fix(objectql): judge each readonlyWhen lock against the row the update stores, not a value another lock drops (#19911)

[objectstack-fleet]

Fixes #19911

Clause-②: no

What a caller could do before, and what happens now

Before. When one field's readonlyWhen read a field that carries its own readonlyWhen, a caller with edit rights could change a locked field by sending a new value for the other field in the same update. The conditional strip judged every lock against one record view built before any lock was judged. That view still held every value the strip then dropped. With status locked by previous.status == 'closed' and amount by record.status == 'closed', update(c1, { status: 'open', amount: 999 }) on a closed row stored { status: 'closed', amount: 999 }. The same happened on bulk (multi: true) updates, for isSystem callers (a readonlyWhen lock binds them), and for a master-detail FK's own lock judged inside the PR #19877 settlement: the line moved to another invoice although its own lock held on the row it kept. The mechanism also over-locked in reverse: a dropped value that WOULD lock another field locked it, though the row never took that value.

Now. Each lock is judged with its own incoming value and every OTHER dropped key reverted to the stored value. No field is written while its lock is TRUE on the row the update stores. The strip then releases, in one step, every dropped key that is unlocked on that row, and keeps the result only when every dropped key is locked and every kept key is unlocked on the row it stores. When that step does not settle the set (locks that read each other in a cycle, or a cascade where releasing one key moves another's verdict, #19927), the larger fail-safe drop set stands. This applies on all four conditional call sites (by-id and bulk strips, and the by-id and bulk FK judgements of the settlement).

A1: reproduction on origin/main a34c27cbe5, before any edit

Real ObjectQL engine, the PR #19905 suite's in-memory driver shape. Object case_x: status readonlyWhen: "previous.status == 'closed'", amount readonlyWhen: "record.status == 'closed'", rows c1, c2 closed with amount: 100.

• By id: update(c1, { status: 'open', amount: 999 }) stored { status: 'closed', amount: 999 }, event [{ status, readonly_when }]. Control update(c1, { amount: 555 }) stayed 100. Reproduced.
• Bulk (where: { note: 'k' }, multi: true): both rows stored amount: 999. Control stayed 100. Reproduced.

A2: the candidates, measured

All five variants (the base and four candidates) were run on the same 20 files: the nine suites PR #19905's body names (its own engine-readonly-when-stored-view included) and every other objectql test file that mentions readonlyWhen (674 tests). Every variant passes 674/674, so the existing pins do not tell the candidates apart. They were compared on the card's shape and on a 27-row probe of every movement class instead. The expected rows were checked against a brute-force enumeration of every drop set that agrees with the stored row.

shape (exact drop sets by enumeration)	base	(i) monotone fixpoint	(iv) fixpoint + exact release (chosen)	(ii) judge against the stored row	(iii) refuse when locks interact
the card, by id and bulk ({status, amount})	amount 999 under-lock	100	100	100	refused
legitimate reopen: status unlocked, amount edited ({})	both land	both land	both land	amount dropped over-lock	both land
close + edit on an open row ({amount})	amount dropped	dropped	dropped	amount 999 lands on a closed row under-lock	dropped
REVERSE: frozen status, closed + amount ({status})	amount dropped over-lock	dropped over-lock	amount lands	lands	refused
lock reading its own field, open row over the cap	dropped	dropped	dropped	5000 lands under-lock	dropped
three-lock chain ({c,b})	{c,a}	{c,b,a} over-locks a	{c,b}	{c,b}	refused
four-lock chain ({c,z,y})	{c}	{c,x,z,y} over-locks x	{c,z,y}	{c,z,y}	refused
two-lock cycle (NONE)	{a}	{a,b}	{a,b} fail-safe	{b}	refused
FK's own record lock reads a stage its own lock keeps	FK lands, amount 999	FK stays, 100	FK stays, 100	FK stays, 100	refused

