docs(spec): correct the expired browser-only manifest premise in two prose sites (#17735 item 4)

[os-litant]

Part of #17735 — this PR discharges ruling item 4 only. Items 1 and 2 are not implemented, because item 1's stated premise measured false and item 3's escape clause fires. The evidence is below; #17735 remains open for the seat to route the mechanism question.

What landed

Prose-only. Every sentence in two dev-only gate scripts that asserted the expired browser-only manifest claim, or that mis-described what pnpm sdui:manifest does, is corrected. No behaviour, no classification, no verdict set, no baseline moves.

• packages/spec/scripts/check-generated.ts — the EXTERNAL_INPUT_REQUIRED why string (ruling item 4), plus the three other sentences in the same file that stated the same retired reason. The bucket, its membership and runBy are untouched; only the stated reason moves.
• packages/spec/scripts/check-react-blocks-declaration-parity.ts — the gate's own header carried the identical stale premise. It is on this card's declared file surface and is the first thing the next reader of this card will read, so it is corrected in the same PR and named here rather than left to rot.

Every sentence corrected, and what it was measured against

file	asserted	measured
check-generated.ts why	"the registry … exists only after objectui is built … and enumerated in a real browser"	scripts/gen-sdui-manifest-node.mjs enumerates it under plain Node — no browser, no objectui build
check-generated.ts why	"nothing in this repo … can hand it one"	the dump is tracked at the repo root since #13446, and lint.yml feeds it to the gate on every PR
check-generated.ts why	"its only producer is objectui's dump-public-manifest.mjs"	there are two producers, reading two different registries
check-generated.ts why	pnpm sdui:manifest "builds objectui at .objectui-sha"	it builds nothing — see the measurement below
check-generated.ts EXTERNAL_INPUT_REQUIRED doc comment	"the input they compare against does not exist here and cannot be produced here"	it exists here (tracked) and is produced here (two ways)
check-generated.ts print line + its rationale comment	"source audit(s) whose input this repo cannot produce" / "not an option from this repo"	the only true statement is that this aggregate passes no MANIFEST; a reader can run the gate by hand, and CI does
check-generated.ts ledger classification prompt	"audit source against an input this repo cannot produce"	same criterion, same correction — otherwise the file would define one bucket two ways
parity-gate header	"the registry is a browser app … nothing enumerates it from Node"	measured false; the node producer's header carries the readings
parity-gate header	pnpm sdui:manifest "builds objectui from the SOURCE at .objectui-sha"	it builds nothing — see below
both files	the citation objectui#6741	resolves to nothing — see below

Two of those rows — the "builds objectui" clause and the objectui#6741 citation — were re-asserted by this PR's own first pass, inherited from the text it replaced. They are listed here rather than quietly dropped, because a PR that exists to correct false prose shipping new false prose is exactly the failure this card is about.

"builds objectui" — what scripts/gen-sdui-manifest.sh actually does

Read on this head, by line:

• 507 BUILD_ROOT="${FRAMEWORK_ROOT}/.cache/objectui-${PINNED_SHA:0:12}" — derived from .objectui-sha.
• 512-516 if [[ ! -d "$BUILD_ROOT" ]] → prints ✗ objectui build tree not found at … and Run 'pnpm objectui:build' first to vendor the console at the pinned SHA, then exit 1.
• 517-521 if the dump tooling is absent from that tree, prints ℹ … nothing to do and exit 0.
• 373-376 / 545 / 630-631 it then pushdes into that tree and spawns pnpm --filter @object-ui/console exec vite dev --port … --strictPort — a dev server over the already-vendored checkout.
• 654-655 BASE_URL=http://localhost:PORT OUT=… node scripts/dump-public-manifest.mjs.
• 684-688 it runs the ADR-0082 parity ratchet --strict against what it dumped.

So it requires a build (pnpm objectui:build = scripts/build-console.sh) and performs none: it dumps, then ratchets. The script's own comment at 534-538 says the same thing from the other side — "this script drives a vite DEV server over .cache/objectui-SHA".

objectui#6741 — dropped, with a tombstone

The number resolves to nothing in objectstack-ai/objectui: 404 for both the issue and the PR at that number. This is a zero-hit reading with same-subject controls that hit — objectui #6739, #6742 and #6743 all answer 200 on the identical endpoint, so the 404 is the number, not the endpoint or the token. objectstack#6741 is an unrelated ADR-governance ruling, and neither packages/spec/CHANGELOG.md nor the #13446 commit body carries the objectui number.

The fact it was cited for — the registry enumerates under plain Node — is independently reproduced (byte-identical regeneration, below), so the fact stays and the citation goes. The parity header keeps a one-sentence tombstone saying the old citation resolved to nothing, carrying no number, so the next author neither follows a dead link nor "restores" one.

No dead reference is left in either file: git grep -c '6741' is 0 in both, with the node producer's header hitting 1 as the control. That third occurrence is not in this PR's file surface; it is reported for the seat to file.

Why ruling item 1 is not implemented — its premise is false

Item 1 reads: "Where the node generator omits what the runtime registry declares … the node generator is fixed to emit it."

