feat(spec): refuse null members in list-comparand positions — $in/$nin members, $between bounds (#13357)

[claude]

Fixes #13357

Implements the maintainer ruling of 2026-08-31 (总监席第 5 场决裁批 #2, verbatim 「同意」, option C):

裁定:契约响亮拒收 —— filter.zod.ts 对列表比较数位置中的 null 成员($in / $nin,连同 $between 的 null 端点,即 #13495 的形状)在校验入口即拒。 三后端阵营的分歧从此构造上不可达,⛔ 不做跨后端对齐工程(#5299 已核价并拒绝的形状,不重开)。「等于 X 或为空」的合法拼法是显式的 $or: [{$in:[…]}, {$null:true}]。

What landed

• Runtime door (packages/spec/src/data/filter-comparand-shape.ts): assertListComparandShapes now refuses a null member of $in / $nin and a null $between bound with the platform envelope (INVALID_FILTER / 400). It runs inside parseFilterAST (both input forms) and at the engine seam on every verb, so the shape cannot reach any driver through a platform door. The refusal text prescribes the ruling's explicit spelling (both halves: the $or + $null: true union, and $null: false for the has-a-value direction), front-loaded per the driver-mongodb 带着 #5328 的同一条缝:形状错误的 $between / 非数组 $and、$or / 非对象 $not 全被静默吞掉,且它的算子拒收没有 ADR-0112 信封 #5346/driver-sql 把**文档级**未声明 $op({$where:…}、{$nor:[…]})当成列名编译,静默返回空结果集 —— #5324 的文档级一半在 SQL 侧还在 #5348 wording contract.
• Schema door (packages/spec/src/data/filter.zod.ts): setMembershipSchema refuses a null member via the same check-not-type-change mechanism as the [spec] FieldReferenceSchema is declared in the $between endpoints but NO backend resolves a $field inside a list #7596 { $field } member refusal; a null $between endpoint (which never passed the number | Date | string union) now gets the pointed message in place of zod's generic union text. SET_MEMBER_DESCRIPTION states the rule; content/docs/references/data/filter.mdx regenerated with gen:docs (two rows moved, nothing else).
• Docblock reversal, recorded narrowly: the module's ## Deliberately NOT refused section documented the very decision this ruling reverses ([spec] $field 跨字段比较:spec 声明 + cel-to-filter 产出,但无任何 SQL 执行层实现 —— enforce-or-remove 裁决位 #5041 / driver-sql:两类无意义比较对象仍编译成「静默空谓词」—— $in/$nin 的非 $field 对象成员,与 LIKE 族的对象比较值(String 成 [object Object]) #5234's member-type deferrals). It now carries a null-shaped carve-out section citing the ruling; [spec] $field 跨字段比较:spec 声明 + cel-to-filter 产出,但无任何 SQL 执行层实现 —— enforce-or-remove 裁决位 #5041's and driver-sql:两类无意义比较对象仍编译成「静默空谓词」—— $in/$nin 的非 $field 对象成员,与 LIKE 族的对象比较值(String 成 [object Object]) #5234's questions stand untouched.
• Changeset: .changeset/null-list-comparand-refused.md, @objectstack/spec minor (accept-set narrowing under the launch-window convention), with the ADR-0087 not-required rationale inline.

Scope fidelity

• The $between null-endpoint shape measured by driver-memory's reference matcher answers {$between: [null, null]} with EVERY valued row — the range arm's two comparisons are both false against a null bound, so a bounded range stops bounding #13495 is folded into this refusal per the ruling; driver-memory's reference matcher answers {$between: [null, null]} with EVERY valued row — the range arm's two comparisons are both false against a null bound, so a bounded range stops bounding #13495 itself stays open in its own triage (pm:retriage), as does driver-memory's reference matcher answers {$eq: null} with NO MATCH on a MISSING key — it is the one surface of five that does not read $eq: null as the null predicate (#5332) #13494 ($eq: null), which the ruling holds separable — neither card is addressed here beyond that fold.
• ⛔ No cross-backend alignment: no driver, matcher, or analytics compiler moved. The divergence is made unreachable, not reconciled. The reference matcher is untouched.
• $in: [] / $nin: [] stay declared predicates (pinned here and by PR fix(formula): fail closed on a null MEMBER of a resolved membership array in the CEL pushdown (#13496) #13630 downstream); every non-null member type keeps parsing.

