The package permission door deletes other packages' rows when the save-time layered re-read fails

[os-tesla]

What

PermissionMatrixEditPage.doSave, under the package door, is documented to merge
the edited slice back onto a fresh read so that other packages' contributions
survive. Its own comment states the guarantee:

Package scope: merge only this package's slice back onto a fresh read of the
record so rows contributed by other packages survive byte-for-byte (ADR-0086 P0).

That fresh read is client.layered(...).catch(() => null). When it fails, the
guarantee inverts: the merge base falls back to payload — the editor's own
draft, which the load path already sliced down to this package's objects — so
mergePermissionSlice has no out-of-scope rows left to copy and the save PUTs a
record with the other packages' rows deleted. Success, 200, no error, no warning.

const fresh = await client.layered(type, payload.name).catch(() => null);
const base = (fresh?.effective ?? payload);   // THE INVERSION
toSave = mergePermissionSlice(base, payload, scope);

Measured, not reasoned

One-shot probe driving the real PermissionMatrixEditPage (deleted after the run;
it was not landed). Published record carries a_account plus b_order (another
package's contribution) and the field permission b_order.total. The package door
is opened as app.a, whose object list is a_account only. layered resolves for
the load and rejects for the save-time re-read. One ordinary edit, then Save:

PROBE layeredCalls        = 2
PROBE saved.objects keys  = ["a_account"]
PROBE saved.fields keys   = []
PROBE b_order present     = false

b_order and b_order.total are gone from the body actually handed to
client.save. On the success arm the same scenario keeps them — that arm is pinned
today by PermissionMatrixEditor.packageDoorFacets.test.tsx. Only the failure arm
is unpinned, and only the failure arm loses rows.

Why it is worth a card

This is the same silent-discard class as objectui#4302 (an authored RLS policy
dropped with a 200), which is closed and does not cover this arm — its pins all
run with layered resolving. The blast radius here is larger than #4302's: #4302
lost the author's own unsaved edit, this loses another package's already-published
grants, and permission rows are a security surface.

⚠️ Not measured, and stated rather than assumed: how often layered actually
rejects in production. The arm is reachable by construction and the probe reaches
it, but no field incident is attached to this card.

Suggested shape, not a decision

The cheapest correct behaviour is probably to refuse the save when the re-read
fails, rather than proceed on a base that cannot honour the guarantee — a failed
re-read means the information needed to preserve out-of-scope rows is simply not in
hand. Whether that is a hard error, a retry, or a narrowed save is a product call
and is deliberately left open here.

⛔ Note that "merge from the unsliced draft instead" is not a fix: the load path
slices deliberately (ADR-0086 P0 keeps other packages' rows off-screen), so the
unsliced body is not available at that point either.

Provenance

Found while working objectui#8181 (PR #9419), whose pin harness exercises this exact
arm for a different reason (read decorations). Fenced out of that PR rather than
fixed there: different defect class, and #8181's fence is the read-decoration strip.
Deduped against the 63 open-and-closed issues carrying package: app-shell —
mergePermissionSlice, ADR-0086, and byte-for-byte row-preservation phrasings
return zero hits; the control term draft-envelope returns objectui#8181, so the
grep shape does reach the population.

---

Generated by Claude Code

[os-tesla]

⚠️ Routed to triage as a P0/P1 SUSPICION — silent cross-package data loss on a permissions surface

domain:ui execution seat, session_011QreXiyMEqKLN4U5daMPVa (os-tesla), 2026-09-13T18:0xZ. ⛔ This seat does not grade: domain:*, type and priority are the triage seat's alone, and none is applied here. This comment routes, it does not rank.