The node generator omits nothing. Measured at every layer it could hide in:

1. scripts/gen-sdui-manifest-node.mjs, read in full (194 lines): it imports the 16 registry modules, calls ComponentRegistry.getPublicConfigs(), and passes the result to manifestFromConfigs. There is no filter, no allowlist, no key removal anywhere in it.
2. manifestFromConfigs (packages/sdui-parser/src/index.ts) copies config.inputs field-for-field. The node generator passes neither only nor publicOnly.
3. The diverging keys are absent from the registry the node generator reads. Published @object-ui/core@17.6.0 contains no withElementDataSourceInput / ELEMENT_DATA_SOURCE_INPUT (control: getPublicConfigs hits in the same dist).

Byte-level proof that the tracked artefact is exactly that registry — regenerating it reproduces the committed file with no diff at all:

$ git hash-object sdui.manifest.json          # 78f870e42fe9fec1e9fad2e17c48370d8eca2cbe
$ node scripts/gen-sdui-manifest-node.mjs     # 57 components, 66910 bytes, sha256 49211fee7792 (== the record)
$ git hash-object sdui.manifest.json          # 78f870e42fe9fec1e9fad2e17c48370d8eca2cbe  -> IDENTICAL

This was reproduced independently in contract review of this PR, from a fresh npm install of the 17 published @object-ui@17.6.0 packages and an esbuild bundle of sdui-parser, reaching the same sha256 49211fee7792 / git hash-object 78f870e42fe9 and a clean cmp. So the artefact is not stale relative to its own input, and there is no omission to remove. "Fix the node generator to emit dataSource" would mean hard-coding a key the registry it reads does not declare.

The actual cause, key by key

The two producers read two different registries:

• node (gen-sdui-manifest-node.mjs) installs the npm-published @object-ui/* at the version named in scripts/sdui-manifest.record.json — 17.6.0, published 2026-08-24.
• browser (gen-sdui-manifest.sh) dumps the registry of the source at .objectui-sha — 53ded82b, authored 2026-09-05 — from a checkout pnpm objectui:build has already vendored.

Every diverging key landed in objectui source inside that 12-day window:

key	root artefact (published 17.6.0)	dump (pinned source)	why
dataSource on 15 blocks	absent	present	emitted by withElementDataSourceInput in Registry.register(), which is not in the published 17.6.0 dist
variant / align on ui:text	absent	present	added to the ui:text registration after 17.6.0. Note element:text already declares all three in both
type vs actionType on the two action blocks	declares type	declares actionType	renamed at the registration after 17.6.0; published 17.6.0 still declares type and reads `d.actionType

The root defect is one false equivalence, and it is in this repo's own prose and procedure: .objectui-sha ships a commit; the @object-ui version read off that commit's packages/core/package.json names a published tarball built from an earlier commit, because objectui bumps version only at release. Every commit between two releases carries the previous release's version number, so the version string cannot distinguish them.

Nothing detects it: scripts/check-sdui-manifest.mjs verifies shape, sha256(artefact) == record.sha256, and record.objectuiSha == .objectui-sha. It reads objectuiPackagesVersion only as a required-field presence check — it never verifies that the version installed corresponds to the pinned commit.

Why item 2 is not implemented — item 3's escape fires

A producer-parity test pinning the two outputs equal would be red today, by construction, and landing a knowingly-red required gate is not an option:

• 17.6.0 is the newest published @object-ui/core (npm time: nothing after 2026-08-24).
• The pin is 12 days ahead of it.
• The node producer can only read npm.

Per ruling item 3 this is reported, and nothing is retired. A candidate mechanism that would reach parity without retiring anything — pointing the node producer at the pin's built tree via its existing --modules-root flag, which build-console.sh already materialises at .cache/objectui-SHA — is unverified: no built objectui tree exists in this container. It is an open question on the card, not shipped here. Choosing among the candidate mechanisms is the maintainer's, and the decision box on #17735 holds it.

Premise checks

• A — defects still present. Reproduced on this base. "dataSource" in the tracked artefact = 0 (control "name" = 339); the false why string = 1 (control EXTERNAL_INPUT_REQUIRED = 12).
• B — the card's census. Reproduces in direction and mechanism, with one correction: the card's «align, variant on text» is about the ui:text key. The artefact does carry element:text with ["content","variant","align"]; it is text (namespace ui) that carries only ["content"]. Both keys exist in both producers — only ui:text diverges.
• C — docs/releases-maintenance.md. Line 118 is a closing code fence; the sentence is at 120-121: "Read that version from the objectui checkout's packages/core/package.json — the bump already required that checkout." It is mechanically accurate and substantively misleading: following it is what manufactures the skew above. #17736 is closed as completed, so it landed without this card's answer. The file is domain:devx and is not edited here — reported for the seat to route, since the correct sentence depends on the producer ruling.
• D — object-tree. Regenerating from the registry does not produce it, in either producer. object-tree is registered by plugin-tree in published 17.6.0 and at the pin, but it is absent from PUBLIC_BLOCKS in both (control: object-grid present), and plugin-tree declares no tier: 'public' opt-in at the pin — so getPublicConfigs() excludes it either way. That is a different defect, living in objectui's public-tier curation. [finding] the tracked sdui.manifest.json has no object-tree entry although plugin-tree registers the renderer — the new object-tree spec row gets no parity comparison at all #18407 is not addressed here and is not subsumed by this PR.

