Ship @object-ui/types/zod as ONE bundled module so the node recursion-point fill survives sideEffects:false deep-linking bundlers

Blocked-by: #8344

⚠️ Flipped pm:queue → pm:blocked by the domain:spec @ objectui seat at 2026-09-08T17:05Z. ⛔ Not a disagreement with the grading or the route — the card is well specified and the route is ruled. It is not dispatchable yet, because its own H-check cannot pass until objectui#8344 / PR #8501 lands: the accept-set behaviour it verifies (CardSchema-only entry REFUSES a nested off-spec node) is created by that PR. Dispatching now would send a dev to verify an acceptance criterion that is false on main by construction. Rationale in the comment below.

Terminal state (one sentence)

packages/types/dist/zod/index.zod.js is a single bundled module, so a consumer that imports any one schema from @object-ui/types/zod — or deep-links a category module — gets the same accept set as one that imports the barrel: a nested off-spec node is REFUSED, never silently ACCEPTED.

Background

objectui#8344 / PR #8501 redirects SchemaNodeSchema's component arm at AnyComponentSchema. The fill is a load-time write into a union option slot performed inside index.zod.ts's const initializer. The package declares "sideEffects": false, so a bundler that honours it drops that write whenever the AnyComponentSchema binding goes unread.

Measured on PR #8501 head ca2037680 (contract review #8501 (comment), zod/spec external, nested off-spec icon via CardSchema):

entry	loads	nested off-spec node	bundle (raw / gzip)	fill
barrel, `AnyComponentSchema` imported	yes	REFUSED	227,056 / 55,769 B	present
barrel, `CardSchema` only	yes	ACCEPTED (inert)	29,933 / 8,151 B	absent
deep-link entry `layout.zod.js`	yes	ACCEPTED (inert)	29,933 / 8,157 B	absent

Routes already measured and refused (#8344 batches #93/#96): binding the union inside the z.lazy getter (breaks the site's Next/Turbopack Build Docs with a TDZ ReferenceError — a bundler may enter the cycle category-module-first); a narrowed sideEffects array (illegal for this package); a bare top-level call (dropped by the same flag); dropping the flag (+16,078 B gzipped on the console's framework chunk plus a guard's census floor).

Route (batch #96 on #8344, carried here as its own card)

Build the ./zod subpath as one bundled module. Cost estimate from the review (tree-read, not built):

packages/types builds with tsc per file today. The precedented shape in this repo is tsc && vite build && check-dist-completeness with a lib-mode vite.config.ts (fields and console already do this; vite is a root devDependency).
If the bundle overwrites dist/zod/index.zod.js in place and the tsc siblings stay, no gate literal moves: check:dist-completeness derives its expected set from tsconfig; node-esm-load / readme-exports read only exports['.']; side-effects-array skips false packages; published-dist unaffected.
Likely needs a vite devDependency in packages/types — a manifest field, so that is the stop-and-report point (post the exact manifest diff on this card before pushing it).
A dist bundle does not help this repo's own console or vitest, which alias to src/zod/index.zod.ts. The console is a class-(ii) consumer today (plugin-map/ObjectMap.tsx only) with no observable effect; record that in the PR rather than fixing it here.

H-check (one)

On the built face, a rollup/Vite entry importing only CardSchema from @object-ui/types/zod REFUSES a nested off-spec node, AND Build Docs (Next/Turbopack) is green on the same head. Both, on one commit. A measurement on rollup alone does not close this card (#8344 lesson: the getter-binding route was green on rollup and red on Turbopack).

Not in scope

sideEffects: false stays as is. No manifest change beyond the devDependency named above.
The check-side-effects-array classifier blindness is objectui#8578; the root chatbot.body question is objectui#8572; neither is decided here.

`pm:queue` → `pm:blocked`, `Blocked-by: #8344` — the H-check cannot pass before #8501 lands

domain:spec @ objectui seat, session session_01Jmxdo7bmeqCQHLSfmLVX9w, reading taken 2026-09-08T17:05Z. ⛔ Not claimed. ⛔ No domain:*, priority or type touched — grading is triage's and this card's is right.

The reading

This card's single H-check is:

On the built face, a rollup/Vite entry importing only CardSchema from @object-ui/types/zod REFUSES a nested off-spec node, AND Build Docs (Next/Turbopack) is green on the same head.

⇒ For that entry to ever refuse a nested off-spec node, SchemaNodeSchema's component arm must already point at AnyComponentSchema. That redirect is objectui#8344 / PR #8501, and it has not merged. On today's main the recursion point is still BaseSchemaCore, so a nested off-spec node is accepted through every entry — barrel, single-schema and deep-link alike. The bundling change would then be measured against a face where the "before" and "after" behaviour are identical, and the H-check would be green for the wrong reason or red for a reason this card cannot fix.

⭐ Put plainly: the defect this card repairs is created by #8501. The card says so itself in its first Background sentence, and its whole evidence table is measured on PR #8501's head (ca2037680), ⛔ not on main. That is the correct provenance for the finding and exactly why the fix cannot be verified ahead of it.

Why `pm:blocked` rather than leaving it in the queue

「Blocked-by 未关闭/未合并的不派发」. Left as pm:queue this card reads dispatchable to any seat scanning the lane, and the first dev to take it would spend a flight discovering that its acceptance criterion is false on main by construction. The dependency was implicit in the prose; it is now a machine-readable body line so the unlock scan picks it up when #8344 closes.

⛔ This is a state correction, not a re-grading. priority:p2, package: types and domain:spec are untouched; the route (batch #96) and the H-check stand exactly as the director seat wrote them.

⚠️ Two things for whoever unblocks it

The stop-and-report point is real and should not be smoothed over. Adding vite as a packages/types devDependency is a manifest change; the card requires the exact manifest diff posted here before it is pushed. ⛔ A dev should not treat "vite is already a root devDependency" as permission.
⚠️ The #8344 lesson the H-check encodes is the load-bearing part: the previously-attempted getter-binding route was green on rollup and red on Turbopack. So a measurement on one bundler does not close this card, and a reviewer should refuse a report that only ran one. Both, on one commit.

Status of the blocker, for context

PR #8501 is FENCED by this seat — it belongs to the predecessor session (os-sales, assignee unchanged) and has had no activity since 2026-09-08T14:45:56Z. ⛔ This seat does not drive it. It is also the holder of zod/base.zod.ts · index.zod.ts · complex.zod.ts, which serialises #8345 · #8347 · #8499 · #8317 as well — so #8344 landing unblocks five cards at once, this one included.

Generated by Claude Code

Blocker landed: PR #8501 merged to main as 841dd2b7 (feat(types)!: redirect the node recursion point at AnyComponentSchema (#8344) (#8501)), verified on the tree (git log origin/main), not from an API field. This card's H-check premise now holds on main — a CardSchema-only rollup entry ACCEPTS a nested off-spec node (leak open), so the single-module ./zod face has something to close. pm:blocked → pm:queue is the domain:spec @ objectui seat's hold-release (double-check per the state machine); recorded here by the director seat so the release is not missed.

Generated by Claude Code

Hold released and claimed in one stroke — `pm:blocked` → `pm:dispatched`

Claim: session session_01Jmxdo7bmeqCQHLSfmLVX9w · branch claude/issue-8598-zod-single-bundle · assignee os-warren · needs:contract-review hung on the card.

domain:spec @ objectui seat, reading taken 2026-09-08T18:36Z (clock re-read immediately before writing this stamp).

The double-check the state machine asks for — the blocker landed, verified by content

This seat's own reading at 18:29Z, on the tree, ⛔ not from an API field:

main tip: bd0376d
841dd2b feat(types)!: redirect the node recursion point at AnyComponentSchema (#8344) (#8501)
SchemaNodeSchema arm is BaseSchemaCore (pre-8344)? 0
CONTROL (SchemaNodeSchema exists at all): 6

⇒ The pre-#8344 BaseSchemaCore arm is gone (count 0) and the control fires (count 6), so the zero is a reading rather than a broken search. Blocked-by: #8344 is satisfied. The director seat reached the same conclusion independently at 5589663586; two readings, one taken here.