• (i) monotone fixpoint. Never opens a lock, and its first pass is the old single pass. It over-locks: a key dropped in an early pass whose own lock is FALSE on the final row. Measured on a two-field shape (REVERSE, where the base over-locks too, so (i) keeps that defect) and on the three- and four-lock chains (the four-lock x is a new over-lock).
• (ii) judge every lock against the stored row. It reverts every judged key, a key's own value included. It over-locks the legitimate reopen, and it opens locks the base holds: close + edit writes the amount onto a now-closed row, and a lock that reads its own field lets 5000 past a 1000 cap. It also reads record as previous for those fields. Rejected.
• (iii) refuse. Moves accept to refuse on every interacting shape. The documented behaviour is to ignore a locked write, not refuse it (quoted below). Rejected. Measured in its narrowest form, refusing only when a drop moves another lock's verdict; a syntactic "reads a dropped field" rule would refuse a superset.
• (iv) chosen. (i), then ONE release: every dropped key that is unlocked on (i)'s row is released at once. The result is kept only if it is exact (every dropped key locked, every kept key unlocked, on the row it stores). Otherwise (i)'s answer stands. It never opens a lock: no field is written while its lock is TRUE on the row the update stores. It reaches the unique exact drop set on every measured shape except the three-lock cascade (A three-lock readonlyWhen cascade drops a write whose own lock is FALSE on the stored row: a legitimate edit is silently ignored #19927): there, one release step does not settle the set (releasing x moves y's verdict), so the fixpoint's larger drop set {c, x, y} stands instead of {c, y}, and a field whose own lock is FALSE on the stored row is still dropped, as on base. The cycle has no exact set at all and takes the same fail-safe answer. It changes no documented behaviour: every movement below brings the stored row into line with the documented rule. It keeps every existing pin, as all candidates do.

The fix

• packages/objectql/src/validation/rule-validator.ts
  • settleReadonlyWhenDrops: the fixpoint and the exact release. Each key is judged with its own incoming value (a lock reading its own field judges the write) and the other drops reverted. Views are memoised per drop set.
  • stripReadonlyWhenFields and stripReadonlyWhenFieldsMulti now route through it, with no new arguments at existing call sites. Bulk means "locked in at least one matched row" for every evaluation.
  • Warnings come from each key's deciding evaluation and are logged once, in declaration order.
  • ReadonlyWhenStripOptions.only: the strip judges every key but takes, and speaks about, only that key, and only when it is taken.
  • readonlyWhenFkJudgementReadsParent: the settlement's question "does judging the FK's lock need the header it names?".
  • The parent-root reader now caches roots per source, so it answers for record too.
  • Neither the options type nor the strips are exported from the package (index.ts and core.ts export none of them), so no published surface moves.
• packages/objectql/src/engine.ts, settleMasterDetailLanding and its two judgeFkLock closures
  • The FK's own lock is judged with every caller-supplied lock (only: fk) instead of alone. This is how a value another lock drops is reverted before the FK's lock reads it.
  • The named header is read when the FK's lock reads parent (as before), or reads record while another payload key's lock reads parent.
  • A landing FK now stays in supplied, so the strip that judges the rest re-judges it on the same landing and settles the same drop set. An FK that does not land keeps PR fix(objectql): judge a parent-scoped readonlyWhen against the header the write stores (#19853) #19877's rule: its verdict is final and never re-asked against the header the row keeps.
  • One new import line, placed away from the rule-validator import line PR feat(formula,objectql): read one hop through a lookup in a validation predicate #19728 edits.

Every movement (base a34c27cbe5 to head d01b912f3d; patch round 1 changed no code, so 611a2fc561 answers the same)