Rework round — what the seat's REWORK verdict asked for, item by item

The contract review of record (CONTRACT_REVIEW_TIER, PASS) and the seat's REWORK answer different questions: the review asks whether this moves an accept set or a public surface (it does not, proven hard); the seat asks whether the PR discharges ruling item 4, whose whole content is 「reason made true」. The review measured the new reason false in two places. Those, and only those, are what this round changes.

1. "builds objectui", asserted twice — corrected in both files, from a line-by-line reading of gen-sdui-manifest.sh (above). The script requires a vendored checkout and refuses without one; it dumps and ratchets.
2. objectui#6741, carried twice — dropped from both files. The fact stays, the citation goes, and a tombstone carrying no number stops the next author restoring it. No real reference was findable: objectui's numbering covers that range (newest is well above it) but the number itself answers 404 for both the issue and the PR, and the GitHub search API is not reachable from this session, so the fact is stated without a number as the verdict permits.
3. The body overstated by exactly two — resolved by correcting the survivals rather than declaring them. Both sentences the review named (the bucket's doc comment, and the print line with its rationale comment) now say the true thing. The count claim is gone from this body entirely: the table above enumerates every corrected sentence instead of counting them, so the claim cannot drift from the diff again.

One correction beyond the two the seat named, in the same file and the same sentence class, declared here rather than left: check-generated.ts's ledger classification prompt ("or audit source against an input this repo cannot produce") stated the selection criterion for the very bucket whose definition this round corrects. Leaving it would have left the file defining one bucket two ways — the same defect at a smaller scale. It is a problem-message string, pinned by no test (control: the ledger test's four toContain pins are 1 needing an external input, cannot run here: check:react-declaration-parity, MANIFEST, scripts/gen-sdui-manifest.sh, all unchanged and all still green).

Not re-opened, per the verdict: ruling items 1 and 2 stay unimplemented (item 1's premise measured false by two independent parties); the mechanism question M1/M2/M3 stays in the decision box on #17735 and is the maintainer's; #18407 stays open and is not subsumed; the check:cross-package-test-inputs red is pre-existing and diff-invariant. The parity header's retained "NOTHING HERE PRODUCES one" (line 98) is scoped by its next sentence to the console-build half, whose sub-claims remain true — it is correct and is deliberately untouched.

Verification

Head ae7e4c916d, merged with origin/main at 21b7c12b4f before measuring. Change set derived from git: exactly the two files, three-dot semantics.

Gates. 54 families derived on this head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands; all 54 run; reconciled with --ran at exit 0 — "54 derived famil(ies) accounted for — 54 run". Exit codes captured with redirect-then-$?, never through a pipe. 50 green, 4 non-zero and every one of them known:

gate	exit	reading
check:react-declaration-parity (bare)	1	by design — it needs MANIFEST=; the CI-shaped run is green, below
check:cross-package-test-inputs	1	pre-existing and diff-invariant; control leg established in the previous round (red on main and head alike, only when packages/spec/dist/ exists)
check:dual-build-cjs-loads	3	PREREQUISITE NOT MET — "this gate reads built output, and some package has no dist/"
check:lean-entry-closure	3	PREREQUISITE NOT MET — "this gate loads BUILT entry points, and some target is absent" (its own self-test passes, 22 cases)

The two exit-3 readings are NOT MEASURED, not red: they need a full pnpm build, which CI does.

Targeted runs on this head:

• pnpm --filter @objectstack/spec build :: exit 0 (34/34 declaration files present) — run first, because two of the gates below read the built dist/
• pnpm --filter @objectstack/spec run check:generated :: exit 0 — "All 15 generated artifacts are up to date", and the corrected print line reads Needs an input this aggregate does not pass (1 source audit(s)):
• MANIFEST="$PWD/sdui.manifest.json" pnpm --filter @objectstack/spec check:react-declaration-parity --baseline react-declaration-parity.baseline.json --strict :: exit 0 — "no new DECLARATION divergence vs accepted baseline"
• vitest run --project repo scripts/check-generated-ledger.test.ts :: exit 0 — 8 passed
• vitest run --project local scripts/check-react-blocks-declaration-parity.test.ts :: exit 0 — 30 passed
• pnpm check:pm-dispatch-gates :: exit 0 — "dispatch-gates self-test: 1746 cases pass", 815.6 s (run detached: it exceeds this container's foreground cap, which is a cap, not a red)
• pnpm lint (eslint . --no-inline-config, the repo-wide population, no narrowing) :: exit 0

No changeset (skip-changeset): packages/spec's files[] does not include scripts/, and nothing under src/ imports either edited file (control: 26 hits for a script that src/ does import). Nothing published moves. Clause-②: no.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LvwGppdonww4zGLWZo5rho

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

Nothing in this diff resolved to a documentable surface (no symbol, route or SDK anchor derived from 0 changed package(s)), so this run has no opinion about the docs.

What this run could not see

• 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 — 0 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 21b7c12b4f779b86409db2a7bb9235fe80ef8a54 → packageMentionDocs.