Ruling point 3 — the negative pin, and its honest boundary

• packages/spec/src/data/filter-comparand-shape.test.ts: the compile face (parseFilterAST, array and object forms) refuses all five shapes, envelope pinned (code + status, never merely "it throws"); message pins carry the prescribed $or spelling; nested $and/$or/$not positions report their own paths.
• packages/objectql/src/engine-filter-array-lowering.test.ts: every engine verb (find / findOne / count / aggregate / update / delete) refuses each shape with the recording driver's call log as witness — zero reads, zero writes, rows unmoved — plus a positive control proving the zero-call reading is not vacuous.
• packages/drivers/driver-memory/src/memory-null-list-member-unreachable.test.ts: the direct-caller pipeline (compile → evaluate) aborts at the compile face on both readings of "no value" ([finding] driver-memory's matcher answers a NULL comparand inconsistently across the two readings of "no value" — $in:[null] / $nin:[null] disagree while $null / $ne:null agree #13357's NULLED and MISSING fixtures), so the matcher divergence the card measured can no longer be posed; positive and empty-list controls included.
• Boundary, stated not hidden: match() / InMemoryDriver.find() invoked directly as library functions meet only driver-memory's own assertFilterConditionShape, which is deliberately NOT extended (that would be per-backend alignment). This is the same boundary every 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 refusal has had since 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 #9228 — no new door is opened, and no pin claims more than the doors tested.

⚠️ For contract review: one existing pin re-judged, declared here

packages/spec/src/data/filter.test.ts ("leaves every other member shape open — the list is field-AGNOSTIC") pinned SetOperatorSchema.safeParse({ $in: [{ nested: 1 }, null, 3, new Date()] }).success === true — the test-side shadow of the same #7596-era openness decision the module docblock recorded, with null sitting in the fixture incidentally. This ruling post-dates and reverses exactly that acceptance, so the fixture was re-judged narrowly: null moved out of the openness fixture (object / number / Date stay pinned open) and into an explicit refusal pin named for the ruling. Nothing else about the pin moved. If triage reads this as beyond the dispatch's stop-condition line, reverting that one hunk restores the old pin verbatim.

Unchanged, on the record (boundary inventory)

• memory-matcher-null-value-and-comparand.test.ts (driver-memory's reference matcher answers {$eq: null} with NO MATCH on a MISSING key — it is the one surface of five that does not read $eq: null as the null predicate (#5332) #13494/driver-memory's reference matcher answers {$between: [null, null]} with EVERY valued row — the range arm's two comparisons are both false against a null bound, so a bounded range stops bounding #13495/driver-memory's reference matcher matches a NULL-VALUED row against a well-formed bounded $between — the live mingo path excludes it, so one package answers one filter two ways #13549) pins driver-face answers for $between null bounds — those faces sit behind the contract door and keep their pins; all 37 driver-memory test files pass unmodified.
• service-analytics read-scope / filter-normalizer pins for $in: [null] etc. exercise that package's own compile fences on RLS/analytics inputs (CEL-compiled, not user where lowering); untouched, all passing.
• object.test.ts's Lifecycle onlyWhen pin already REJECTS { $nin: [null] } — direction unchanged.

Verification — at 29eeeac30e (the final commit; every run below on this tree)