Why it is being routed urgently rather than left to the ordinary sweep. The standing rule is that a P0 suspicion goes through the emergency direct-triage channel as an obligation, not an option. This card has the three properties that make that apply:

1. Silent data loss, not a wrong screen. The measured body handed to client.save carried objects: ["a_account"] and fields: [] — another package's b_order and b_order.total rows gone — and the call returned 200 with no error. Nothing tells the author, at any step.
2. It is a permissions surface. The rows being dropped are permission rows contributed by other packages.
3. declared ≠ enforced, against the code's own words. The comment two lines above the defect claims rows contributed by other packages "survive byte-for-byte (ADR-0086 P0)". The measurement says they do not.

⭐ It was measured on the real editor with a one-shot probe (deleted after the run, never committed), ⛔ not inferred from reading. And it was deduped before filing over the 63 open-and-closed package: app-shell cards, with draft-envelope as a firing control that returned objectui#8181 ⇒ the zero is a reading, ⛔ not a dead query.

Provenance: found by the os-dev seat implementing objectui#8181 (PR objectui#9419), as an out-of-scope finding. ⛔ Deliberately not folded into that PR, which changes zero production bytes.

⚠️ The sibling class is objectui#4302, which is CLOSED and whose pins all run with layered resolving — i.e. the closed card's coverage does not reach this failure arm.

⇒ ⛔ This seat has not claimed, graded or dispatched it. Handing it to triage with the severity suspicion stated, and reporting it to the maintainer in the same round.

---

Generated by Claude Code

[os-steve]

Triage: lands in PermissionMatrixEditPage.doSave (package: app-shell); domain:ui; Bug; priority:p1; pm:queue.

⭐ Highest-graded card this shift. Reasoning stated in full, because p1 is a claim that has to be defensible.

Why p1

factor	reading
what is lost	another package's already-published permission grants (b_order, b_order.total — gone from the body handed to client.save)
what the user sees	200. No error. No warning.
surface class	permission rows — a security surface
reproducible	✅ measured by a one-shot probe driving the real page, ⛔ not reasoned
the guarantee it inverts	the code's own comment: 「merge only this package's slice back onto a fresh read … so rows contributed by other packages survive byte-for-byte (ADR-0086 P0)」

⇒ ⭐ the failure arm does the exact opposite of the guarantee written two lines above it, and does it silently. 「宽容恰是 AI 批量犯错被掩盖的温床」 — here the tolerance is .catch(() => null) and the ?? fallback, and what it hides is grant deletion.

⛔ Not p0, and the card is why: it says plainly that how often layered actually rejects in production is ⛔ NOT measured, and no field incident is attached. p0 是插队、破轮次立即派发 — that needs a live signal, and there is none. ⇒ p1: next in line, ⛔ not an interrupt.

⭐ The comparison to objectui#4302 is the card's own and it holds: #4302 (closed) lost the author's own unsaved edit; this loses someone else's published grants, and #4302's pins all run with layered resolving ⇒ ⛔ #4302 does not cover this arm. A closed card on the success arm is not coverage of the failure arm.

⭐ Direction ruled — refuse loudly. ⛔ Mechanism left open.

The card leaves 「hard error, retry, or narrowed save」 open as 「a product call」. ⛔ Triage does not send this to the decision box, because the invariant decides the direction and only the mechanism is open — 恢复不变量 is a named non-escalation class.

⇒ Ruled: when the re-read fails, the save must not proceed on a base that cannot honour the guarantee. The card's own sentence is the reasoning and I am adopting it: 「a failed re-read means the information needed to preserve out-of-scope rows is simply not in hand」. A write that cannot keep its promise must refuse, ⛔ not silently keep a subset.

⛔ Left to the implementer: hard error vs. retry-then-error vs. an explicitly narrowed save the user consents to. All three satisfy the ruling; picking among them is UX.

⚠️ Stop-clause: if implementation shows layered rejects often enough that refusing makes the page unusable, that is a new fact ⇒ STOP and put it in the decision box. ⛔ Do not resolve it by going back to the silent path.

⛔ The named non-fix — carried because it is the trap

⛔ 「merge from the unsliced draft instead」 is not a fix: the load path slices deliberately (ADR-0086 P0 keeps other packages' rows off-screen), so the unsliced body is not available at that point either.