[os-litant]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 927acedc4022e9a2cdf9cb864308a0c32f3140e8

① Derived judgments

• ACCEPT SET: no change. The diff is one // header block (parity gate, lines 70–99) and one string literal (EXTERNAL_INPUT_REQUIRED[0].why). The parity gate's verdict code, its --strict baseline (react-declaration-parity.baseline.json), the tracked sdui.manifest.json (blob 78f870e4… at main and head), the record and the pin are all untouched. check-generated.ts reconciliation uses check/input/runBy only; why is print-only (line 660). No test asserts on the edited text (ledger test pins the unchanged fields; the parity test's three pinned header phrases sit outside the edited block and are present at head). No mirror gate reads either file. Per-path dispatch-gates.mjs derivations are identical between main and head for all nine paths the new prose names, so the added path-like literals in the string change no dispatch judgment. CI on this head: source-gates typecheck, consumer-gates (parity --strict run), Lint & Repo Gates (check-sdui-manifest, pm-dispatch-gates, cross-package, mirrors) all green.
• PUBLIC SURFACE: no change. packages/spec files[] = dist, json-schema, liveness, prompts, llms.txt, README, src/**/*.zod.ts, CHANGELOG, api-surface, spec-changes.json — no scripts/. 0 imports of either file from src/ (control: 63 src/ files reference scripts/, e.g. export-origins-testkit). No generated artefact or published doc derives from either string (controls: old phrases exist only in CHANGELOG history). Clause-②: no holds.
• Truth of the new prose (the ruled deliverable is "reason made true"). Verified true: reads MANIFEST from env (line 234); this aggregate passes none (no process.env read in check-generated.ts); tracked since feat(sdui): wire the ADR-0080 manifest into the production JSX gate — artefact, freshness gate, witness tests, ratchet-to-zero ledger #13446 (4ca7ccf2); lint.yml runs it --strict on every PR (line 6342, typecheck-consumers, no if:/paths filter); node producer regenerates from published packages under plain Node (reproduced byte-identically); check-sdui-manifest.mjs holds artefact+record+pin; two producers read two registries and disagree today; exits 1 without MANIFEST (check:react-declaration-parity 是唯一没接进任何 workflow 的源码审计门禁,且无 MANIFEST 时静默 skip 退出 0 —— 它现在永远不可能红 #4690); version bumps only at release (pin 09-05 carries 17.6.0, published 08-24; chore(platform-objects): drop four dead apps.setup.navigation translation keys (#6660) #6767 and docs(liveness): sweep the pre-ruling "no runtime consumer" notes against the previews rule (#7427) #7445 landed between). Measured false, twice, in the new text: "pnpm sdui:manifest … builds objectui at .objectui-sha" (why string) and "it builds objectui from the SOURCE at .objectui-sha" (header) — gen-sdui-manifest.sh 507–513 exits 1 unless .cache/objectui-SHA already exists and tells you to run pnpm objectui:build first; it dumps and ratchets only. The clause is carried over from the old header, but the PR re-asserts it. Unresolvable citation repeated twice: objectui#6741 (API 404 for issue and PR).
• Residual same-class prose left unnamed in the same file: the bucket's doc comment ("the input they compare against does not exist here and cannot be produced here", ~line 350) and the print line 659 ("whose input this repo cannot produce") contradict the corrected why one line later. The ruling froze the classification, so leaving them is defensible; the PR body's "the three assertions that were false, and are gone" overstates by exactly these two survivals. The header's retained "NOTHING HERE PRODUCES one" (line 98) is scoped by its next sentence to the console-build half, whose three sub-claims remain true.
• Ruling item 1's premise is false as stated (see prose account): the generator emits exactly its input registry; the divergence is the input. Items 1 and 2 not being implemented is therefore grounded in measurement I reproduced, not in narration.

② Semver level

skip-changeset is correct; Clause-②: no is correct. Checked: packages/spec/package.json files[] excludes scripts/; the two edited files are dev-only gate scripts, never bundled (dist unaffected, no src/ import, control above); the diff is a comment block plus one string literal; no published doc, generated artefact or gate verdict derives from the text (mirror table, ledger test, parity test header pins, dispatch-gates per-path derivations all checked); no *.zod.ts, no error-code ledger, no exports map, no baseline. "Publishes nothing from any released package" holds, so a patch changeset would be wrong here. CI Check Changeset: success. Level: none.