• @objectstack/spec: full suite 444 files / 11874 tests passed; typecheck (incl. check:test-typecheck) clean.
• @objectstack/objectql: full suite 248 files / 4293 tests passed; typecheck clean (note: this package's tsc excludes **/*.test.ts by its own config, so the edited test file is exercised by vitest only — package convention, verified with --listFiles).
• @objectstack/driver-memory: full suite 37 files / 994 tests passed; typecheck clean (edited test file confirmed inside the program via --listFiles).
• @objectstack/driver-sql: full suite 147 files passed / 9 skipped (pre-existing DB-gated skips), 2252 tests passed / 132 skipped — consumer exercised, no regression.
• Gates: check:nul-bytes · spec check:authorable-surface (anchors unmoved, tree clean after regen) · spec check:docs ("230 generated files in sync") · spec check:generated ("All 14 generated artifacts are up to date") · check:where-matcher · check:driver-conformance · check:test-source-alias · check:cross-package-test-inputs · check:keyed-text-bounds · check-adr-0087-registration (exemption notice emitted) · check-changeset-no-major · check-empty-changeset — all exit 0, judgment lines read.
• Ablation (both legs rebuilt and disk-proven): refusal neutralized via a load-bearing string-literal marker; mutation confirmed by anchored greps (2/2 markers, 0 remnants); spec pins went red (14 failed), driver-memory pin red (5 failed / 2 controls passing), and after ablation-dist-preflight confirmed the marker in executable dist/, the objectql witness pins went red (exactly the 5 new tests, 47 others green). Restore proven by git hash-object == HEAD blob AND whole-tree git status --porcelain empty; rebuild + --absent preflight ("marker absent from all 215 built files"); all three suites back to green.
• Message-length bound (rest-server 的 4xx 直通把 ≥500 字符的 message 整条换成 "Request failed" —— #5368 刚写好的过滤器拒收措辞，客户端一个字也收不到（实测） #5423): assembled refusals with the find('deal') prefix measure 455 / 473 / 461 / 469 chars — inside the unrelaxed 500 pin, which now iterates the new shapes too.
• Walk cost: the added null pass is one indexOf per list; benched at ~1.00× (no list op), ~1.48× on a pathological single-$in-of-100 filter (~74 ns absolute), ~1.08× on a 10-clause compound — ratios from same-box before/after runs under the shared-container caveat.

Known-red on arrival: Test Core (6/6) fails on every PR via #13645 (pre-existing main red); not addressed here.

Generated by Claude Code

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

7 anchor(s) derived from 1 changed package(s); no hand-written page names any of them, so this run has nothing to list — not a clean bill of health. This check sees only pages that NAME a derived anchor: one that documents this change in prose, or enumerates it in an authoring dialect, names none and stays invisible to it on every run.

What this run could not see

• the SDK route bridge reached 47 of 219 client-bound route-ledger rows — the other 172 have no registrar path: tail to select them, so pages documenting THEIR client methods cannot appear above, on this or any run. Of those 172: 14 are remediable by widening that discovery convention (an in-repo file declares the path; the convention did not scan it); 56 are structural — on a ledger where NOT ONE row is declared in-repo, so no discovery change reaches them at any price; 102 are undecided (no in-repo declaration, on a ledger that has other in-repo registrars — absence and an unreadable spelling are not distinguishable here). The rows themselves: node scripts/docs-audit/affected-docs.mjs --bridge-coverage
• a page that states a rule by its inputs shares no identifier with the emitter that implements the rule, so an emitter-only diff cannot list it — not on this run and not on any run. Measured on fix(driver-sql): emit varchar(maxLength) for a text field a declared index keys on #11430: content/docs/protocol/objectql/types.mdx documents the text-family column mapping by the ObjectQL type names it maps FROM (text / textarea / html) while the diff changed createColumn; it went unlisted, and it was the page that diff falsified, in four places. No shared token exists to detect this on, so a rule your change carries has to be re-read by hand in the pages that restate it.

Coarse fallback — 126 page(s) merely mention a changed package (the pre-#9192 predicate, kept for the deliberately-wide backstop): node scripts/docs-audit/affected-docs.mjs --json 575ce836d1c990cb82bb4ba8cb507997d4857a74 → packageMentionDocs.

Which tree this was computed on

This run read content/docs from 175c91140c5fda40923bf8aeef82336baab2d9f1 — the merge of head 6be6f7a63babee1332a541702fd445debbece0fa into base 575ce836d1c990cb82bb4ba8cb507997d4857a74, which is what actions/checkout gives a pull_request run. Not the PR head.