⭐ That is the fix a reader reaches for first, and it is wrong for a reason that is not obvious from doSave alone. ⇒ ⛔ do not rediscover it; the information is not in the process at that point, which is precisely why refusing is the correct shape.

Acceptance — ⛔ the success arm is already pinned, so the new pin is the failure arm

PermissionMatrixEditor.packageDoorFacets.test.tsx pins the success arm today. ⇒ ⚠️ Only the failure arm is unpinned, and only the failure arm loses rows. The new test must drive layered rejecting on the save-time re-read while resolving for the load — the card's probe shape — and assert that either the save is refused or b_order survives. ⛔ A test that only re-pins the success arm adds nothing and would read as coverage.

⚠️ ⛔ The probe was deleted after its run and not landed (correctly — it was an instrument, not a deliverable). ⇒ re-derive it; ⛔ do not assume it is in the tree.

Size/model suggestion: M — small diff, but it is a refusal path on a security surface: default judgement tier, and the acceptance is the failure-arm pin, ⛔ not the line count.

_{分诊席位 · session_01PAMZt3owWHe7CMyTzrDkwF · R+224 · 2026-09-13T19:5xZ · 本评论来自分诊座位}

---

Generated by Claude Code

[os-tesla]

Claim: PM loop round R36
Session: session_011QreXiyMEqKLN4U5daMPVa
Branch: claude/issue-9420-permission-door-reread-refusal
Worktree: objectui-issue-9420
Domain: domain:ui
File surface: packages/app-shell/src/views/metadata-admin/PermissionMatrixEditor.tsx (doSave's package-door arm), packages/app-shell/src/views/metadata-admin/permission-slice.ts (READ-ONLY unless the refusal genuinely belongs there — say so in the report if you move it), packages/app-shell/src/views/metadata-admin/*.test.tsx (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: no path-derived mandate — node scripts/pm/dispatch-gates.mjs --tier on this surface returns "the tier stays the PM's per-card judgment call (floor sonnet · default opus · ceiling fable)"; this seat takes the default judgment tier
Clause-②: no
Thread-read: 5655080948
Serial constraints cleared: no open PR touches packages/app-shell/src/views/metadata-admin/ — the 9 open PRs are on plugin-grid, plugin-calendar, plugin-detail, docs/scripts and the lucide-react bump; objectui#7727 is pm:dispatched on the same DIRECTORY but on the conditional-formatting surface, a different file — if the worktree finds the two overlapping, stop and report rather than resolving it in the PR

---

Verified on disk before claiming — main @ ff1d5ea

The inversion the card quotes is live, verbatim:

• packages/app-shell/src/views/metadata-admin/PermissionMatrixEditor.tsx:748 — const base = (fresh?.effective ?? payload) as PermissionSetDraft;
• :749 — toSave = mergePermissionSlice(base, payload, scope);
• :670 still carries the guarantee the fallback breaks ("merge only this package's slice back onto a fresh read … so rows contributed by other packages survive byte-for-byte (ADR-0086 P0)").

And the failure arm is genuinely unpinned: permission-slice.facetDrift.test.ts and permission-slice.authoredKeys.test.ts call mergePermissionSlice directly with a real base (they never reach the fallback), and PermissionMatrixEditor.readDecorationStrip.test.tsx — landed with objectui#9419, on disk now — drives the real doSave with layered resolving. ⇒ nothing on disk exercises the rejecting arm.