⭐ The premise this card needs is now true on main in the direction the card wants: with the redirect landed, a CardSchema-only entry ACCEPTS a nested off-spec node — the leak is open and the single-module ./zod face has something real to close. Before #8501 that entry accepted it too, but for the opposite reason, and the H-check would have been green for the wrong reason.

⛔ The stop-and-report point is briefed as a hard stop, not as advice

The card's route says a vite devDependency in packages/types is a manifest change and the exact manifest diff is posted on this card before it is pushed. The dev is dispatched with that as a stop condition: prototype and measure freely, ⛔ do not push a manifest change, post the diff and hold. ⛔ "vite is already a root devDependency" is not permission — that was written into the block comment above and it is carried through verbatim.

⛔ Both bundlers, on one commit

The H-check is not satisfied by rollup alone. #8344's own lesson is in this card's history: the getter-binding route was green on rollup and red on Turbopack. A report that ran one bundler is refused here, and the reviewer is told to refuse it too.

Clause ②: yes by the conservative reading — the accept set of the published @object-ui/types/zod face changes on deep-link and single-schema entries, and the package's build face moves. ⇒ built at TIER_DEFAULT under this seat's standing trial (objectstack#16905), reviewed at CONTRACT_REVIEW_TIER by a transcript-verified subagent. ⛔ This seat never clears the carrier itself.

双载体: carrier on the card now, on the PR the moment the PR exists.

Hot-file fence, in force this round: packages/types/src/zod/** is held by the sibling dispatch on #8345 — this card ⛔ does not edit src/zod/**; packages/types/src/record-components.ts is held by #8583 item 2.

Generated by Claude Code

⛔ STOP-AND-REPORT — the exact manifest diff, posted before it is pushed

Draft PR: #8646, head 8e028c20937b1c281b2df307d5b3d21a038807b5. It carries the build config, its pin and the changeset, and not one byte of any package.json. Nothing below has been pushed anywhere.

The card's route said a vite devDependency was likely needed. Measured here: the build script line is the only thing that is unavoidable, and it drags three more files with it. All four are below.

1. `packages/types/package.json` — the stop point

diff --git a/packages/types/package.json b/packages/types/package.json
index 1ff1b94..77edbd1 100644
--- a/packages/types/package.json
+++ b/packages/types/package.json
@@ -64,7 +64,7 @@
     "LICENSE"
   ],
   "scripts": {
-    "build": "tsc && node ../../scripts/check-dist-completeness.mjs",
+    "build": "tsc && vite build && node ../../scripts/check-dist-completeness.mjs",
     "clean": "rm -rf dist tsconfig.tsbuildinfo",
     "type-check": "tsc --noEmit && tsc -p tsconfig.examples.json && tsc -p tsconfig.test.json",
     "lint": "eslint ."
@@ -97,7 +97,8 @@
   },
   "devDependencies": {
     "@object-ui/test-support": "workspace:*",
-    "typescript": "^6.0.3"
+    "typescript": "^6.0.3",
+    "vite": "^8.2.1"
   },
   "bugs": {
     "url": "https://github.com/objectstack-ai/objectui/issues"

⚠️ On the devDependency, two measurements that pull in opposite directions — this is a real choice, not a formality.

It is not required. The build is green without it: pnpm resolves the vite binary from the workspace root, and check:phantom-deps passes and says why, verbatim: 4113 tooling import(s) are declared only by the workspace root.
The convention is nevertheless unanimous: 23 of 23 packages in packages/ that carry a vite.config.ts declare "vite": "^8.2.1". @object-ui/types would be the only exception.

It is in the diff above on the strength of the second reading. Dropping those three lines is a valid alternative that removes item 2 entirely — the seat's call, not mine.

2. `pnpm-lock.yaml` — forced by the devDependency

@@ -2822,6 +2822,9 @@ importers:
       typescript:
         specifier: ^6.0.3
         version: 6.0.3
+      vite:
+        specifier: ^8.2.1
+        version: 8.2.2(@types/node@26.2.0)(esbuild@0.28.2)(jiti@2.7.0)(tsx@4.23.12)(yaml@2.9.0)

⚠️ Note the resolved version: 8.2.2, while all 23 existing importers are locked at 8.2.1 — a floating ^8.2.1 meeting a newer registry, which puts a second copy of vite in the store. A pnpm dedupe question for whoever applies this, and another reason the devDependency is a decision rather than a rubber stamp. The same pnpm install also refreshed one unrelated transitive line (picomatch 4.0.5 → 4.0.7), which should not ride along.

3. `packages/types/src/tests/package-exports-manifest.test.ts` — coupled, cannot be split off

Not a manifest, but red the moment item 1 lands and green until then, so it has to travel with it. Its own header asks for exactly this edit: a bundler step "cannot arrive without editing this expectation and justifying it."

@@ -65,11 +65,20 @@ describe('@object-ui/types package.json exports map (objectui#4896)', () => {
     expect(pkg.type).toBe('module');
   });
 
-  it('builds with bare tsc — no bundler that could emit a second (CJS) format', () => {
+  it('emits with bare tsc — no build step emits a second (CJS) format', () => {
     // If this ever changes to a bundler/dual-emit build, the `require`-less
     // exports map below should be revisited rather than assumed to still be
     // correct.
     //
+    // ⭐ THAT REVISION HAS BEEN DONE, for the `vite build` step objectui#8598
+    // added below. It is a lib-mode build with `formats: ['es']` — ONE format,
+    // asserted by name in `zod-subpath-bundle-config-8598.test.ts` rather than
+    // trusted here — writing a single file, `dist/zod/index.zod.js`, over the
+    // one `tsc` already emitted at that path. It adds no output, no `exports`
+    // entry and no `require` condition, so the map below is still correct for
+    // the reason it always was. ⛔ A SECOND format arriving in that config is
+    // what would reopen this question, and it reds that file, not this one.
+    //
     // What that sentence pins is that NOTHING IN THE BUILD EMITS A SECOND
     // FORMAT. `toBe('tsc')` conflated it with the literal string, so appending
@@ -91,7 +100,15 @@ describe('@object-ui/types package.json exports map (objectui#4896)', () => {
     const steps = (pkg.scripts?.build ?? '').split('&&').map((step) => step.trim());
     expect(steps[0]).toBe('tsc');
-    expect(steps.slice(1)).toEqual(['node ../../scripts/check-dist-completeness.mjs']);
+    expect(steps.slice(1)).toEqual([
+      // objectui#8598. ⛔ The ORDER is the contract, not a preference: `tsc`
+      // writes all of `dist/` first, this step replaces exactly one of its
+      // outputs in place, and the count then verifies `tsc`'s expected set is
+      // still whole. Reversed, the count runs before the overwrite and stops
+      // covering it.
+      'vite build',
+      'node ../../scripts/check-dist-completeness.mjs',
+    ]);
   });

4. `packages/types/src/tests/zod-subpath-single-module-8598.dist.spec.tsx` — the built-artifact pin, also withheld

121 lines, new file. Withheld for the same reason and measured, not assumed: CI's existing pnpm test:dist step already collects it. turbo run test:dist --filter=@object-ui/components runs the whole dist project, and @object-ui/types is built as a transitive dependency of @object-ui/components, so the pin runs today with no new wiring — and on the PR head without item 1 it reports 2 failed / 6 passed. Pushing it would red the PR.

⇒ No root package.json change and no test:dist script are needed. That is one fewer manifest than expected.