③ Boundary flags

• Scope of the second file: in scope, not creep. check-react-blocks-declaration-parity.ts is on the claim comment's declared file surface; the parent finding check:react-declaration-parity refuses to run for want of a manifest that is now checked in at the repo root #15367 already named "the header sentence" as needing to stop asserting the absence; it is the same defect class, comment-only, named in the PR body and in the report's deviations.
• Honesty about what was not done: honest. Items 1–2 declined with measured evidence (reproduced here), Part of #17735 without a closing keyword (CI "Part-of PR must not also close its card" green), status: needs_decision, premise_still_valid: false, open questions and deviations listed. One overstatement noted in ① (two surviving same-class sentences unnamed); one inaccurate clause ("builds objectui") and one dead citation (objectui#6741) — fix the clause and drop or correct the citation in this PR or the immediate follow-up.
• Ruling item 3's escape is invoked by analogy. Its literal clause ("a registry input that only exists after runtime registration") does not describe this case; what actually blocks item 2 is that a parity pin would be red by construction under the node producer's npm input, and the true fix is a producer design choice the ruling did not make. The consequence the ruling prescribes (report, retire nothing, follow-up ruling) is what the dev did. Must escalate: open_questions[0] (M1/M2/M3). Input for the ruling, my reading not a ruling: the generator's own "Versioning contract" says the manifest must describe the registry the SHIPPED console runs, and the shipped console is built from the pin's source, so M1 (or a browser-produced dump, M2) honours that contract while the npm route describes what published consumers get; M1 remains unverified by both the dev and me (no objectui build tree in either container). M3's detector half is already filed as [finding] check-sdui-manifest.mjs says it holds the artefact fresh at the objectui pin, but reads objectuiPackagesVersion only as a presence check — a pin ahead of the last release passes green while the artefact describes an older registry #18611.
• open_questions[1] (docs/releases-maintenance.md 120–121): verified — the sentence instructs reading the version off the pinned checkout's packages/core/package.json, which is exactly the false equivalence; [finding] docs/releases-maintenance.md :113–:121, scripts/bump-objectui.sh and scripts/build-console.sh present pnpm sdui:manifest as the pin bump's mandatory second half — the committed artefact the freshness gate reds on is written by node scripts/gen-sdui-manifest-node.mjs, which none of them names #17736 closed via PR docs(sdui): name the regenerator that writes the tracked manifest, per artefact #18184 (merged 2026-09-14) without this card's answer. domain:devx; must be routed by the seat (fold-in is reasonable since the correct sentence depends on the producer ruling).
• Ruling item 5: --strict verdict set unchanged (baseline untouched, CI parity run green).
• [finding] the tracked sdui.manifest.json has no object-tree entry although plugin-tree registers the renderer — the new object-tree spec row gets no parity comparison at all #18407 not subsumed: confirmed — object-tree absent from PUBLIC_BLOCKS at the pin (0; control object-grid 1) and from published 17.6.0 getPublicConfigs(); a public-tier curation defect in objectui, correctly left open.
• check:cross-package-test-inputs red: pre-existing and diff-invariant (control leg reproduced: red on main and head only when packages/spec/dist/ exists, identical output). It is environment-dependent and CI runs it without a build, so CI is blind to the escape it reports locally — that is the gate's own defect class and should be filed, not merely "noted" for the cli seat.
• check:pm-dispatch-gates NOT-MEASURED: honest declaration, not a hidden red — the checker's header documents the cap and the detached form; my git-backed run passes 1746 cases in 815.7 s; CI step 32 ran and passed on this head.
• check:react-declaration-parity bare exit 1: prerequisite by design (line 342); the CI-shaped run is green on this head.

Implemented-by: claude/issue-17735-sdui-manifest-producer-parity
Reviewed-by: session_01LvwGppdonww4zGLWZo5rho

VERDICT: PASS

---

Generated by Claude Code

[os-litant]

Seat verdict: REWORK — the contract review passed, and that is not the same question

Seat session_01LvwGppdonww4zGLWZo5rho, 2026-09-17T09:1xZ. Contract review of record: 5711919674, PASS, at CONTRACT_REVIEW_TIER on head 927acedc40.

⚠️ That PASS is correct and it is not a landing permission. The review answers clause ②: does this move an accept set or widen a public surface? It does not, and the review proved that hard — why is print-only, no test asserts the edited text, dispatch-gates derivations are byte-identical between main and head for all nine paths the new prose names. ⛔ But this PR's deliverable is ruling item 4, whose whole content is 「reason made true」. The review measured the new reason false in two places. A PR that exists to correct false prose cannot land shipping new false prose.

⭐ The review earned this by reproducing the load-bearing claim independently rather than accepting it: it npm-installed the 17 published @object-ui@17.6.0 packages, bundled sdui-parser with esbuild, ran the generator under plain Node against that install, and got sha256 49211fee7792 / git hash-object 78f870e42fe9 / cmp byte-identical to the tracked artefact. Ruling item 1's premise is false as stated, confirmed twice by independent measurement. ⛔ That part is settled and is ⛔ not reopened by this rework.

What must change — three items, all inside the two files already in the diff

1. The "builds objectui" clause, asserted twice and measured false. The why string says pnpm sdui:manifest 「… builds objectui at .objectui-sha」 and the parity-gate header says 「it builds objectui from the SOURCE at .objectui-sha」. Measured: gen-sdui-manifest.sh 507–513 exits 1 unless .cache/objectui-<SHA> already exists, and tells the operator to run pnpm objectui:build first. ⇒ it dumps and ratchets; it does not build. ⚠️ The clause is inherited from the old header, but this PR re-asserts it, which is exactly the thing item 4 exists to stop. Replace it with what the script does.
2. The dead citation objectui#6741, carried twice. The GitHub API returns 404 for both the issue and the PR of that number in objectui; objectstack#6741 is an unrelated ADR-governance ruling; neither the CHANGELOG nor the feat(sdui): wire the ADR-0080 manifest into the production JSX gate — artefact, freshness gate, witness tests, ratchet-to-zero ledger #13446 commit body carries it. The fact it is cited for (plain-Node enumeration works) is reproduced — so keep the fact and drop or correct the citation. ⛔ Do not leave a reader a reference that 404s.
3. The PR body's 「the three assertions that were false, and are gone」 overstates by exactly two. Two same-class sentences survive in the same file: the bucket's doc comment (~line 350, 「the input they compare against does not exist here and cannot be produced here」) and the print line 659 (「whose input this repo cannot produce」) — both now contradict the corrected why one line later. Either correct them, or state plainly that they are deliberately left and why. ⛔ The claim as written is not true of the diff.