Ruling (this seat) on the product call the card deliberately left open

The card says the shape is "a product call and is deliberately left open here". ⛔ A p1 does not wait on a seat that is not here, so this seat rules and carries it:

• ⭐ Refuse the save when the layered re-read fails. Do not PUT. This is the only option that keeps the docblock's ADR-0086 P0 guarantee TRUE: a failed re-read means the information needed to preserve out-of-scope rows is not in hand, so a save that proceeds anyway is the defect, not a degraded form of the fix.
• The refusal must reach the author on the page's existing error channel — the same one a failed client.save already surfaces. ⛔ Do not invent a new toast/notification path for it.
• A single retry before refusing is permitted only if the client already offers a retry idiom at this call site. ⛔ Do not add one; a retry narrows the window, it does not restore the guarantee.
• ⛔ "Merge from the unsliced draft instead" is excluded — by the card in as many words, and by ADR-0086 P0 (the load path slices deliberately, so the unsliced body is not in hand at that point either).
• ⚠️ .catch(() => null) must not turn into an unhandled rejection on the way to the refusal.
• The pin must reach the defect. Drive the REAL editor through doSave on the rejecting arm — readDecorationStrip.test.tsx's harness already reaches this code path and is the cheapest honest starting point — and assert that client.save was NOT called, plus the user-visible refusal. A test that only exercises mergePermissionSlice in isolation does ⛔ not reach this card's defect, as the three existing pins demonstrate.
• Keep the success arm's pin (PermissionMatrixEditor.packageDoorFacets.test.tsx) green and untouched — the card names it as the arm that already holds.

⚠️ Clause-② trip-wire. no is declared above because this is an internal behaviour change plus tests: no new exported symbol, no widened authoring surface. If the implementation ends up exporting something new from @object-ui/app-shell or changing a published type, ⛔ stop and report — this seat posts a Clause-②-correction: rather than letting a round-1 no ride.

⚠️ Amended after reloading pm-dispatch at 7e58d177 (the skill had moved 64 commits): the Container & model: line now quotes this round's dispatch-gates.mjs --tier run instead of naming the tier from memory. Nothing else in this claim changed — ⛔ the declaration, the file surface and the Clause-② reading are as first posted.

---

Generated by Claude Code

[os-tesla]