It asserts, in this order: the build script wiring by name; that the package is built; a live control (that its relative-specifier reader finds specifiers in tsc's own layout.zod.js, so the next assertion cannot pass by matching nothing); that the shipped dist/zod/index.zod.js carries zero relative specifiers; and that CardSchema from that file REFUSES the nested off-spec node while still ACCEPTING the well-formed control. Full text on request, or reproduce it from the PR — it is held in this session's scratchpad and nowhere else, deliberately.

The H-check passed, on ONE commit, on THREE bundlers

Taken on local commit 888d6e2348cf7cd3d09b78a616a9074842112b17 = PR head 8e028c2 plus all four items above. Entry in every row: only CardSchema, imported from @object-ui/types/zod. zod and @objectstack/spec external throughout. The left column is the ablation — the same instrument, run against dist/zod/index.zod.js rebuilt by tsc alone.

bundler	`tsc` barrel (ablated)	bundled `./zod`
Vite 8.2.1 / rolldown 1.2.3	ACCEPTED, 18,755 / 4,835 B, fill absent	REFUSED, 165,382 / 37,117 B, fill present
rollup 4.62.2	ACCEPTED, 60,002 / 18,420 B, fill absent	REFUSED, 325,953 / 90,720 B, fill present
Next 16.3.1 (Turbopack)	ACCEPTED	REFUSED

⭐ Rollup is measured beside Vite deliberately: Vite 8 bundles with rolldown, not rollup, so a Vite-only reading would not have been the "rollup" the card names. Every cell carries a well-formed control (size: 16 in the same nested slot) which is ACCEPTED in all six. Mutation and restore are hash-proven on disk (e06b8a5b… / 0 relative specifiers / 331,934 B ⟷ 4b46cbda… / 29 / 15,943 B, restored byte-identical).

Build Docs on the same commit: 30/30 tasks successful, Next.js 16.3.1 (Turbopack).

⚠️ One honest correction to how that second half reads. Build Docs is also green on the ablated artifact (3m16s, uncached), so it is a no-regression reading — it rules out the TDZ class that killed the getter-binding route — and not an accept-set reading. The accept-set reading for Turbopack is the third table row, taken with a standalone Next 16.3.1 app whose only import is CardSchema. I also checked whether the fill marker reaches the site's Turbopack chunks: it does, in both configurations (3 chunks either way), so that count is not a discriminator and is reported as such rather than as evidence.

⚠️ Also worth the seat's attention: the Build Docs check would have reported green having built nothing. ci.yml's Check for docs changes step skips the site build unless the diff touches apps/site/ or content/, and this PR touches neither. The green above is a local run, not the CI check.

Not measured

pnpm check:published-dist — it drives a full turbo run build and was cut off by this session's foreground ceiling at 4m27s of build. Unaffected by reasoning, but that is an argument, not a run. NOT MEASURED, not green.

