The #5869 list-comparand rule is enforced only at the engine seam: a scalar `in`/`nin` comparand still reaches a driver, and on mingo >= 7.2.3 it escapes as a raw `TypeError`

[os-project-manager]

Surfaced by the weekly Dependabot production group #9212 (mingo 7.2.2 -> 7.2.4), which is red on this and nothing else. The bump did not break us — it removed the third-party coercion that was hiding a coverage hole in an already declared rule.

Measurement (verified on GitHub, not recalled)

PR #9212 @ 45eaf8db3a03fe30d8798069439aad74e4c219c5, two independent jobs, same single failure:

FAIL packages/drivers/driver-memory/src/memory-filter-ast-vocabulary.test.ts
 > InMemoryDriver filter vocabulary <-> VALID_AST_OPERATORS > expresses notin without dropping it
AssertionError: promise rejected "TypeError: b.filter is not a function" instead of resolving
 > $in node_modules/.pnpm/mingo@7.2.4/.../operators/_predicates.js:95
 > $nin node_modules/.pnpm/mingo@7.2.4/.../operators/_predicates.js:99

• Test Core (2/3): https://github.com/objectstack-ai/objectstack/actions/runs/31987846882/job/95265845071
• Temporal Conformance (live PG + MySQL): https://github.com/objectstack-ai/objectstack/actions/runs/31987846882/job/95265844998
• main CI was green in the same window (runs 31987144246 / 31989563215 / 31991738619, all success), so the red is introduced by the PR, not inherited from the base.

mingo's own changelog for 7.2.4 names the change: "make $elemMatch with empty field and $in/$nin with regex behave correctly (#613)". 7.2.2 tolerated a non-array $in/$nin operand; 7.2.4 does not.

Why this is our defect, not mingo's

The contract is already declared, and its gate already exists — it just is not on every path in.

1. FieldOperatorsSchema (packages/spec/src/data/filter.zod.ts) declares $in: z.array(z.any()), $nin: z.array(z.any()), $between: z.tuple([min, max]).
2. packages/objectql/src/filter-comparand-shape.ts (数据 API:集合算子 not_in/in 的比较值是标量(非数组)时答 500 DATABASE_ERROR,而不是带信封的 400 —— 而这个形状是 spec 合法的 ViewFilterRule #5869 / PR fix(objectql): 集合算子的标量比较值答 400 INVALID_FILTER 并点名期望形状,不再 500 (#5869) #6209) enforces it — assertListComparandShapes, called from engine.ts, answers 400 INVALID_FILTER. Its header states the rule verbatim: "parseFilterAST lowers ['status', 'not_in', 'done'] to { status: { $nin: 'done' } } without complaint — so a scalar reached the driver", and it explains why the gate went to the engine seam rather than into each driver (both driver families are under the [裁决] driver-memory / driver-mongodb 投入冻结 —— 维护者 2026-08-05 口径（跨单锚点） #5499 investment freeze).
3. The spec-side comparand-type door added by PR feat(spec): comparand-type door — the accepted literal comparand set, enforced once at the shared compile face for all five drivers (#7872) #8234 (driver-memory has no policy for an unsupported comparand TYPE: a BigInt crashes with a raw mingo TypeError, and four other types silently answer zero rows #7872) sits at exactly the seam where a shape check would belong, and deliberately routes past it — packages/spec/src/data/filter-comparand-type.ts:

} else if (LIST_COMPARAND_OPERATORS.has(op) && Array.isArray(comparand)) {
 // Each member is a comparand in its own right (#5234). A non-array
 // comparand here is a SHAPE defect owned by the engine's #5869 gate.

So: parseFilterAST() lowers a scalar membership comparand without complaint, and only a query that goes through the objectql engine meets the gate. A caller that reaches a driver directly — InMemoryDriver.find(), which is what this driver's own conformance suite does and what an embedder does — meets nothing. Until now that path was carried by mingo's coercion; from mingo 7.2.3 on it is an unhandled third-party TypeError with no code and no status, reaching the caller.