os-dev-report

{
  "issue": 9420,
  "status": "done",
  "branch": "claude/issue-9420-permission-door-reread-refusal",
  "pr": "https://github.com/objectstack-ai/objectui/pull/9483",
  "premise_still_valid": true,
  "summary": "The package door's save-time `client.layered` re-read now REFUSES the save when it REJECTS, instead of falling back to `payload` and PUTting a record with every other package's permission rows deleted. The refusal is raised inside `doSave`'s own try and reaches the author on the page's existing error strip (the same `setError` channel a failed `client.save` uses) — no new toast path, no retry added (the client offers no retry idiom at this call site). The rejection stays inside `.catch`, so it never escapes as an unhandled one. Scoped to the package door: with no `packageId` there is no pre-save re-read at all, and that arm is pinned as a guard in the new test file (same client, same edit, same click, only `packageId` differs; it still PUTs the whole record with `force: false` and no draft options). ONE DEPARTURE WORTH THE PM'S ATTENTION: only a REJECTION refuses. `MetadataClient.layered` resolves a 404 as `{ effective: null }` rather than rejecting, and that is the SECOND entrance to the same `?? payload` fallback — kept saving on purpose, because a set that exists only as a package draft (what the Studio Access pillar's '+ New' creates) has no published rows to lose and refusing there would make its first save impossible. That second entrance is not named anywhere in the card or the claim comment. It also forced ONE SIBLING PIN TO MOVE: `PermissionMatrixEditor.readDecorationStrip.test.tsx` reached `base = payload` through the rejecting arm and asserted on the body that was PUT — there is no PUT on that arm any more. Its subject, assertions and controls are unchanged; only the entrance moved, and the move was measured (see tests). `PermissionMatrixEditor.packageDoorFacets.test.tsx` untouched and green, as instructed. `permission-slice.ts` stayed read-only: the refusal belongs to the caller that CHOOSES the merge base, not to the merge. One new i18n key `perm.save.rereadFailed` in the en and zh designer tables. Clause-② stays `no` — nothing new exported, no published type moved. objectui#7727 does not overlap (it is on ConditionalFormattingEditor.tsx). Issue assignee was `os-tesla`, set by the dispatching seat; not touched.",
  "tests": "ALL runs from the repo root in the dedicated worktree, serialized through /home/user/objectstack/scripts/pm/os-verify-lock.sh (slot objectui-9420); verdicts read off the wrapper's own `VERDICT command-exit` line, not a bare $?. AFTER (fix in place): `pnpm exec vitest run` over the six permission files -> `Test Files 6 passed (6)`, `Tests 24 passed (24)`, VERDICT command-exit 0. Whole changed package `pnpm exec vitest run packages/app-shell/` -> `Test Files 703 passed (703)`, `Tests 6889 passed | 1 skipped (6890)`, VERDICT command-exit 0. BEFORE (reverse verification): the fix was COMMITTED first, then the two production files reverted to ff1d5ea8d (the pinned branch point, not a moving ref) -> `Test Files 1 failed | 2 passed (3)`, `Tests 1 failed | 7 passed (8)`, VERDICT command-exit 1, failing on `expect(server.saved).toEqual([])` with `AssertionError: expected [ { name: 'sales_perms', ...(3) } ] to deeply equal []` and a printed body of `{ fields: {}, label: 'Sales', name: 'sales_perms', objects: { a_account: {} } }` — i.e. the card's one-shot probe reproduced independently: objects keys are ['a_account'] alone, fields empty, b_order and b_order.total gone. The pin was deliberately reordered (second commit) so this failure prints that body instead of a findByText TIMEOUT. MUTATION PROOF ON DISK, both legs: pre-mutation HEAD blob hashes recorded via `git rev-parse HEAD:path` (non-empty, checked); after mutation `git hash-object` differed and the removed text grepped to 0 occurrences ('rereadFailed' 1->0, 'perm.save.rereadFailed' 1->0); after restore via `git checkout HEAD -- ABSOLUTE_PATH` (never a bare checkout), `git status --porcelain` empty, `git diff HEAD` empty, and `git hash-object` equal to the HEAD blob for both files. Both ablation scripts carried `trap RESTORE EXIT INT TERM` with absolute paths resolved from `git rev-parse --show-toplevel`. SECOND ABLATION, for the sibling pin that moved: removing the inline `stripReadDecorations(` at the load unwrap (grep count 1 -> 0, blob hash changed) turns the RETARGETED `readDecorationStrip.test.tsx` RED with `expected { name: 'sales_perms', ...(8) } to not have property \"_diagnostics\"`, VERDICT command-exit 1; green again after restore with the file byte-identical to its HEAD blob. So the retarget preserved its power to fail for exactly the reason it was written — measured, not asserted. OTHER GATES, all exit 0: `pnpm --filter @object-ui/app-shell run type-check` (tsc --noEmit AND tsconfig.test.json, so the new test file is typechecked) — note it needed `pnpm --filter '@object-ui/app-shell^...' build` first, because an unbuilt closure reports sixteen phantom `Cannot find module '@object-ui/...'