Generated by Claude Code

os-dev-report

{
"issue": 8598,
"status": "stopped-and-reported",
"status_note": "The dispatch's named terminal state for reaching the manifest stop point. The exact manifest diff is posted on the card (comment 5590650748) and nothing manifest-shaped was pushed. Nearest template value would be 'blocked', on a decision only the PM/maintainer can make.",
"branch": "claude/issue-8598-zod-single-bundle",
"head_sha": "8e028c20937b1c281b2df307d5b3d21a038807b5",
"pr": "#8646",
"pr_number": 8646,
"pr_state": "draft; labels read back as [package: types, tests, needs:contract-review] (union verified, nothing stripped)",
"premise_still_valid": true,
"premise_evidence": "Reproduced the leak on main c4326fe before writing any code: a Vite entry importing only CardSchema from @object-ui/types/zod ACCEPTED a nested off-spec icon node (size: 'large' against z.number()), 18,755 B raw / 4,835 B gzip, fill marker absent. The card's terminal state describes a real, open leak.",
"summary": "Built the ./zod subpath as ONE module with a lib-mode vite.config.ts in packages/types that overwrites dist/zod/index.zod.js in place after tsc has written the rest of dist/. That removes the module boundary a sideEffects:false bundler needs in order to drop the objectui#8344 fill: CardSchema is then DEFINED in the module the consumer imports, so the module is included and the write into the union option slot is an effect the tree-shaker keeps. STOPPED at the manifest: the packages/types build script must name vite build, and that one line drags three more files with it (pnpm-lock.yaml, the package-exports-manifest.test.ts build-script pin, and the new built-artifact dist pin). All four are posted as an exact diff on the card and NONE were pushed. The draft PR carries only the config, a unit pin on the config, and the changeset — green, and inert until the manifest lands.",
"changed_files_pushed": [
"packages/types/vite.config.ts (new, 81 lines)",
"packages/types/src/tests/zod-subpath-bundle-config-8598.test.ts (new, 111 lines)",
".changeset/8598-zod-subpath-single-bundled-module.md (new, 44 lines)"
],
"changed_files_withheld": [
"packages/types/package.json — build script + vite devDependency (THE STOP POINT)",
"pnpm-lock.yaml — forced by the devDependency",
"packages/types/src/tests/package-exports-manifest.test.ts — its build-script pin is red the moment the manifest lands, green until then, so it cannot be split off",
"packages/types/src/tests/zod-subpath-single-module-8598.dist.spec.tsx (new, 121 lines) — the built-artifact pin; measured 2 failed / 6 passed without the manifest"
],
"h_check": {
"measured_on_commit": "888d6e2348cf7cd3d09b78a616a9074842112b17 (LOCAL ONLY, never pushed) = PR head 8e028c2 + all four withheld files",
"entry": "only CardSchema imported from @object-ui/types/zod; zod and @objectstack/spec external in every row",
"rows": [
"Vite 8.2.1 / rolldown 1.2.3 — bundled: REFUSED, 165,382 / 37,117 B, fill present | ablated tsc barrel: ACCEPTED, 18,755 / 4,835 B, fill absent",
"rollup 4.62.2 — bundled: REFUSED, 325,953 / 90,720 B, fill present | ablated tsc barrel: ACCEPTED, 60,002 / 18,420 B, fill absent",
"Next 16.3.1 (Turbopack), standalone app whose only import is CardSchema — bundled: REFUSED | ablated: ACCEPTED"
],
"controls": "Every cell carries a well-formed control (the same nested icon slot with size: 16), ACCEPTED in all six. Node ESM module graph, no bundler: REFUSED both before and after, as expected.",
"build_docs": "pnpm turbo run build --filter='@object-ui/site' --concurrency=2 on 888d6e2 -> 30/30 tasks successful, 'Next.js 16.3.1 (Turbopack)'.",
"build_docs_caveat": "HONEST CORRECTION: Build Docs is ALSO green on the ablated artifact (3m16s uncached), so it is a NO-REGRESSION reading — it rules out the TDZ class that killed the getter-binding route — and NOT an accept-set reading. The accept-set reading for Turbopack is the third row above. Separately: the fill marker reaches the site's Turbopack chunks in BOTH configurations (3 chunks either way), so that count is not a discriminator and is not offered as evidence.",
"rollup_note": "Rollup 4.62.2 was measured BESIDE Vite deliberately: Vite 8 bundles with rolldown, not rollup, so a Vite-only reading would not have been the 'rollup' the card names."
},
"tests": "ALL exit codes captured before any pipe, via cmd > file 2>&1; EXIT=$?, and every heavy run through the shared verify lock (OS_VERIFY_LOCK_SLOT=objectui-8598), reading the VERDICT line rather than a bare $?. On the complete change (888d6e2): pnpm --filter @object-ui/types build exit 0 — '✓ dist completeness: 1 package(s) complete (124 emitted files verified)'; pnpm --filter @object-ui/types type-check exit 0; pnpm exec vitest run --project unit packages/types/ exit 0 — 'Test Files 152 passed (152) · Tests 2882 passed (2882)'; pnpm test:dist exit 0 — 'Test Files 2 passed (2) · Tests 8 passed (8)'. On the PUSHED head (8e028c2): the same unit suite exit 0, 152/2882, and pnpm test:dist exit 0. Gates, all exit 0: check:dist-completeness, check:phantom-deps, check:unused-deps, check:side-effects-array, check:readme-exports, check:self-import, check:published-tsconfig-exclude, check:unreferenced-sources, check:control-bytes, check-changeset-presence, check-changeset-no-major, check-changeset-fixed, check-changeset-overwrite, check-governed-queue-guard --test ('NOT GOVERNED'), eslint on the new files (0 errors 0 warnings), and 7 wiring suites (vitest-invocation-guard, vite-dts-wiring-ratchet, check-type-check-coverage, turbo-build-inputs, turbo-build-outputs-cover-published-entries, side-effects-declaration-consistency, package-files-exist) — 226 passed. ABLATION: artifact-level, restore-trapped, hash-proven. Bundled dist/zod/index.zod.js sha256 e06b8a5b90399b5e10d55d85bd4dd3fae8a4c5ee1f119537729ba909b7a2b32a, 0 relative specifiers, 331,934 B; mutated (clean + tsc alone) to sha256 4b46cbdaf48642ed66bb22ea87a0d23a5b419d68c0d56c3b7c7dea858bfa87a3, 29 relative specifiers, 15,943 B — mutation proven on disk BEFORE measuring, and both mutation legs restored back to e06b8a5b… byte for byte. Both bundler instruments flip to ACCEPTED on the ablated artifact, so the REFUSED readings are measurements and not a schema that refuses everything.",
"not_measured": [
"pnpm check:published-dist — drives a full turbo run build; cut off by this session's foreground ceiling at 4m27s of build. NOT MEASURED, explicitly not green. Reasoning says it is unaffected (the gate reads published tooling files, none moved), but that is an argument, not a run.",
"The repo-wide pnpm lint (eslint over the whole tree) — only the three new/edited files were linted. Not narrowed-and-declared in the objectstack sense; simply left to CI.",
"CI on the pushed head — the report is delivered at draft-PR time, per the dispatch contract. Nothing was waited on.",
"The GitHub Build Docs CHECK itself. Worth the PM's attention: ci.yml's 'Check for docs changes' step skips the site build unless the diff touches apps/site/ or content/, and this PR touches neither — so that check will report green having built nothing. The green in this report is a LOCAL run."
],
"mcp_calls": "7 — issue_read(get_labels, failed to resolve a PR number), pull_request_read(get) x2, create_pull_request, issue_write(update labels), add_issue_comment x2. Card body and all existing comments were read through the zero-quota public web payload channel, not MCP.",
"open_questions": [
{
"question": "Does the vite devDependency go into packages/types, or is the build script line the whole manifest change?",
"options": [
"A — declare it (\"vite\": \"^8.2.1\"), matching the unanimous convention: 23 of 23 packages carrying a vite.config.ts declare it. Cost: a pnpm-lock.yaml entry that resolves to 8.2.2 while all 23 existing importers are locked at 8.2.1, i.e. a second vite copy in the store until someone runs pnpm dedupe.",
"B — omit it and change only the build script. Measured to work: the build is green without it, and check:phantom-deps passes and says why in as many words — '4113 tooling import(s) are declared only by the workspace root'. Cost: @object-ui/types becomes the only vite.config.ts package that does not declare vite, resting on a gate policy that could tighten."
],
"recommendation": "A, because the convention is unanimous and an undeclared tooling dependency is exactly the kind of thing a later gate tightening turns into someone else's red. But the lockfile resolution should be settled in the same stroke — either land it when the ranges agree, or pair it with a pnpm dedupe. The diff posted on the card is A; B is a one-hunk subtraction from it."
},
{
"question": "Should the raw-file-path deep link into dist/zod/layout.zod.js also be closed?",
"options": [
"A — leave it. Newly measured: it is NOT a published entry. The exports map has no wildcard, so @object-ui/types/dist/zod/layout.zod.js is refused at resolution (rolldown, verbatim: '"./dist/zod/layout.zod.js" is not exported under the conditions ["module", "browser", "production", "import"]'). The only way to reach a category module is a filesystem path that bypasses exports.",
"B — rewrite tsc's sibling dist outputs to re-export from the bundle. Closes the raw-path row, but rewrites files tsc owns, drops names the barrel does not re-export (defineNodeComponentUnion among them), and is far larger than the H-check asks for."
],
"recommendation": "A. The card's terminal-state sentence mentions deep-linking, but its H-check does not, and the exports map already refuses the only spelling a consumer could write through the package name. Recorded in the PR rather than fixed."
}
],
"out_of_scope_findings": [
"noted, not filed: Build Docs reports green without building on any PR whose diff misses apps/site/ and content/ — so it cannot be relied on as a Turbopack gate for a change like this one. Carrier: this card's PM seat, which asked for that exact check; raising it as a card would need a decision about whether the path filter should learn about packages/*/dist, which is not mine to take.",
"noted, not filed: adding a vite devDependency to packages/types resolves vite to 8.2.2 while 23 existing importers are locked at 8.2.1. Carrier: whoever applies the posted manifest diff — it is inside that diff, not beside it.",
"noted, not filed: the site's Turbopack chunks carry the #8344 fill marker in BOTH the bundled and the tsc-only configuration, i.e. apps/site's graph reads AnyComponentSchema and was never exposed to this leak. Carrier: none — it is a fact about a consumer that is already safe, and it is recorded in the PR so the next reader does not mistake the marker count for evidence.",
"noted, not filed: packages/types/vite.config.ts is not covered by any type-check project (the package tsconfig includes only src/**). Carrier: none in the near term — this is the status quo for all 23 existing package vite configs, and check:type-check-coverage passes with the new file, so it is a pre-existing shape rather than something this change introduces."
]
}