write	before	now
the card, by id and bulk	amount: 999; event [status]	amount: 100; one event [status, amount] readonly_when
the card under strictReadonlyWrites	refused, fields: ['status']	refused, fields: ['status', 'amount'], drops one readonly_when event (a refusal before and now)
the card, isSystem / preserveAudit / a hook echoing the caller's status	amount: 999	amount: 100
REVERSE by id and bulk (frozen status)	amount dropped; event [status, amount]	amount: 999 lands; event [status]
REVERSE under strictReadonlyWrites	refused, ['status', 'amount']	refused, ['status']
own-field lock beside a dropped reopen	amount: 500 on a closed row	amount: 100
three-lock chain	stored {c: L, b: x, a: 1}	stored {c: L, b: y, a: 2}
two-lock cycle	{a} dropped	{a, b} dropped (no exact set exists)
requiredWhen requiring note while the amount is above 500, on the card	refused VALIDATION_FAILED	commits, amount 100
requiredWhen requiring note while the amount is below 500, the card clearing note	commits {closed, 999, null}	refused VALIDATION_FAILED, nothing lands
settlement, FK lock reads a stage its own lock keeps, by id and bulk	line moves to inv_b, amount 999	stays under inv_a, amount 100; event [stage, invoice, amount]
settlement, stays to moves: the FK's own record lock reads a value that is itself locked under the header the update names (invoice locked by record.amount == 'big', amount by parent.status == 'paid'; line under open inv_b, the update names paid inv_a with amount: 'big'), by id and bulk	line stays under inv_b and stores amount: 'big'; event [invoice]; header reads [inv_b]	line moves to inv_a and keeps amount: 'small'; event [amount]; header reads [inv_a], then the repoint's reference check on inv_a. Both rows agree with their locks
the same write under strictReadonlyWrites	refused, fields: ['invoice']	refused, fields: ['amount'], nothing lands
three-lock cascade (c locked by previous.c == 'L', x by record.c == 'open', y by record.x == 'xv'; row c: 'L'; the update sets all three), by id and bulk	drops {c, x, y}	drops {c, x, y}, unchanged; the exact set is {c, y} (#19927)
same FK lock, no parent-scoped lock (plain strip)	FK moves	FK stays
PR #19877's inv_line_moored shape (stage not in the payload)	header reads [inv_a]	[inv_b, inv_a]: one extra read, same row, same event

Unmoved (pinned): both controls; a hook-written status and a hook overwriting the caller's status (both land and are read); a hook-written amount; the legitimate reopen by id, bulk and strict (nothing dropped); close + edit; an own-field lock over and under its cap; an FK lock reading only previous (one header read); a landing FK's fail-open fault warning, said once. PR #19905's 18 pins and PR #19877's 38 (its whole file) are green.

Warn lines: a newly dropped key gains its drop line and a released key loses it. A key re-judged in a later pass speaks from its deciding evaluation, once. A landing FK's fault warnings now print with the rest, in declaration order, instead of first. When the FK stands on its own lock but the static strip takes it, its conditional fault warning is no longer printed.

A3: all four call sites

By-id strip, bulk strip, by-id judgeFkLock and bulk judgeFkLock all run the same settlement; no site is exempt. The FK's verdict is consistent when it lands: the re-judge is the same computation over the same header. When the FK's lock reads neither record nor parent, its verdict does not depend on the view. PR #19905's stored view is the base every view is built from, so a value the static strip takes is never read either.

Tests

New packages/objectql/src/engine-readonly-when-interdependent-locks.test.ts, 35 cases (32 in round 1, plus three stays-to-moves pins by id, bulk and under strictReadonlyWrites in patch round 1): the card by id and bulk with both controls, the report and warn lines, the strict envelope (code ERR_READONLY_FIELD_REJECTED, name, fields, drops; the engine error carries no HTTP status, which is mapped downstream and not measured here), isSystem, preserveAudit, four hook shapes, the legitimate reopen (by id; bulk plus strict), close + edit, REVERSE by id / bulk / strict, the own-field lock (two cases), the chain, the cycle, both requiredWhen directions (refusal asserted on code VALIDATION_FAILED + fields), the settlement by id and bulk with its header reads, the plain-strip FK, the moored read count, the previous-only FK, the landing FK's single warning, and two unit cases for only.