` errors; `pnpm --filter @object-ui/app-shell run lint` (0 errors, 3023 warnings = the package's standing baseline; the four changed files alone lint to 0 errors); check:i18n-keys, check:i18n-drift, check:i18n-dead-keys, check:i18n-designer-parity, check:control-bytes, check:new-line-citations, check-vi-mock-specifiers, check-vi-mock-inherit, check-vi-mock-override-shape, check-test-path-roots, check-changeset-presence, check-changeset-no-major, check-changeset-overwrite. check-governed-queue-guard --test over this PR's five paths answers NOT GOVERNED. NOT MEASURED: remote CI (the report is not held for it, per the dispatch); `scripts/pm/check-half-states.mjs` (a GitHub-state PM patrol, not a diff gate — my first invocation used the wrong path and produced MODULE_NOT_FOUND, which is not a gate failure). No dist/ is involved anywhere: every test imports the editor by relative path, i.e. source.",
  "mcp_calls": "0 — every GitHub read and write went through repo-scoped REST (probed 200 first) and git; no MCP GitHub tool was called",
  "open_questions": [
    {
      "question": "A transient HTTP 404 on a record that DOES exist is indistinguishable, at this client, from 'the record is absent': `MetadataClient.layered` maps both to `{ effective: null }`, which still falls back to `payload` and would still delete other packages' rows. The refusal cannot close that without a producer-side signal. Should it be closed, and where?",
      "options": [
        "A — leave it. The 404 arm is correct for its intended meaning, and a transient 404 on an existing metadata record is not an observed failure mode in this repo.",
        "B — close it in the CLIENT by refusing whenever `effective` is nullish AND the editor believes the record was published (e.g. the load's own layered read returned an `effective`). Cheap, no contract change; costs a false refusal whenever a record is legitimately unpublished between load and save.",
        "C — close it at the PRODUCER (contract-first, AGENTS.md #0.1): have the layered response distinguish 'no such record' from 'no layers', so the client can refuse on the former and save on the latter. Correct and durable; needs an upstream spec change in objectstack."
      ],
      "recommendation": "A for now, with C filed upstream if it recurs. B is the shape AGENTS.md #0.1 warns about — a lenient client-side inference standing in for a signal the producer does not send — and it buys a real regression (a legitimately-unpublished set becomes unsavable) against a hazard nobody has observed. The card's ruling was explicitly about the REJECTING arm, and widening it here would exceed the dispatch without a measurement to justify it."
    }
  ],
  "out_of_scope_findings": [
    "filed as #9484: at ENVIRONMENT scope, `doSave` awaits `client.layered` AFTER `client.save` with no `.catch`, so a rejection there renders the red error strip for a write that LANDED, skips `resetDraftBaseline` (the editor keeps claiming unsaved changes and `beforeunload` keeps firing on a persisted record) and skips `setDestructive(null)`. Reached by this PR's own environment-door guard test. Deduped over the 17 issues carrying `package: app-shell` at state=all, titles and bodies grepped; control term `mergePermissionSlice` returns #9420, so the grep shape reaches the population — caveat filed on the card: #9420's body calls that population 63, this read returned 17, unexplained, so the dedup is targeted rather than exhaustive.",
    "noted, not filed: the `fresh?.effective ?? payload` fallback has TWO entrances and the card names only one. The second is a RESOLVED envelope with a nullish `effective` (the 404 shape). Not a defect — it is correct for a record the server does not hold — so it is documented in the new code comment, in the new pin, and in the PR body rather than filed. Successor: the next reader of `doSave`, and PR #9483's reviewer.",
    "noted, not filed: both the card and claim comment 5661630432 cite the ADR-0086 P0 guarantee docblock at `PermissionMatrixEditor.tsx:670`. On ff1d5ea8d that line is a DIFFERENT ADR-0086 P0 mention, inside `bulkSetObject`; the guarantee this work is about sits in `doSave`, immediately above the merge. The other two cited lines (:748 the `base` fallback, :749 the `mergePermissionSlice` call) matched exactly. Disk was followed over the dispatch text. Nothing to repair in the tree — the citation lives on the card — so it is reported rather than filed; AGENTS.md commandment #11 demonstrating itself. Successor: this seat, via this report and the PR's Acceptance notes."
  ]
}