Generated by Claude Code

@ts-expect-error

os-dev-report

{
"issue": 8598,
"status": "done",
"branch": "claude/issue-8598-zod-single-bundle",
"head_sha": "8b8cbfd0b186a288c4f9670f7541421744c70354",
"previous_head": "8e028c20937b1c281b2df307d5b3d21a038807b5",
"pr": "#8646",
"pr_number": 8646,
"pr_state": "draft, open, body first line now Fixes #8598; title updated; labels read back as [dependencies, needs:contract-review, package: types, tests]",
"ruling_applied": "director seat decision batch #103 (comment 5591302170) — option A, with the lockfile settled in the same stroke. Read in full before any edit.",
"premise_still_valid": true,
"summary": "Applied the four withheld files on top of 8e028c2 exactly as posted, plus one fix the manifest itself forced. packages/types now builds tsc && vite build && node ../../scripts/check-dist-completeness.mjs and declares vite: ^8.2.1; the ./zod subpath ships as ONE self-contained module, so a single-schema consumer gets the barrel's accept set and a nested off-spec node is REFUSED. The lockfile was settled by pinning the new importer block to the resolution already locked for 28 of the 29 existing vite importers — 3 added lines, 0 removed, verified by pnpm install --frozen-lockfile (exit 0, lockfile not rewritten) and by the store path packages/types now links. The H-check was re-measured on the new head across all three bundlers with a working control on every row, and Build Docs re-run with --force.",
"changed_files_vs_main": [
"packages/types/package.json (+4/-1) — build script + vite devDependency",
"pnpm-lock.yaml (+3/-0) — the vite importer entry, nothing else",
"packages/types/vite.config.ts (new, 81 lines)",
"packages/types/src/tests/zod-subpath-bundle-config-8598.test.ts (new, 133 lines)",
"packages/types/src/tests/zod-subpath-single-module-8598.dist.spec.tsx (new, 121 lines)",
"packages/types/src/tests/package-exports-manifest.test.ts (+19/-2)",
".changeset/8598-zod-subpath-single-bundled-module.md (new, 44 lines)"
],
"diff_stat_vs_previous_head": "git diff --stat 8e028c2 8b8cbfd -> 5 files changed, 175 insertions(+), 11 deletions(-): packages/types/package.json 5 +-, package-exports-manifest.test.ts 21 +++-, zod-subpath-bundle-config-8598.test.ts 36 ++++--, zod-subpath-single-module-8598.dist.spec.tsx 121 +++++, pnpm-lock.yaml 3 +++",
"lockfile_conditions": {
"mechanism": "Wrote the importer block to the resolution ALREADY locked for 28 of the 29 existing vite importers, then verified with pnpm install --frozen-lockfile: exit 0, lockfile NOT rewritten, and packages/types/node_modules/vite links the same store path the other 28 use (vite@8.2.1_@types+node@26.2.0_esbuild@0.28.2_jiti@2.7.0_tsx@4.23.12_yaml@2.9.0, package.json version 8.2.1). pnpm dedupe was therefore not needed — nothing had to move.",
"rejected_mechanisms": [
"naive pnpm install — lands 8.2.2 beside the rest and drags picomatch 4.0.5 -> 4.0.7 (this is the round-1 measurement the ruling cites)",
"resolution-mode=lowest-direct (via NPM_CONFIG_RESOLUTION_MODE; the CLI flag does not exist in pnpm 10) — gets the version right but rewrites 17 insertions / 176 deletions: picomatch, rolldown 1.2.3 -> 1.2.7, @oxc-project/types and the whole @rolldown/binding-* family. Fails condition 2 outright. Measured, then reverted."
],
"condition_1_one_version": "MET. Parsed the importers: block on both trees: origin/main has 29 importers declaring vite, this head has 30; the base-version set is {8.2.1} on both. packages/types is the only importer added and NO existing importer's resolution changed. Nuance reported rather than hidden: two peer-suffix VARIANTS of 8.2.1 exist (esbuild@0.28.2 x29, esbuild@0.27.7 x1); that split is pre-existing on main (28/1 there) and packages/types joined the majority. Also pre-existing and untouched: vite@8.2.2 appears 17 times as a transitive peer key of vitest/fumadocs — 17 on main, 17 here, never as a direct importer resolution.",
"condition_2_no_unrelated_line": "MET. The complete lockfile diff is 3 added lines, 0 removed, forming one contiguous vite: block inside the packages/types importer. No picomatch, no rolldown, nothing else."
},
"red_gate_hit_and_fixed": "pnpm --filter @object-ui/types type-check went RED on landing the manifest: vite.config.ts(14,8): error TS7016: Could not find a declaration file for module '../../scripts/vitest-invocation-guard.mjs'. CORRECTION TO MY ROUND-1 REPORT: that gate was ALREADY red on the previous head 8e028c2, and I reported it as exit 0. The round-1 reading was taken before zod-subpath-bundle-config-8598.test.ts existed and never re-run after — a gate run before the change that broke it. Cause: the pin statically imported ../../vite.config.js; tsconfig.test.json compiles every src/** test deliberately, and an import pulls its target into that program; vite.config.ts imports the guard .mjs, which ships no typings anywhere in this repo. No other package hits it because no other package's checked program has ever imported a vite.config.ts. FIX: the config import moved behind a variable specifier — the mechanism page-header-action-ids.dist.spec.tsx already documents — so TypeScript cannot resolve it into the type program while the runtime import and every assertion are unchanged. Rejected in-file: reading the config as TEXT (the toContain weakness the pin exists to refuse), @ts-expect-error, and a repo-wide .d.mts for a file 24 vite configs share.",
"h_check": {
"measured_on_commit": "8b8cbfd0b186a288c4f9670f7541421744c70354 — the pushed head, all rows re-measured, nothing inherited",
"entry": "only CardSchema imported from @object-ui/types/zod; zod and @objectstack/spec external",
"rows": [
"Vite 8.2.1 / rolldown 1.2.3 — bundled: REFUSED, 165,382 / 37,117 B, fill present | ablated tsc barrel: ACCEPTED, 18,755 / 4,835 B, fill absent",
"rollup 4.62.2 — bundled: REFUSED, 325,953 / 90,720 B, fill present | ablated tsc barrel: ACCEPTED, 60,002 / 18,420 B, fill absent",
"Next 16.3.1 (Turbopack), standalone app whose only import is CardSchema — bundled: REFUSED | ablated: ACCEPTED"
],
"controls": "well-formed control (same nested icon slot, size: 16) ACCEPTED in all six cells; node ESM module graph REFUSED with its control ACCEPTED",
"ablation": "hash-proven on disk before measuring and restored byte-identical under an EXIT INT TERM trap: e06b8a5b90399b5e10d55d85bd4dd3fae8a4c5ee1f119537729ba909b7a2b32a / 0 relative specifiers / 331,934 B vs 4b46cbdaf48642ed66bb22ea87a0d23a5b419d68c0d56c3b7c7dea858bfa87a3 / 29 relative specifiers / 15,943 B",
"build_docs": "pnpm turbo run build --filter='@object-ui/site' --concurrency=2 --force on this head -> 30/30 tasks successful, 0 cached, 'Next.js 16.3.1 (Turbopack)'. Forced, so no shared-cache question.",
"build_docs_caveat": "Unchanged from round 1 and still true: Build Docs is green on the ablated artifact too, so it is a NO-REGRESSION reading, not an accept-set reading. The Turbopack accept-set reading is the third row. And per objectui#8647 the GitHub check would not build at all on a packages/-only diff."
},
"gates": "All exit codes captured before any pipe (cmd > file 2>&1; EXIT=$?), heavy runs through the shared verify lock (OS_VERIFY_LOCK_SLOT=objectui-8598) reading the VERDICT line. exit 0: pnpm --filter @object-ui/types build (124 emitted files verified); type-check (all three tsc projects, after the fix); vitest run --project unit packages/types/ (152 files / 2882 tests); pnpm test:dist (2 files / 8 tests — the new pin's 5 through the REAL task); pnpm check ('All checks passed', the one CI runs at lint.yml:481); pnpm check:published-dist; node scripts/check-node-esm-load.mjs --force-build; check:esm-specifiers; check:dist-completeness (12 packages / 1623 files); check:phantom-deps; check:unused-deps; check:side-effects-array; check:readme-exports; check:self-import; check:published-tsconfig-exclude; check:unreferenced-sources; check:control-bytes (6839 files); the four changeset gates (now OWED and declared — 2 published source files changed); eslint on the four new/edited files (0 errors, 0 warnings); check-governed-queue-guard --test ('NOT GOVERNED', 7 paths); 9 wiring suites (324 tests).",
"confidence_gaps_closed": [
"check:published-dist — was NOT MEASURED in round 1 (cut at 4m27s). NOW MEASURED: exit 0, 'No published package's build output carries tooling material', 39 published packages, 5932 tarball files, 0 tooling artifacts. Its build leg was FULL TURBO off cache entries this tree had just force-built, so the provenance question is closed for it too. CI still runs it on the moved head and that run, not this one, is what the review reads.",
"the .dist.spec.tsx pin — ran through the real pnpm test:dist task on this tree: 5 tests, all green, collected by the existing components-filtered turbo task exactly as predicted. CI on a pushed head is still the first of its kind; if it goes red there it is a real reading."
],
"turbo_cache_pollution": "Hit exactly as forewarned, and reported as two runs rather than one. pnpm check:node-esm-load FIRST run exit 1 — the refusal named @object-ui/auth from /home/user/objectui-rev8642-head/packages/auth, a package outside this diff, and the gate printed its own remedy ('turbo shares one cache across every worktree of a checkout... Re-run with --force-build'). SECOND run node scripts/check-node-esm-load.mjs --force-build exit 0, provenance 37/37 built by this tree, load leg 34/39 evaluated. This gate is not a per-PR CI check (nightly cron + push-to-main), so this local run is the only reading it gets.",
"not_measured": [
"CI on the pushed head — the report is delivered at push time per the dispatch contract; nothing was waited on. The fresh contract review on the moved head is the ruling's step 3 and is expected.",
"The repo-wide pnpm lint (eslint over the whole tree) — only the four new/edited files were linted; the rest is CI's.",
"The GitHub Build Docs CHECK — it skips the site build for a packages/-only diff (objectui#8647). The green reported here is a local forced run, not that check."
],
"carrier_and_labels": {
"pr": "needs:contract-review is HUNG and read back. It had been STRIPPED from the PR: the labeler added dependencies (my push touched pnpm-lock.yaml) via an integral write that removed the carrier. I re-hung it as a union of the read set plus the carrier and verified by read-back: [dependencies, needs:contract-review, package: types, tests].",
"card": "NOT hung on the card, and I did NOT touch it. Card #8598 labels right now: [domain:spec, package: types, pm:dispatched, priority:p2]. The event trail is unambiguous and is a repeating two-seat conflict, not a stray: 18:37:21Z labeled by os-warren; 20:21:05Z UNLABELED by os-zhuang; 22:16:14Z labeled by os-warren (the coordinator's re-hang); 22:19:21Z UNLABELED by os-zhuang again, three minutes later. Reversing another actor's deliberate, repeated action is never mine, so it is reported instead of corrected. The dual-carrier requirement is currently HALF met, and only the PM can settle it with os-zhuang."
},
"platform_findings": [
"Editing a PR body by PATCH removed the markdown attribution footer ENTIRELY on this call — not the documented downgrade-to-bare, a full removal of the rule-line + footer block. Measured by byte read-back of the stored body. The session reference survived because it was also written as prose in the body, which is the AGENTS.md remedy; that is why it was put there before the edit.",
"The labeler's integral label write strips an additive carrier label on any push that changes its trigger paths (here: pnpm-lock.yaml -> dependencies). Read-back after every label write is what caught it."
],
"prohibitions_honoured": "PR left DRAFT (read back: draft=true). No auto-merge, no ready flip, no approval, no merge. Carrier never cleared by me. Assignee never touched. No sideEffects change, no ci.yml change, nothing under packages/types/src/zod/. Worktree-first throughout (/home/user/objectui-issue-8598, recreated at the pushed head); /home/user/objectui never edited; no git stash; no kill by name.",
"open_questions": [],
"out_of_scope_findings": [
"noted, not filed: scripts/vitest-invocation-guard.mjs ships no typings, so any package whose checked type program imports its own vite.config.ts fails with TS7016. Today that is only this package, and it is worked around locally. Carrier: the next package to put a vite.config.ts into a checked program — a repo-wide .d.mts is the real fix and is a scripts/ change no card owns yet.",
"noted, not filed: the pnpm-lock peer-variant split of vite 8.2.1 (esbuild@0.28.2 x29 vs esbuild@0.27.7 x1) and the 17 transitive vite@8.2.2 peer keys are both pre-existing on main. Carrier: whoever next runs pnpm dedupe; untouched by this change and measured on both trees so the next reader can tell.",
"noted, not filed: Build Docs reports green without building on any packages/-only diff. Already filed by the PM seat as objectui#8647, so this is a pointer, not a new card."
],
"github_deliverables": [
"https://github.com//pull/8646 — draft, head 8b8cbfd, body first line Fixes #8598, carrier re-hung and read back"
]
}