On d01b912f3d (patch round 1):

• pnpm --filter @objectstack/objectql exec vitest run --project local --maxWorkers=2: 307 files / 5158 tests passed, exit 0.
• pnpm --filter @objectstack/objectql typecheck: exit 0, check:test-typecheck: OK. The new file is in the tsconfig.test.json program (--listFiles: 1) with 0 diagnostics. pnpm --filter @objectstack/objectql test:repo: 5 passed.
• Ablation (fix committed; engine.ts and rule-validator.ts restored to their a34c27cbe5 blobs e973ab50ac / 2b3002b8e1 with git restore --source, tree only; on-disk hashes verified and markers settleReadonlyWhenDrops / readonlyWhenFkJudgementReadsParent counted 0 / 0). The three suites ran 23 failed / 68 passed: every new pin failed except the 12 whose behaviour is unchanged, the three stays-to-moves pins included, and PR fix(objectql): judge a record-scoped readonlyWhen against the payload the write stores (#19887) #19905's 18 and PR fix(objectql): judge a parent-scoped readonlyWhen against the header the write stores (#19853) #19877's 38 stayed green. The subject is imported from source (./engine.js), so no dist sits on the path. Restored with git checkout HEAD --: blobs c9b1cfc19e / f3934b80f6 equal HEAD, git diff HEAD empty, rerun 91/91 passed. The script carried an EXIT / INT / TERM trap.

Gates on 611a2fc561

Patch round 1 (d01b912f3d) changed docblocks, the changeset and three pins over the same four paths. It re-ran the objectql tests and typecheck, node scripts/check-issue-citations.mjs (23 citations resolve, #19927 included), node scripts/check-changeset-no-major.mjs --base origin/main, pnpm check:nul-bytes and the narrowed eslint run: exit 0 each. The full derivation below ran on 611a2fc561.

• node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack: 65 commands, each run with its exit code captured before any pipe. --ran: 65 derived famil(ies) accounted for — 65 run, 0 NOT-MEASURED (a DERIVED zero …), exit 0. check:dual-build-cjs-loads, check:lean-entry-closure and check:type-check-debt first answered PREREQUISITE NOT MET (exit 3, no workspace dist). After pnpm exec turbo run build --filter='./packages/*' --filter='./packages/*/*' --concurrency=2 (72/72 tasks) all three exit 0, and the record carries the reruns.
• node scripts/check-issue-citations.mjs (live, the verdict CI blocks on): exit 0, 22 citations resolve.
• node scripts/check-system-context-census.mjs: exit 0 (no new isSystem read). pnpm check:query-options-erasure: exit 0.
• The three roster gates the derivation marks as keeping a roster under this diff's directories: node scripts/check-changeset-fixed.mjs, pnpm check:authz-resolver, pnpm check:filter-alias-parity, exit 0 each.
• Lint, narrowed and proven. eslint's own config (ESLint.isPathIgnored / calculateConfigForFile) ignores the changeset and lints the three TypeScript files with typescript-eslint/parser and no parserOptions.project / projectService. eslint --no-inline-config --format json over them gives 3 results, 0 errors, 0 warnings. The config enables no type-aware linting, and its only file reads are two baselines this diff does not touch, so this diff cannot move a verdict on any untouched file. pnpm lint itself is CI's.

Neighbour PR #19728

This PR stays textually disjoint from it and does not wait on it. Driver-free bare probe (git clone --bare --shared, no merge.* driver registered): merge-tree --write-tree of d01b912f3d (and before it 611a2fc561) against its head 3b9c5f2fca exits 0, and against origin/main beac798026 exits 0. The merged tree carries both changes (readonlyWhenFkJudgementReadsParent 3 in engine.ts, settleReadonlyWhenDrops 6 in rule-validator.ts, #19728's RelatedRecordBinding 2 and 5).

Acceptance notes

• Cycle residue. When locks read each other in a cycle, no drop set can make every dropped key locked AND every kept key unlocked on the stored row, and no rule can promise both halves. This one keeps the fail-safe half (a lock that cannot be settled is not waived), as the bulk strip's "locked in at least one row" rule already does.
• Cost. A write where nothing locks runs one pass, as before. Each drop adds one pass over the standing keys and one re-check of the dropped keys; a full check pass runs only when something is over-locked. The settlement reads one extra header in the moored shape. Noted, not measured.
• In-tree usage. The tree has no case of one readonlyWhen reading another readonlyWhen field. Examples and platform objects were scanned: the showcase invoice locks read status / parent.status, and neither carries a lock. The card's shape is the natural "a closed case stays closed" plus "a closed case's amount is frozen".
• User docs are untouched and still true after the fix; see the declaration note.