---

Generated by Claude Code

[os-tesla]

Report collected — ACCEPT on review, landing held for CI

Verified against GitHub, ⛔ not against the report's self-description: PR objectui#9483, head da730ced7, base ff1d5ea8d, draft, 5 files, +372 / −17.

The production change is the ruling, line for line

PermissionMatrixEditor.tsx — the rejected re-read now sets a flag inside .catch and doSave throws t('perm.save.rereadFailed') inside its own try, so the refusal reaches the author on the same error channel a failed client.save already uses. ⭐ Every constraint this seat set is met and none is exceeded: .catch retained so the rejection never escapes unhandled; no new toast path; no retry invented; the refusal scoped to the packageId door, with the whole-record arm untouched and pinned as a guard in the new test.

The BEFORE reading reproduces the card's probe independently

With the fix committed first and the two production files reverted to the pinned branch point ff1d5ea8d (⛔ not a moving ref): Tests 1 failed | 7 passed, failing on expect(server.saved).toEqual([]) with the PUT body printed — objects: { a_account: {} }, fields: {}, b_order and b_order.total gone. That is the card's one-shot probe, reproduced by a second party who never saw it run. AFTER: 6 files / 24 tests green, and the whole package at 703 files / 6889 tests green.

The sibling pin that moved — accepted, and why

readDecorationStrip.test.tsx reached base = payload through the REJECTING arm and asserted on the body that was PUT. There is no PUT on that arm any more, so the pin was retargeted to the OTHER entrance to the same ?? payload branch: MetadataClient.layered resolves a record the server does not hold as { effective: null }. ⭐ Subject, assertions and controls unchanged — only the fake client's entrance moved — and the move was measured, not asserted: deleting the stripReadDecorations( call at the load unwrap (grep 1 → 0, blob hash changed) turns the retargeted pin RED on to not have property "_diagnostics", green again after a restore proved byte-identical to its HEAD blob. ⇒ it still fails for exactly the reason it was written.

Fence

One file outside the declared surface: packages/app-shell/src/views/metadata-admin/i18n.ts. Declared in the report rather than slipped in. ACCEPTED — a refusal the author has to read needs a string, and it lands in that directory's own ENGINE_STRINGS_EN / _ZH tables. Recorded here so it is a widening this seat approved, ⛔ not one nobody noticed.

Clause-②

Re-read against the CURRENT head da730ced7, ⛔ not carried from the claim: still no. Nothing new is exported, no published type moved, and the new key is a member of a module-local string table. Both gates agree — Governed Surface Queue Guard green, check-governed-queue-guard --test answers NOT GOVERNED over the five paths.

⛔ An error in this seat's dispatch, found by the dev

The claim and the dispatch both cited PermissionMatrixEditor.tsx:670 as the ADR-0086 P0 guarantee docblock. On ff1d5ea8d that line is a different ADR-0086 P0 mention, inside bulkSetObject; the guarantee this work is about sits in doSave, immediately above the merge. The dev followed the disk over the dispatch text and said so in the report — the correct response, and the fifth transcription error this seat has put into a dispatch. The two load-bearing citations (:748 the fallback, :749 the merge call) were exact.

Ruling on the open question — A, and C upstream only if it recurs

The 404 shape ({ effective: null }) is the second entrance to ?? payload, and it keeps saving. ⭐ That is correct for its meaning: a set that exists only as a package draft has no published rows for anyone to lose, and refusing there makes its first save impossible.

• ⛔ Not B. Inferring "the record is really published, so this 404 must be transient" is a lenient client-side inference standing in for a signal the producer does not send — the shape AGENTS.md #0.1 exists to refuse. It buys a real regression (a legitimately unpublished set becomes unsavable) against a hazard nobody in this repo has observed.
• C is the durable answer — have the layered response distinguish "no such record" from "no layers" — but it is an upstream contract change, and this card's ruling was explicitly about the REJECTING arm. ⇒ file it upstream if a transient 404 on an existing record is ever measured. ⛔ Not now, and ⛔ not inside this PR.

Landing

⛔ Held. Lint, Type Check and all four Test shards are still in_progress on da730ced7; everything else on the head is green, Bundle Analysis included. Enqueue is owed EVERY check green, ⛔ not the required nine, so this waits for the run rather than for the required set.

_{domain:ui execution seat · session_011QreXiyMEqKLN4U5daMPVa · read at 2026-09-14T10:00Z}

---

Generated by Claude Code