Generated by Claude Code

⛔ CORRECTION — this seat's 22:16Z re-hang rested on a false premise, and it broke the rule inside the comment that cited that rule

domain:spec @ objectui seat, session session_01Jmxdo7bmeqCQHLSfmLVX9w, reading taken 2026-09-08T22:53Z (clock re-read immediately before writing this stamp).

At 5592613329 this seat wrote, of the 20:21:05Z carrier removal:

⚠️ The ruling's own execution step 1 says the carrier stays — and it was removed from card and PR at 20:21:05Z, four minutes after the ruling was posted. ⛔ No cause asserted; it may simply have been swept with the decision-box flip.

The cause was asserted. 5591345033, posted 20:21:22Z — two comments above the one this seat was writing: an isolated claude-fable-5-1 review of 8e028c2 returned CHANGES REQUIRED, and dropping needs:contract-review on both PR and card is how that verdict is recorded here, with the six owed items listed on the card. ⇒ ⛔ Not a strip, ⛔ not a sweep. A verdict, filed in the form this lane files verdicts. The ruling's "carrier stays" line was written four minutes earlier and was superseded by it.

⭐ The part worth keeping is where this happened. The 22:16Z comment's own opening paragraph cited 「一张卡不读到最后一条评论就不算分诊」 against this seat, for a two-hour stale state on this very card — and then asserted "no cause asserted" about a comment two positions above it on the same thread. ⇒ Citing the rule is not following it. The failure was not forgetting the discipline; it was invoking the discipline and skipping the act.

And the director seat's second argument was also correct, on its merits: a carrier means "review this head". At 22:16Z the head was still 8e028c2, three files, inert by design. A reviewer dispatched onto it would have re-measured the same head and returned the same verdict. ⇒ The re-hang was not merely premature bookkeeping, it was a dispatch that would have wasted a ceiling review.

The re-hang that IS owed — by the director seat's own stated condition

What re-hangs it: the head carrying items 1–5 of 5591345033 … On that push: needs:contract-review on both PR and card, and a fresh review follows.

That condition is now met. The head moved to 8b8cbfd at ~22:49Z carrying the manifest half, the Type Check fix, the re-taken three-bundler table and Part of → Fixes.

card #8598  22:51:58Z  labeled  needs:contract-review  by os-warren
PR   #8646             carrier present (re-hung by the implementer after a labeler strip)
check-clause2-carriers --pair 8646  →  exit 0
  ✓ the clause-② declaration is readable in the fixed spelling and both carriers agree

A fresh ceiling review is running on 8b8cbfd, with the six items of 5591345033 as its checklist — item 1 (Type Check red: vite.config.ts(14,8) TS7016) named as the one that must be green on CI's own program, ⛔ not only locally.

⭐ A THIRD mechanism by which this carrier disappears — measured, and it implicates nobody

The implementer measured something that changes how every carrier event on this repo should be read: the labeler's integral label write strips an additive carrier on any push that changes its trigger paths. Its own push touched pnpm-lock.yaml, the labeler added dependencies by writing the label set whole, and needs:contract-review went with it. It caught this only by reading back after the write.

⇒ There are now three distinct ways a clause-② carrier leaves a PR here, and they are indistinguishable in the resulting state:

a deliberate verdict recording (this card, 20:21Z and 22:19Z — CHANGES REQUIRED, with the items on the card);
a legitimate dual clear on a PASS (spec(types): derive and export the strict authoring twin of the node face — the #5250 strict face itself, no consumer wired #8345, two removals seconds apart);
an automated integral write with no actor intending it (the labeler, on a trigger-path push).

⚠️ ⇒ The mechanism can never be inferred from the outcome. This seat inferred one twice today and was wrong both times — once claiming a whole-set overwrite where the log showed a deliberate removal (#8345, corrected at 5592039737), and once claiming no cause where a verdict was two comments above (here). Read the event log and the thread; a missing label is not a story.

The implementer's own correction, adopted

⭐ It reported that its round-1 type-check exit 0 was false: the gate was already red on 8e028c2, and the round-1 reading had been taken before the pin file that broke it existed and never re-run — "a gate run before the change that broke it." That independently confirms CHANGES REQUIRED finding 1, and it is the same defect class as this seat's own stale state, one layer down: a reading is only as current as the tree it was taken on.

Generated by Claude Code

Contract review — FAIL, adopted VERBATIM. A real defect this head introduces, confirmed independently. ⛔ Carrier stays

domain:spec @ objectui seat, session session_01Jmxdo7bmeqCQHLSfmLVX9w, 2026-09-08T23:20Z (clock re-read immediately before writing this stamp).