Declaration note for the contract review

The claim declares no widening; copied above as given. The fix moves a refuse/accept answer only through the validation rules that run on the stripped payload, and moves strip verdicts toward the documented rule:

• Accept direction. A write refused VALIDATION_FAILED only because an amount the lock should have held raised a requirement now commits without it. The REVERSE shape's amount, which the base silently dropped, now lands. Both remove a verdict the published text already negated. content/docs/data-modeling/fields.mdx ("Conditional Logic") says readonlyWhen is a "CEL predicate; field is read-only when TRUE", and "The server enforces requiredWhen on submit and ignores writes to fields whose readonlyWhen predicate is TRUE". ("Who the lock applies to") says "A TRUE readonlyWhen predicate locks the field for every API-boundary caller". content/docs/data-modeling/formulas.mdx binds record to "the row being evaluated". On the row the update stores, the REVERSE amount's predicate is FALSE and the card's is TRUE.
• Refuse direction. A write that cleared a field the kept amount requires is now refused: a restored guarantee.
• strictReadonlyWrites: every shape that moves carries a conditional drop, so it was a refusal before and still is; only fields / drops move, and they grow on the card and shrink on REVERSE.

Every other movement is a strip (the write still commits). No authorable key, export or error code moves.

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

15 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

• 1 anchor(s) matched too much of the corpus to be a work list: ObjectQL (symbol, 68 pages)
• 2 name(s) were too generic to anchor anything (single lowercase words)
• the SDK route bridge reached 60 of 215 client-bound route-ledger rows — the other 155 have no registrar path: tail to select them, so pages documenting THEIR client methods cannot appear above, on this or any run. Of those 155: 0 are remediable by widening that discovery convention (an in-repo file declares the path; the convention did not scan it); 55 are structural — on a ledger where NOT ONE row is declared in-repo, so no discovery change reaches them at any price; 100 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.
• a key NAME is not a key, so the hand re-read the line above prescribes can land on the wrong schema. The same spelling is authorable on one governed type and a [REMOVED] tombstone on another for each of active, aria, joins, objects, template, tools and version (censused on [finding] tools is a key on BOTH AgentSchema (tombstoned, dead) and SkillSchema (live, cloud-attested), so a name-based search attributes skill examples to the agent key — it produced a false stop-the-line alarm on PR #19059 #19093 over the liveness ledger's governed types, top-level keys); nothing in a search result distinguishes the two, so a grep hit on a LIVE example reads as evidence about the DEAD key. Measured on fix(spec): the agent.tools liveness row says dead — it claimed live on a key the schema tombstoned #19059: content/docs/ai/agents.mdx was reported as contradicting the agent.tools tombstone over its tools: example at :161, which is inside the defineSkill({ block opened at :155 — the page was already correct. Settle ownership by PARSING the value against both schemas, never by the name: that literal PASSES SkillSchema, and as an AgentSchema it FAILS at tools with the tombstone prescription. ⛔ These names are not the whole class — a key retired through a .strict() guidance map leaves no tombstone in the walked shape and none of them here (tool.category, live as AIToolDefinition.category).

Coarse fallback — 17 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 beac7980267a361c5979ca7aa75c034624955af0 → packageMentionDocs.

Which tree this was computed on

This run read content/docs from e4b2798101416c9b090455002b43ca472c360f0e — the merge of head d01b912f3da551e13c6f561eda44ac7bd1008d20 into base beac7980267a361c5979ca7aa75c034624955af0, 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 e4b2798101416c9b090455002b43ca472c360f0e && git checkout e4b2798101416c9b090455002b43ca472c360f0e
# afterwards, rebuild it from the two parents, which stay fetchable
git fetch origin beac7980267a361c5979ca7aa75c034624955af0 d01b912f3da551e13c6f561eda44ac7bd1008d20 && git checkout -B drift-repro beac7980267a361c5979ca7aa75c034624955af0 && git merge --no-ff d01b912f3da551e13c6f561eda44ac7bd1008d20

node scripts/docs-audit/affected-docs.mjs --json beac7980267a361c5979ca7aa75c034624955af0

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