That is the same class of escape #7872 closed for comparand types, one branch over: types are judged at the shared door, shape is not.

packages/drivers/driver-memory/src/memory-driver.ts:820-821 is where it lands ({ [field]: { $nin: store(value) } }, no normalization, no refusal). The test only exercises it because valueFor (memory-filter-ast-vocabulary.test.ts:80-87) hands an array to in / nin / not_in but a scalar to the equally-accepted notin and not in spellings — an accident that turned out to be the one probe covering this hole.

Ruling (not open for re-adjudication by the dev)

• Fix it at the shared face, one implementation. Do not add a per-driver patch: driver-memory and driver-mongodb are under the [裁决] driver-memory / driver-mongodb 投入冻结 —— 维护者 2026-08-05 口径（跨单锚点） #5499 investment freeze, and PR feat(spec): comparand-type door — the accepted literal comparand set, enforced once at the shared compile face for all five drivers (#7872) #8234 already set the precedent for this exact question ("enforced once at the shared compile face for all five drivers").
• Do not create a second copy of the rule. Exactly one implementation of "a list operator takes a list" may exist when you are done; assertListComparandShapes either becomes the single implementation the door calls, or delegates to the one that moved.
• driver-memory changes are test-only here (construct valid input for all membership spellings). No driver behaviour patch.
• Do not touch mingo versions, package.json or pnpm-lock.yaml in this PR. The bump belongs to Dependabot chore(deps)(deps): Bump the production-dependencies group with 19 updates #9212, which is expected to go green because of this card.

Premises — verify before you build, and report a fork instead of working around them

1. Dependency direction allows the sink. The rule has to end up somewhere parseFilterAST can reach, i.e. in packages/spec, with @objectstack/objectql delegating to it. objectql depends on spec, so the reverse import is not available. If the assertion cannot move without dragging engine-only concepts into spec, that is a fork — report it, do not invent a third home.
2. Nothing in-tree authors a scalar membership comparand. Tightening parseFilterAST refuses input that today survives to the engine gate anyway (which already answers 400), but stored view metadata, examples, docs fixtures and objectui view definitions must be checked, not assumed. Grep for authored in / nin / not_in / notin / not in rules with non-array values across examples/**, content/docs/**, packages/**/fixtures/** and the objectui view metadata. If real authored usage exists, stop and report the fork with the sites: normalizing a scalar to a one-element list is then a live option and that is the maintainer's call, not yours.

If either premise fails: report premise_still_valid: false with the measurement. Do not hard-do it, and do not quietly switch to the other option.

Suggested route (advisory — measurement beats it)

In filter-comparand-type.ts the LIST_COMPARAND_OPERATORS branch already knows the operator is list-shaped; the Array.isArray(comparand) guard silently falls through today. Make the else arm a loud refusal reusing the existing envelope (INVALID_FILTER, 400) and the #5869 wording, then have assertListComparandShapes delegate. Keep the message's front-loaded shape — the 500-char client bound is real (#5423).

Acceptance

• parseFilterAST([['name', 'notin', 'alpha']]) (and the nin / not_in / not in spellings) answers a coded INVALID_FILTER / 400 refusal, with a test that asserts both code and status, not just that it throws.
• pnpm --filter @objectstack/driver-memory test green, with memory-filter-ast-vocabulary.test.ts exercising every membership spelling on valid input.
• The engine path keeps its existing behaviour and its existing tests (engine-comparand-type-door.test.ts, the 数据 API:集合算子 not_in/in 的比较值是标量(非数组)时答 500 DATABASE_ERROR,而不是带信封的 400 —— 而这个形状是 spec 合法的 ViewFilterRule #5869 suites) — one implementation, unchanged verdicts.
• Measured, not assumed: run the driver-memory suite once against mingo@7.2.4 via a temporary local override to confirm the escape is gone, then revert the override — report the reading in the PR body, commit no manifest or lockfile change.
• Local gates for this surface (taken from node scripts/pm/dispatch-gates.mjs at dispatch time, re-take for your actual diff): pnpm check:test-source-alias, pnpm check:type-source-resolution, node scripts/docs-audit/check-affected-docs.mjs; and because a test file moves: pnpm check:query-options-erasure, pnpm check:type-check-coverage, pnpm check:type-check-debt (needs the workspace built first), pnpm check:engine-double-contract, pnpm check:where-matcher.
• A patch changeset is required (published packages change).