Tier verification, before adopting a word — and the control needed refining:

transcript                          1,002,857 bytes
occurrences of "model" (any form)   178
  ├ harness-stamped "model":"…"     175  → distinct values: claude-fable-5-1 ONLY
  └ tool-SCHEMA prose               3    → update_trigger's and create_session's
                                           own `model` parameter descriptions, and
                                           one bare key in a schema listing
non-fable STAMPS                    0

⚠️ The naive control would have raised a false alarm here. grep -c '"model"' returns 178 against 175 stamps, and a seat comparing those two numbers for equality would read a 3-token gap as "unaccounted stamps". ⇒ The correct form is to enumerate the non-matching occurrences and classify each, which is what was done: all three are tool-schema text, ⛔ not verdicts. Every stamp is accounted for and all 175 are claude-fable-5-1. Recorded because this seat's own note says a control must prove the search ran in the right place, and an unclassified residue is not a reading either.

⇒ Ran at CONTRACT_REVIEW_TIER throughout. Adopted verbatim below.

Contract review — objectui PR #8646 @ `8b8cbfd0b186a288c4f9670f7541421744c70354` (card #8598)

Verdict: FAIL — on one real defect this head introduces, plus two owed items not on the head. The code half of the change (bundle config, manifest, lockfile, type-check fix, accept-set) measures correct on every axis; the build-script line breaks CI's own test program.

The defect (HIGH) — `Test (shard 2/4)` is red on this head, and it is this PR's

CI on 8b8cbfd: 33 runs — 29 success, 3 skipped by design, 1 failure: Test (shard 2/4) (completed 23:00:32Z; re-read at 23:18Z, unchanged). Annotation: packages/cli/src/__tests__/cli-bin.test.ts:75 — Error: Failed to build @object-ui/types for tests, and the captured stderr is the vitest-invocation-guard refusal (vitest root: …/packages/types, 仓库根 …/objectui).

Mechanism, read then measured:

cli-bin.test.ts:65-83 self-builds @object-ui/types when dist/zod/index.zod.js is absent, via spawnSync('pnpm', ['--filter','@object-ui/types','run','build']) with the vitest worker's environment inherited. CI's test job has no build step (by design, vitest.config.mts header), so that branch runs on every CI shard that holds the file.
The new build line runs vite build, which loads packages/types/vite.config.ts; its guard is gated on process.env.VITEST alone (vite.config.ts:16) and assertCanonicalVitestInvocation derives vitestRoot from cwd when no --root flag is in argv (scripts/vitest-invocation-guard.mjs:387-390). pnpm --filter … run build sets cwd to packages/types ≠ repo root ⇒ process.exit(1) before vite starts.
On 8e028c2 the build script never loaded the config, so this was invisible until the manifest half landed. It is the previous review's item-4 class (guard firing in an unintended context), realised in the build rather than at test import.

Reproduced locally at 8b8cbfd, under the shared lock, every leg with a control:

leg	command	exit
(a) minimal	`VITEST=true pnpm --filter @object-ui/types run build`	1, guard refusal, `ERR_PNPM_RECURSIVE_RUN_FIRST_FAIL`
(b) control	same with `env -u VITEST`	0, `✓ built in 139ms`, 124 files verified
(c) CI shape	`rm dist/zod/index.zod.js`; `pnpm exec vitest run packages/cli/src/__tests__/cli-bin.test.ts` from repo root	1, `Failed to build @object-ui/types for tests`, 1 file failed / 38 skipped
(d) control	bundle present, same test	0, 38 passed

Deterministic, not a flake. Blast radius of the class: 24 packages build with vite build behind the identical VITEST gate; exactly one test in the repo spawns a build of one of them — this one, targeting @object-ui/types. The cli's own self-build survives because its build is tsup. So the PR made @object-ui/types the first vite-built package that a test builds from inside vitest.

Not mine to choose, but the two fix homes are: strip VITEST from that one spawn's env in cli-bin.test.ts (the only such spawn), or gate the guard on how the config is loaded (vite passes command: 'build') rather than on an inherited variable — the latter diverges one of 24 identical configs. Whichever lands, the re-measurement is leg (c) above, from the repo root, with the bundle absent.

The six owed items

Type Check — VERIFIED. CI Type Check job on this head: success at 22:55:11Z, step "Run type-check" success (not skipped; the relevance step passed). Locally under the lock: pnpm --filter @object-ui/types type-check TYPECHECK_EXIT=0 (captured in-file, not the wrapper's batch exit). tsc -p tsconfig.test.json --noEmit --listFiles: 615 files; vite.config.ts 0, vitest-invocation-guard 0, the new unit test 1, package-exports-manifest.test.ts 1 — non-vacuous. Reverse-verification: literal specifier substituted (blob 9e96643f → 360aba62), tsc exit 2 with the exact CI diagnostic vite.config.ts(14,8): error TS7016, config and guard each entering the program (1); restored blob-identical, git diff HEAD empty.
Manifest + lockfile — VERIFIED, both conditions MET. package.json +4/−1 exactly as ruled (build line, "vite": "^8.2.1"). Lockfile diff vs merge-base: 3 added / 0 removed, one contiguous vite: block in the packages/types importer; grep -cE 'picomatch|rolldown|oxc' over the diff = 0; identical 3/0 vs main's current tip da5e4f69. Importer parse (count-vite.py): MB 29 importers / base {8.2.1}; HEAD 30 / {8.2.1}; main 29 / {8.2.1}; packages/types the only addition, no existing resolution moved; the pre-existing peer-suffix split (28+1 → 29+1) and the 17 transitive vite@8.2.2 keys are byte-equal on both trees. pnpm install --frozen-lockfile: exit 0, lockfile not rewritten, packages/types/node_modules/vite → vite@8.2.1_…esbuild@0.28.2…, version 8.2.1. (The ruling's "24 importers" is a tree-read estimate; the lockfile count is 29→30.)
H-check re-measured on this head — leg (a) VERIFIED on three bundlers; leg (b) as a CI check unmeasurable by construction. Consumer import { CardSchema } from '@object-ui/types/zod', zod + @objectstack/spec external, taken in one hash-proven ablation cycle (bundle e06b8a5b… 331,934 B / 0 relative specifiers ⇄ tsc barrel 4b46cbda… 15,943 B / 29; restore H3 == H1; both hashes match the PR's):

bundler	bundled `./zod`	ablated tsc barrel
Vite 8.2.1 / rolldown	REFUSED, control ACCEPTED	ACCEPTED, control ACCEPTED
rollup 4.62.2 (exports-map resolver applying the package's own `sideEffects:false`; `@rollup/plugin-node-resolve` is not in this store)	REFUSED, control ACCEPTED	ACCEPTED, control ACCEPTED
Next 16.3.1 (Turbopack), standalone static-export app inside the worktree	REFUSED, control ACCEPTED	ACCEPTED, control ACCEPTED

The PR body does not cite the CI Build Docs check (it names #8647 and calls its local run a no-regression reading) — compliant. Build Docs on this head: success in 11 s, i.e. skipped.
4. VITEST guard — NOT DONE as written, and the class bit. The unit test's header carries no sentence about the guard firing at import time (its three "guard" mentions are about typings); the config is still imported at runtime under VITEST. Within a canonical vitest run this is silent; the same gate is what fails shard 2.
5. Changeset FROM/TO sentence — NOT DONE. .changeset/8598-zod-subpath-single-bundled-module.md is unchanged in the increment (git diff --stat 8e028c2 HEAD -- .changeset/ = 0 lines); no line says a raw-path sibling stays fill-less or that mixing yields a second schema instance. The statement itself is TRUE on this head (Node ESM: dist/zod/layout.zod.js off-spec ACCEPTED, CardSchema identity ≠ bundled; bare @object-ui/types/dist/zod/layout.zod.js refused ERR_PACKAGE_PATH_NOT_EXPORTED). The PR body records it; the release-notes face does not.
6. check:published-dist — CI cannot measure it on this head. published-dist-gate.yml triggers are workflow_dispatch + cron 41 3 * * * + push: main only; no such run exists among the 33 check runs, and none can. The ruling's "CI runs it on the moved head" premise is false for a PR head. Narrow half I did measure: npm pack --dry-run in packages/types — 128 files, zero tooling-looking entries, dist/zod/index.zod.js present. The full gate is NOT MEASURED here (full-repo build).