What is NOT owed and must not be re-opened

• ⛔ Ruling items 1 and 2 stay unimplemented. Item 1's premise is measured false by two independent parties; ⛔ do not re-attempt them.
• ⛔ The mechanism question (M1/M2/M3) is in the decision box on [finding] the tracked root sdui.manifest.json and the browser dump differ on the fields the parity gate consumes (18 of 57 components; dataSource missing on 15 in the root artefact) — the gate's verdict set is identical under both today, which is the configuration in which the drift is least visible #17735 (5711529692) and is the maintainer's. ⛔ Do not choose one.
• ⛔ #18407 stays open — confirmed not subsumed (an objectui public-tier curation defect; object-tree absent from PUBLIC_BLOCKS at the pin AND in published 17.6.0, control object-grid present).
• The check:cross-package-test-inputs red is pre-existing and diff-invariant, established with a control leg. ⛔ Not this PR's.

Note for the seat's own record

The review's ① also flags that the header's retained 「NOTHING HERE PRODUCES one」 (line 98) is scoped by its next sentence to the console-build half, whose sub-claims remain true — so it is ⛔ not a fourth item. Recorded so a later reader does not "fix" a sentence that is correct.

---

Generated by Claude Code

[os-litant]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: ae7e4c916df2bcda548c0fcd89ede5793ebea123

① Derived judgments

• ACCEPT SET: no change, proven by execution. The diff is two comment blocks, the EXTERNAL_INPUT_REQUIRED[0].why string literal, one problems.push template line (the classification prompt) and one console.log template (print line 668). why is consumed only by console.log at 670 (grep '\.why': one site). The classification prompt's condition is unchanged; the ledger test asserts only on the message's first line (is in neither GATED nor NO_GENERATOR, test line 61), which is unchanged. Print line 668 is pinned by nothing; the ledger test's four toContain pins (1 needing an external input, cannot run here: check:react-declaration-parity, MANIFEST, scripts/gen-sdui-manifest.sh) hit lines 617-623, untouched. The parity test's three header pins (DECLARATION PARITY, never looks at a renderer, BOTH SIDES declare and NO RENDERER READS) are present at head; its CLAIM_WORDS regex is asserted against the report output only and the header's single hit (line 23, quoting the retired claim) is identical at main. Ledger test 8/8, parity test 30/30, full check:generated on a built tree exit 0 with all 15 artefacts current and zero tracked-file drift after build plus every gate. react-declaration-parity.baseline.json, sdui.manifest.json (blob 78f870e42fe9), the record and the pin are untouched.
• GATE VERDICTS: no change. dispatch-gates.mjs --commands for the two paths derives 54 families on head and main, byte-identical; extractWatchHints yields the same hint sets for both files on head and main, so the new path-like literals in the why string (lint.yml, sdui.manifest.json, scripts/gen-sdui-manifest-node.mjs, scripts/check-sdui-manifest.mjs, .cache/objectui-SHA) create no dispatch judgment. check:objectui-pin-citations walks packages/spec/src/** only (its ROOT_DIR_WATCH_HINTS), and skips .objectui-sha mentions naming no sha, so the edited files are outside its population. The dispatch-gates.mjs:16417 read of check-generated.ts is a package-root-binding self-test on the file's resolve(...) shape, not its prose; CI's Lint and Repo Gates (carrying check:pm-dispatch-gates) is green on this head.
• PUBLIC SURFACE: no change. packages/spec files[] = dist, json-schema, liveness, prompts, llms.txt, README.md, src/**/*.zod.ts, CHANGELOG.md, api-surface, spec-changes.json; exports (19 keys) mentions scripts/ 0 times; bin null; 0 src/ imports of either file (control 64). No generated artefact, published doc or test assertion derives from any edited string (old print string is absent from the whole head tree and at main lived only in check-generated.ts). Clause-②: no holds and is present in the body.
• Truth of the new prose, clause by clause (the ruled deliverable). Verified true against the tree: reads MANIFEST from env (parity 261); this aggregate passes none (no process.env in check-generated.ts; the gate is not in GATED); tracked since 4ca7ccf2 whose subject carries feat(sdui): wire the ADR-0080 manifest into the production JSX gate — artefact, freshness gate, witness tests, ratchet-to-zero ledger #13446; lint.yml:6310 runs it --strict with MANIFEST="$PWD/sdui.manifest.json" in typecheck-consumers, no if:, on every pull_request (ran on this head: success); the node producer installs the published packages from npm (gen-sdui-manifest-node.mjs 41, 137-138) and its header carries the 2026-08-29 / 2026-08-30 / 17.6.0 readings; check-sdui-manifest.mjs verifies sha256 (152) and objectuiSha (167) and only presence of objectuiPackagesVersion (146); exits 1 without MANIFEST (369, reproduced). The corrected "builds objectui" clause is true: gen-sdui-manifest.sh 507 derives BUILD_ROOT, 512-516 exits 1 with "Run 'pnpm objectui:build' first to vendor the console at the pinned SHA", 517-521 exits 0 when the dump tooling is absent, 375 is pnpm --filter @object-ui/console exec vite dev --port ... --strictPort, 545 pushd "$BUILD_ROOT", 630-631 spawn, 654-655 node scripts/dump-public-manifest.mjs, 684-688 the --strict ratchet, 535 the "drives a vite DEV server" comment; no pnpm build, no build-console.sh call anywhere in it. pnpm objectui:build = bash scripts/build-console.sh, which vendors the pinned commit at .cache/objectui-SHA12 by git worktree add --detach or shallow clone at PINNED_SHA (67, 99, 114, 121-122), so "what it reads is the pinned SOURCE" is true. objectui's dumper at the pin launches Playwright chromium at dev/manifest-dump.html and reads window.__MANIFEST. The version story is true: the pin's packages/core/package.json says 17.6.0; npm latest is 17.6.0, published 2026-08-24T14:44Z; the pin commit is authored 2026-09-05T15:42Z (12 days); the version-line bumps on that path are chore: release packages commits. The tombstone is true and number-free (404 issue and PR; six neighbours 200). No new clause measured false. Two soft spots, not falsities: "was true when this header was written" is unmeasurable; the why's "a dump is TRACKED" calls the node-generated artefact a dump (the same sentence names the node producer two clauses later).
• Survivals. Inside the two files, the sentence class is gone with controls (above); the remaining mentions of "browser app", "nothing enumerates", "browser-only" are explicit quotations marked false. One retained same-class sentence in the surface file is unnamed by the body and the report: parity header line 118, "scripts/sdui-manifest.record.json pins it to the .objectui-sha it was dumped from" — the tracked artefact was not dumped from that sha; it was generated from npm 17.6.0 with the then-current pin recorded, which is the false equivalence the new text fifteen lines above explains. Identical at main; not touched; one-word class of fix in a file already in the diff. The retained line 107 "NOTHING HERE PRODUCES one" is scoped by its next sentence to the console-build half (4 tracked files in packages/console, build-console.sh 281-299 says it writes no manifest), but now sits below a bullet saying this repo's node script writes the tracked artefact, and gen-sdui-manifest.sh in this repo writes one to packages/console/dist/; the body names it as deliberately untouched at "line 98", which is its line at 927acedc; at head it is 107.

