifier-liveness OK — Leg 1: 491 citation(s) over 47 published file(s) checked against 92934 implementation word tokens (3 ledgered exemption(s)); Leg 2: 8 registered exhaustive section(s), 0 ledgered gap(s).
✅ Skill docs in sync
✓ check-skill-compatibility-version: 11 SKILL.md file(s) reconciled against 78 workspace packages

(The ratchet's file count and Leg 2's binding count both dropped by one against round 2 because the data flight landed on main between the rounds and deleted skills/objectstack-data/rules/hooks.md; nothing in this PR moved either number.)

check:skill-examples ran against a freshly rebuilt @objectstack/spec + @objectstack/client-react closure (pnpm --workspace-concurrency=2 --filter '@objectstack/spec...' --filter '@objectstack/client-react...' build, exit 0), so it refused nothing rather than false-greening on a stale dist. Heavy steps ran under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-14302.

Why round 2 shipped a red: round 2 re-ran only the five card-named gates instead of re-deriving the union, and check:role-word is reached by the prose paths rather than by the card's list. Round 3 runs the whole re-derived list, which the new baseline file itself widens.

skip-changeset

Nothing released by any package changes. The change set is 4 files under skills/, one generated file under content/, and one gate baseline under scripts/; scripts/check-empty-changeset.mjs:357-360 enumerates exactly that population:

* It releases nothing (.github/, .claude/, skills/, docs/, content/, examples/, tests-only, and the like) — -> delete the changeset and apply the 'skip-changeset' label (route 2). The label is a gate-level exemption: it produces NO input for changesets/action.

Route 2 taken: no changeset file, label applied.

Out of scope, filed

• [finding] The published skill reference index labels data/query.zod.ts — the whole QueryAST — "Sort Node", because its first JSDoc block is SortNode's #14441 — QRY-B-05, the spec-side twin: packages/spec/src/data/query.zod.ts opens with SortNode's JSDoc, so the generated references/_index.md describes the whole QueryAST file as "Sort Node". Filed unassigned with the finding label after a dedupe pass; the generated artifact is untouched here.

No second out-of-scope card is owed from review round 1: the spec .describe() at packages/spec/src/data/filter.zod.ts:697 was checked against all three faces and is accurate.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

[github-actions]

⛔ merge queue 构建失败 — 先分诊，再决定要不要重排

队列构建 33625162151 红了。队列跑的是全量套件（PR 侧 CI 只跑 affected 子集），
所以失败的测试可能在本 PR 没碰过的包里 —— 那不是重排能修的。每次盲目重排都会让排在后面的所有 PR 重建一轮。

失败的 job（日志抽取，best effort）：

• Test Core (1/6) — 失败步骤: Run this shard's tests

  @objectstack/cli:test:  FAIL   unit  test/vitest-tiers-partition.test.ts > the two tiers of packages/cli (#13504) > INTEGRATION_FILES equals the behavioural predicate over every file on disk
    ↳ 失败原因: @objectstack/cli:test: AssertionError: files that spawn the CLI or boot a kernel/driver but are NOT in INTEGRATION_FILES (add them): expected [ Array(1) ] to deeply equal []
  

↳ 失败原因 是判读的关键：超时（Test timed out in … / Hook timed out in …）多半是负载/时序，不是本 PR 的回归；
断言（AssertionError: …）才指向真实的行为改变。两者的 FAIL 行长得一模一样，只有这一行能区分。

⚠️ 断言这一侧有一类例外，判据是断言在测什么，不是它是不是 AssertionError。 断言的对象是产品行为（一个值、一个形状、一次拒收）⇒ 照上面读：真实的行为改变，去查，⛔ 不要重排掉；
断言的对象是这次实验自身的有效性前提（跑完的耗时、负载下的先后、任何只在时间预算内才成立的条件）⇒ 它跟超时是同一类，同样对负载敏感，重排一次是合法的判别手段。
识别是机械的：断言的消息或它比较的值本身点名了一段时长、一个时间戳、一个耗时计数。实测过的一对 —— AssertionError: SecurityPlugin.init() ran: expected false to be true 测的是产品行为（真回归）；
AssertionError: this run took over a second, so second-precision stamps could have differed too: expected 1006 to be less than 1000 测的是实验前提：它守护的那条不变式当时是绿的，同一个 head 原样重排一次即成功。
穿着 AssertionError 外衣的时间测量，仍然是时间测量。（⛔ 这只改「怎么读一次红」，不改「哪些测试可以重排」——后者由别处管。）

跨 PR 相同签名（24h，按失败测试文件聚合）：

• test/vitest-tiers-partition.test.ts — 24h 内已弹出 5 个 PR（skills(query): optimization flight — seven twice-stated rules merged into rules/, removed-key history collapsed to one table, search/flow-filter/compareTo pointed at their anchors, context and the calling convention taught (net −2,139 tokens) #14443、test(cli): split the suite into named unit and integration vitest tiers #14536、docs(ci): stop hard-coding hook self-test deps in a comment #14538、docs(spec,approvals): ApprovalEscalation.timeoutHours names its clock — calendar (wall-clock) hours #14542、docs(spec): state minApprovals' real per-behavior default in the schema prose #14543），但它们同属 1 条投机栈 ⇒ 1 次独立命中。汇总 issue：Queue-flake anchor: test/vitest-tiers-partition.test.ts #14554（已刷新）
• ⚠️ 24h 评论账本没读完（超过 5 页仍未读到窗口尽头），所以上面的「不同 PR 数」是下界，不是全量。

历史信号：

• 本 PR 过去 24h 无队列失败记录（首次）。
• 过去 24h 队列共有 5 个失败构建（不含本次）。

分诊清单：

1. 失败测试在本 PR 改动的包里 → 真回归，修 PR。
2. 失败测试与本 PR 无关 → 看上面的「跨 PR 相同签名」；已有汇总 issue ⇒ flaky/环境问题实锤，去那张 issue 上谈，修好前重排只会再烧一轮全队列。
3. 两者都不是 → 可能与同组 PR 语义冲突；等前面的 PR 落地或失败出队后再重排一次即可，不要连续重排。

---

Generated by Claude Code · merge-queue-triage workflow (#4859)