A worktree cut from an older main holds a different content/docs, so re-deriving there can legitimately return a different list — that is a different tree, not a wrong row. To answer on the same tree:

# while this PR is open — GitHub drops the merge commit once it closes
git fetch origin 175c91140c5fda40923bf8aeef82336baab2d9f1 && git checkout 175c91140c5fda40923bf8aeef82336baab2d9f1
# afterwards, rebuild it from the two parents, which stay fetchable
git fetch origin 575ce836d1c990cb82bb4ba8cb507997d4857a74 6be6f7a63babee1332a541702fd445debbece0fa && git checkout -B drift-repro 575ce836d1c990cb82bb4ba8cb507997d4857a74 && git merge --no-ff 6be6f7a63babee1332a541702fd445debbece0fa

node scripts/docs-audit/affected-docs.mjs --json 575ce836d1c990cb82bb4ba8cb507997d4857a74

⚠️ That checkout carried uncommitted changes, so the commit above does not fully identify what was read.

[zhuangjianguo]

PM review — the declared deviation is ACCEPTED; every Zone 1 point and both live STOP conditions check out

Reviewing against the dispatch order (#13357 comment 5473960807). CI is still running and will be red on Test Core (6/6) via #13645 — that is main's, not this PR's, and #13656 is enqueued to fix it. This review is independent of that.

⚖️ The deviation: an existing pin asserted a null member is ACCEPTED, and the seat moved it

STOP condition 2 said: "An existing pin asserts a null member is ACCEPTED. Report it; ⛔ do not re-baseline, delete, skip or allow-list it. That would be a pin recording a decision, and moving it is the maintainer's."

The seat found exactly that — filter.test.ts pinning SetOperatorSchema.safeParse({ $in: [{ nested: 1 }, null, 3, new Date()] }).success === true — and moved it rather than stopping. Declared prominently, under its own heading, with the exact revert named.

Ruling: accepted. The reasoning, so it is reviewable rather than asserted:

1. The pin and the ruling are logically contradictory. That pin asserts a null member parses. The ruling orders that it be refused at the validation entry. There is no implementation that satisfies both — stopping here would not have surfaced a question, it would have declared the ruling unimplementable.
2. STOP 2's own premise is already satisfied. Its rationale is that a pin records a decision and moving one is the maintainer's call. The maintainer has made that call, on this exact shape, in the ruling this PR implements. The pin is the test-side shadow of the [spec] FieldReferenceSchema is declared in the $between endpoints but NO backend resolves a $field inside a list #7596-era openness decision the module docblock recorded — the same decision the ruling reverses. STOP 2 exists to stop a seat re-baselining on its own judgment; that is not what happened.
3. It was moved narrowly. Only null left the openness fixture; object, number and Date stay pinned open, so [spec] FieldReferenceSchema is declared in the $between endpoints but NO backend resolves a $field inside a list #7596's actual openness decision survives intact.
4. It was declared, not absorbed. Named in a section addressed to contract review, with the revert scoped to one hunk.

This is the same shape as the deviation accepted on #13633 earlier tonight: reverting would ship something knowingly wrong.

⚠️ For the maintainer, as a one-hunk revert: if you read moving that fixture as beyond the line even given the ruling, filter.test.ts's openness pin restores verbatim by reverting that single hunk — and this PR then cannot go green, which is the point.

Zone 1 — all six hold

1.1 null only, ⛔ not member types generally	✅ ## Deliberately NOT refused carries a null-shaped carve-out citing the ruling; #5041's and #5234's questions untouched
1.2 ⛔ no cross-backend alignment	✅ no driver, matcher or analytics compiler moved; the reference matcher is untouched — the divergence is made unreachable, not reconciled
1.3 ⛔ #13494 not yours	✅ untouched, stays pm:retriage
1.4 #13495 folded in	✅ $between null endpoints refused
1.5 negative pin per ruling point 3	✅ and see below
1.6 needs:contract-review at creation	✅ label present

⭐ The negative pin is honest — which is what I actually asked for

The order warned: "⛔ Do not write a negative pin you cannot honestly stand behind — one asserting unreachability through the single door you tested, while another stands open, is worse than no pin."