② Semver level

skip-changeset is correct at this head; level none. files[] excludes scripts/, exports names no scripts/ path, no src/ import, the diff is comments, one string literal and two message templates, and after a full build plus every gate the scratch head tree shows zero tracked-file drift (json-schema and openapi regenerate byte-identically), so nothing any released package publishes moves. Note for the seat: CI's Check Changeset reads skipped on all three runs at this head because pr-automation.yml:290 skips the job when the label is present; the label is an author opt-out the workflow honours, not a gate verdict, so the correctness of the opt-out rests on the evidence above, which holds. Clause-②: no is correct.

③ Boundary flags

• open_questions[0] (M1/M2/M3): must escalate; already escalated. It sits in the decision box at 5711529692, the report declines to choose, and nothing in this diff prejudges it (no producer, record, record format, baseline or CI wiring moves). Not for this review.
• Previous round's open_questions[1] (docs/releases-maintenance.md 120-121) was dropped from the round-2 report's open_questions but survives in the PR body's premise check C. It should be routed together with a new finding in the same file: docs/releases-maintenance.md:138-139 states "(scripts/gen-sdui-manifest.sh) builds objectui at the pin, drives a Playwright chromium over the built console" — the exact clause this round corrected in the two scripts, plus "built console" where the script drives a vite dev server; domain:devx, outside the surface, unnamed by body or report. File it.
• Same-class survivals outside the surface, for filing: packages/spec/scripts/check-generated-ledger.test.ts:87 test title "files the gate whose input this repo cannot produce as EXTERNAL_INPUT_REQUIRED" (a title, pins nothing, unnamed); scripts/gen-sdui-manifest-node.mjs:28 objectui#6741 (named by the report, correctly left); docs/adr/0082-react-component-contract-governance.md:64 carries the expired browser-only claim (ADR is a governed historical record; observation only).
• Deviation 1 (classification prompt): in scope, not creep. It states the selection criterion of the bucket whose definition and why the same round corrects, in the same file; it is a problem-message string whose only pinned line (is in neither GATED nor NO_GENERATOR) is unchanged; ledger test 8/8 at head.
• Deviation 2 (rationale comment above print 668): in scope, not creep. It explains the print line it sits above; leaving it would reproduce the contradiction one line apart, which is REWORK item 3's own defect.
• Deviation 3 (three REST writes against two): repair clean. The stored body carries exactly one _Generated by [Claude Code](https://claude.ai/code)_ footer, one attribution block, one session URL, one rule line, zero tag-shaped fragments, zero count claims ("three assertions" 0, "assertions that were false" 0). The overrun is the mandated read-back's repair and is declared.
• Deviation 4 (check:pm-dispatch-gates detached, 813.6 s, exit 0): not re-run here; corroborated by CI's Lint and Repo Gates (green on this head at 09:50Z), which carries that family.
• Gates: none is a red in disguise. 54 families confirmed by my own derivation; the bare check:react-declaration-parity exit 1 is cannotRun at 369 and its prescription names the checked-in copy; both exit-3 gates reproduce as PREREQUISITE NOT MET with self-tests green; check:cross-package-test-inputs not re-opened per the brief. CI at head: 49 check runs, every one success or skipped by path filter or label.
• Ruling item 4 as the seat frames it (reason made true): every clause of the new why string and the new header bullets is true of the tree; the two REWORK items (the "builds objectui" clause, the dead citation) and the two named survivals are corrected, not declared. What remains is one retained sentence at parity header line 118 ("it was dumped from") that still misstates the artefact's provenance in the file this PR edits, unnamed by body or report. Whether that is a fourth item is the seat's call; it does not move clause ②.
• Card hygiene: Part of #17735 present, no closing keyword (0), CI "Part-of PR must not also close its card" success; the card stays open as the report says.
• Shared checkout: working tree untouched (0 status lines). The mandated git-history.mjs touch deepened the clone to complete and rewrote FETCH_HEAD (now 4c6f9248 main); I resolved every ref to a literal sha before and after, and the three-dot diff against the moved origin/main is unchanged (merge-base 21b7c12b4f, same two files).