Re-check commands for the premises above

git fetch origin main
git show origin/main:packages/spec/src/data/filter-comparand-type.ts | grep -n "SHAPE defect owned by"
git show origin/main:packages/objectql/src/filter-comparand-shape.ts | grep -n "export function assertListComparandShapes"
git grep -n "notin\|not_in" origin/main -- examples content packages/*/fixtures

Links

• Blocks: chore(deps)(deps): Bump the production-dependencies group with 19 updates #9212 (Dependabot production group; red on this failure alone)
• Sibling branch, same door: driver-memory has no policy for an unsupported comparand TYPE: a BigInt crashes with a raw mingo TypeError, and four other types silently answer zero rows #7872, closed by PR feat(spec): comparand-type door — the accepted literal comparand set, enforced once at the shared compile face for all five drivers (#7872) #8234
• The declared rule and its engine gate: 数据 API:集合算子 not_in/in 的比较值是标量(非数组)时答 500 DATABASE_ERROR,而不是带信封的 400 —— 而这个形状是 spec 合法的 ViewFilterRule #5869 / PR fix(objectql): 集合算子的标量比较值答 400 INVALID_FILTER 并点名期望形状,不再 500 (#5869) #6209
• Driver investment freeze: [裁决] driver-memory / driver-mongodb 投入冻结 —— 维护者 2026-08-05 口径（跨单锚点） #5499

[os-project-manager]

Claim: PM loop round 1
Session: session_01HDA9nN6nXQngoQUAAzRdMb
Branch: claude/issue-9228-list-comparand-door
Worktree: objectstack-issue-9228
Domain: domain:spec
File surface: packages/spec/src/data/** (the door), packages/objectql/src/filter-comparand-shape.ts (delegation only — no second copy of the rule), packages/drivers/driver-memory/src/memory-filter-ast-vocabulary.test.ts (test-only), .changeset/ (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: fable — mandatory, not a judgment call: this card changes contract accept/reject behaviour at the packages/spec filter door.
Serial constraints cleared: #9205 / PR #9224 (spec lane — automation notify + email templates) and #9198 / PR #9221 (spec lane — packages/spec/src/ui element schemas), both in flight under a different seat (os-zhuang) and both disjoint from packages/spec/src/data/**; no open PR touches the filter door or filter-comparand-*. #9212 is the consumer waiting on this card, not a predecessor.

Routing authority: routed and dispatched from a PM session under the maintainer's direct instruction (2026-08-10 standing direct-dispatch channel). Quoted verbatim rather than paraphrased, because the domain:spec seat is this lane's normal owner and the bypass is valid only for the cards the maintainer named:

先立卡修我们这侧

立一个专项卡，你负责派发

---

Generated by Claude Code

[os-project-manager]

Dispatch 1 died on a model quota wall before it produced anything: You've reached your Fable 5 limit. Verified leftovers, so this is a clean re-dispatch and not a hand-over — git ls-remote origin 'refs/heads/claude/issue-9228*' returns nothing and no worktree exists.

Re-dispatched at model: opus, under the maintainer's standing quota exemption (2026-08-13), quoted verbatim:

fable 如果用完了，可以用 opus

Amending the claim's Container & model line accordingly: M, mode:subagent, model: opus — the exemption's floor, nothing below it. The clause ② mandate itself is unchanged; only its dispatch tier is exempted, because fable was measured unavailable rather than judged unnecessary.

Consequence, recorded now rather than discovered at queue time: this dispatch tier is below CONTRACT_REVIEW_TIER (claude-fable-5, single-sourced from scripts/pm/dispatch-gates.mjs). If the delivered diff touches packages/spec/src/**, this seat will not queue the PR — it will record the finding, apply needs:contract-review, and stop. The contract-review lane clears the label at the required tier; the quota exemption covers dispatch only and never the review, which exists precisely to compensate for a dispatch below the floor.

---

Generated by Claude Code