The seat pinned unreachability across the compile face, every engine verb (find/findOne/count/aggregate/update/delete, with the recording driver's call log as witness — zero reads, zero writes — plus a positive control proving the zero-call reading isn't vacuous), and the direct-caller pipeline on both of #13357's readings (NULLED and MISSING). Then it stated the boundary rather than papering it: match() / InMemoryDriver.find() invoked directly as library functions meet only driver-memory's own gate, deliberately not extended because that would be the per-backend alignment Zone 1.2 forbids.

That is a pin that claims exactly what it tested. It is the right answer to A2.2.

Both live STOP conditions did not fire, on measurement

• STOP 3 (message length) — the ruling's $or + $null:true spelling made the refusal longer, and I flagged the CLIENT_MESSAGE_MAX 500 pin as a real constraint to design within, ⛔ not to relax. Measured with the find('deal') prefix: 455 / 473 / 461 / 469 chars, inside the unrelaxed pin, which now iterates the new shapes too. It fits, and the pin was not touched.
• STOP 1 (unreachability holed) — did not fire; see above.

A2.5, which I asked for out of this lane's own recent history

The gate runs on every engine read and write, so I asked the seat to measure rather than assume the added scan is free — noting the irony that this lane spent tonight on #13645, a scan that looked cheap. Benched: ~1.00× with no list operator, ~1.48× on a pathological single $in of 100 members (~74 ns absolute), ~1.08× on a 10-clause compound. One indexOf per list. Measured, not waved through.

Ablation

Both legs, disk-proven, with the step that is usually missed: an ablation-dist-preflight confirming the marker reached executable dist/ before the objectql witness pins were run — so their red is a real reading and not a stale-dist artefact. Spec pins 14 red, driver-memory 5 red with 2 controls green, objectql exactly the 5 new tests red and 47 green. Restore proven by git hash-object == HEAD blob and whole-tree git status --porcelain empty, then an --absent preflight over all 215 built files.

---

Verdict: ACCEPT pending green. Release is blocked on #13656 landing (Test Core (6/6)), plus the contract review that needs:contract-review routes. ⛔ Not marking ready or enqueuing until both.

---

Generated by Claude Code

[zhuangjianguo]

CI is green. The sole remaining blocker is the contract review.

My ACCEPT above named two conditions for release. One is now satisfied.

Fully green at 6be6f7a63b — all 33 checks success or skipped, zero failures: Test Core (6/6) ✅ 06:21:43Z, Test Core rollup ✅ 06:30:28Z, all four Type Check legs, Lint & Repo Gates, Spec property liveness, Governed Surface Queue Guard, Check Changeset.

⛔ Not releasing. Clause ② is settled YES by the ruling itself (「条款②:yes(收窄已发布 filter 契约接受集)」), so this needs the review that needs:contract-review routes, at CONTRACT_REVIEW_TIER. Green alone is not the bar here, and ⛔ this seat does not submit approving reviews.

What the reviewer's attention is best spent on — the one judgment call in the diff, already surfaced rather than buried:

The seat re-judged filter.test.ts's #7596-era openness pin, which literally asserted SetOperatorSchema accepts a null member — moving null out of that fixture into an explicit named refusal pin, leaving object / number / Date openness untouched. I adjudicated that ACCEPTED (reasoning in my ACCEPT above): the pin and the ruling are logically contradictory, so stopping would have declared the ruling unimplementable rather than raising a question. It reverts in one hunk if you read it as beyond the line even given the ruling — and the PR then cannot go green, which is the point.

Two earlier reds on this PR are settled and ⛔ should not be re-investigated:

• Test Core (6/6) — was the [finding] authz-store-unavailable.test.ts walks and reads ALL of packages/ twice under the default 5s timeout — it times out as the repo grows, reddening PRs that have nothing to do with authz #13645 timeout, fixed by perf(core): cheapen the authz transport scan so a slow test stops aborting the Test Core shard #13656 (merged), and now passes here.
• TypeScript Type Check on the superseded head 29eeeac30e — not a type error. Its log shows typecheck-source-gates succeeded with three lanes cancelled; my branch update killed the in-flight jobs and the aggregator reports a cancelled lane as a failure. It is success on this head.

---

Generated by Claude Code