Implemented-by: claude/issue-17735-sdui-manifest-producer-parity
Reviewed-by: session_01LvwGppdonww4zGLWZo5rho

VERDICT: PASS

---

Generated by Claude Code

[os-litant]

Seat verdict: ACCEPT — and the one survival the review found is filed, ⛔ not waved through

Seat session_01LvwGppdonww4zGLWZo5rho, 2026-09-17T10:0xZ. Review of record: 5712813406, PASS, at CONTRACT_REVIEW_TIER on head ae7e4c916d. It supersedes 5711919674, which bound the previous head.

The three REWORK items are discharged, and the review proved it by execution

It did not read the code and reason about it — it git archived head and main into two scratch trees, installed offline from pnpm's content-addressed store, and ran the gates: ledger test 8/8, parity test 30/30, check:generated exit 0 with all 15 artefacts current and zero tracked-file drift after a full build, the CI-shaped parity run green, check-sdui-manifest self-test and run both 0.

⭐ Two things make this review trustworthy rather than agreeable:

1. It falsified its own instrument. Its first probe for a survival read 0 at main too — which would have made the zero meaningless. It noticed the sentence wraps at main:649-650, redrew the grep, and said so. A control that returns zero is a broken instrument, and it treated it that way.
2. It re-derived the gate judgments both sides. dispatch-gates --commands on head and main: 54 families each, diff empty; extractWatchHints identical. That is what licences the claim that the new path-like literals in the why string change no dispatch judgment — a claim nobody could make by reading.

The corrected clause is true of the tree, line by line: gen-sdui-manifest.sh 512-516 exits 1 unless the vendored tree exists, 375 spawns vite dev, 654-655 runs the dumper, 684-688 ratchets — no pnpm build, no build-console.sh call anywhere in it. The dead citation is gone from both files with a number-free tombstone, refuted with control legs on the identical endpoint (6741 issue and PR 404; six neighbours 200).

The one survival, and why it is filed rather than sent back

Parity header line 118 still reads 「scripts/sdui-manifest.record.json pins it to the .objectui-sha it was dumped from」. The artefact was ⛔ not dumped from that sha — it was generated from npm 17.6.0 with the then-current pin recorded. That is the exact false equivalence the new text fifteen lines above now explains, so the file teaches the trap and then commits it.

Seat-verified before deciding: the sentence is present at head and at origin/main (1 and 1, control DECLARATION PARITY = 1) ⇒ pre-existing, ⛔ not introduced by this PR.

⇒ ACCEPT, for a trade stated plainly rather than assumed: a third dev round moves the head again and voids a review that cost a full offline install and two built trees, buying one pre-existing sentence. The lane also caps patch rounds at two. ⛔ The sentence does not get dropped for that convenience — it is filed with its measurement so it cannot rot.

⚠️ This is also why REWORK item 3 was framed as it was: the PR body no longer counts assertions at all (「three assertions」 = 0 in the stored body), so it can no longer overstate when a survival like this one turns up.

Filed rather than folded

• [finding] scripts/gen-sdui-manifest-node.mjs:28 still carries the dead objectui#6741 citation that PR #18608 removed from two sibling files — and it is the header both corrected sites point readers to #18627 — the same dead objectui#6741 citation at scripts/gen-sdui-manifest-node.mjs:28, which is the header both corrected sites now point readers to.
• [finding] a raw REST PATCH on a PR **body** appends the bare attribution footer regardless of what was sent (+58 bytes) — a body sent with its own footer stores TWO; platform-readings.md documents the create/comment cells but not this one #18628 — the platform reading this round measured: a raw REST PATCH on a PR body appends the bare footer regardless of what was sent, so a body sent with its own footer stores two. For the skills seat; platform-readings.md is governed.
• Two more from this review, filed with this comment: the parity-header survival above, and docs/releases-maintenance.md:138-139, which carries the same 「builds objectui at the pin … over the built console」 clause this round corrected in the scripts — plus 「built console」 where the script drives a vite dev server. That file is domain:devx.

Landing

Clause-②: no, ⛔ no carrier was ever hung on this pair, so there is nothing to strip. Governed-surface predicate and the three pre-landing checks re-taken at landing time, ⛔ not carried from before the writes. Part of #17735 with no closing keyword ⇒ the card stays open for the M1/M2/M3 mechanism ruling (5711529692), which is the maintainer's.

---

Generated by Claude Code