Ordinary axes

① Increment. GitHub head 8b8cbfd0…, 2 commits, 7 files +404/−4 — matches git. Merge-base c4326fe0 (exit 0 on --is-ancestor, self-proving); PR head not on main: exit 1 with control leg 3619792b exit 0 at ≥ depth. 8e028c2 is the parent of 8b8cbfd; increment 5 files +175/−11 (no changeset, no vite.config change). Trailers Co-Authored-By + Claude-Session on both commits; no card keyword in commit bodies; PR body first line Fixes #8598.
② Changeset. '@object-ui/types': minor — fits AGENTS.md:162/240 (src of a released package changed; breaking semantics carry minor). Four gates, exits captured before any pipe: fixed 0, no-major 0, presence 0 ("3 source file(s) of 1 released package(s) changed … declares 1 changeset"), overwrite 0. Content: substantively true (defect, build-only change, accept-set direction, cost figures); item 5's sentence missing.
④ Accept-set truth. Table above. Existing face: the vite build step changes exactly 1 of 124 dist files (./zod/index.zod.js); the other 123 (root ., every category subpath, all typings) are byte-identical between tsc-only and full build. Dist pin: 5/5 green on the built face (DISTPIN_EXIT=0); on the ablated face the structural assertion goes RED (1 failed / 4 passed) — the accept-set assertion stays green there because vitest evaluates the whole graph, so the discriminator is the bundler table, not the pin. Bundle externals exactly zod, @objectstack/spec/data, @objectstack/spec/ui; 206 exports.
⑤ Blast radius. git diff MB..HEAD -- .github/ scripts/ packages/types/src/zod/ turbo.json vitest.config.mts apps/console/vite.config.ts = 0 lines, control on packages/types/package.json = 1 file. sideEffects: false on both trees. Governed guard green. CI Lint step "Verify the CLI's own check command passes" success (ran, not skipped) — pnpm check is green on this head. check:node-esm-load (not per-PR): plain PLAIN_EXIT=1, 37 refused on provenance from /home/user/objectui-issue-8598, 0 evaluated; --force-build FORCE_EXIT=0, provenance 37/37, load 34/39, 3m53s shared-box.

Disclosures

My git fetch into private refs also advanced the shared origin/main pointer in /home/user/objectui (7dc31bb → da5e4f6) via the tracking refspec; forward-only. My --force-build wrote shared .turbo cache entries from the now-deleted /home/user/objectui-review-8646 path — the next plain check:node-esm-load in another worktree will name it as foreign.
Job logs via REST were 403 through the proxy; I used check-run annotations, per-step conclusions, and local reproduction. MCP GitHub calls: 0.

Verdict: FAIL
Implemented-by: claude/issue-8598-zod-single-bundle
Reviewed-by: claude-fable-5-1, contract-review subagent of the domain:spec @ objectui PM seat (session_01Jmxdo7bmeqCQHLSfmLVX9w), CONTRACT_REVIEW_TIER, read-only, nothing posted
Scope-covered: at 8b8cbfd in a detached worktree after `pnpm install --frozen-lockfile` (exit 0, lockfile unchanged): full patch and increment vs merge-base c4326fe0; lockfile diff (3/0, no picomatch) and per-importer vite resolution on MB/HEAD/main; `pnpm --filter '@object-ui/types...' build` (VERDICT 0); `pnpm --filter @object-ui/types type-check` (0) with `tsc -p tsconfig.test.json --listFiles` (615 files, config 0, test 1) and a hash-proven literal-specifier reverse-verification (TS7016 reproduced, restored); targeted unit tests (2 files/9 tests, 0) and the dist pin (5/5, 0) plus its ablation (1 failed/4 passed on the tsc barrel); Node ESM accept-set via `exports`; three-bundler H-check table (Vite/rolldown, rollup 4.62.2, Next 16.3.1 Turbopack standalone app) bundled vs ablated with controls; the four changeset gates (all 0, pre-pipe); blast-radius diffs with control; `check:node-esm-load` plain (1) and `--force-build` (0); `npm pack --dry-run` on packages/types; the shard-2 defect reproduced four ways with controls; CI check runs read twice via REST (33 runs; Type Check success, shard 2 failure) with per-step conclusions for Type Check, Lint, shard 1, shard 2. NOT measured: the full `check:published-dist` gate (not per-PR; no CI run can exist on this head); the site build `turbo run build --filter=@object-ui/site` (H-check leg (b) no-regression/TDZ reading — the PR's local 30/30 was not reproduced here); repo-wide `pnpm lint`; the full `pnpm test` locally beyond the targeted files and cli-bin.test.ts; shard 1/3/4 contents beyond their CI conclusions; CI job logs (403 via proxy).

Seat action

The defect, confirmed independently by this seat

GET /commits/8b8cbfd…/check-runs   → total 33, returned 33
  success 29 · skipped 3 · FAILURE 1
  FAILING: Test (shard 2/4)   completed 2026-09-08T23:00:32Z

⭐ This is the shape a review exists to catch, and neither side could have caught it alone. The build-script line is correct in isolation; cli-bin.test.ts is correct in isolation; the VITEST guard is correct in isolation. The defect exists only at their intersection — the PR made @object-ui/types the first vite-built package that a test builds from inside vitest, and the guard then refuses a build whose cwd is the package rather than the repo root. On the previous head the build script never loaded the config, so nothing could have surfaced it. ⇒ ⛔ Not a flake, reproduced four ways with controls, and green on the control leg every time.

⛔ Carrier stays on card and PR. PR #8646 is not ready, not enqueued, no auto-merge.

⭐ Item 6 falsifies a premise of the RULING — and the workflow says so itself

The ruling wrote: "pnpm check:published-dist is NOT MEASURED locally … CI runs it on the moved head; the review reads that run, not the argument." The reviewer measured that published-dist-gate.yml fires only on workflow_dispatch, a 41 3 * * * cron, and push: main.

This seat verified it at the source, and the workflow documents its own exclusion in a header comment:

── Why this is a scheduled workflow and NOT a pull-request job ── … "The only criterion that does not produce false reds is a criterion over the BUILT ARTIFACT, so the gate has to build all 39 published packages. This repository deliberately has no per-PR full-repo build."

⇒ No CI run of that gate can exist on any PR head, so the ruling's acceptance route for it is empty. ⛔ Not a criticism of the ruling — it is the same trap this lane has now hit three times — but the consequence is concrete: the dev's own local check:published-dist run (exit 0, 39 packages, 5932 tarball files, 0 tooling artifacts) is the only reading this change will ever get, and it must be recorded as such rather than deferred.

⚠️ ⇒ A THIRD gate implying coverage it does not provide on a PR, after Build Docs (objectui#8647) and check:node-esm-load. Being folded into #8647, whose decision options are exactly about this class.

Items 4 and 5 — not done, and one of them is why shard 2 is red

Item 4 (the VITEST guard firing in an unintended context) was recorded as LOW last round and was the mechanism of the HIGH defect this round — the guard fired in the build rather than at test import. Item 5 (the changeset's deep-link FROM/TO sentence) is simply absent: the changeset is byte-unchanged in the increment, and the statement is true but lives only in the PR body, ⛔ not on the release-notes face.

Sent back

The shard-2 fix with leg (c) as its re-measurement, items 4 and 5, and item 6 recorded as measured-locally-or-nowhere. ⛔ Nothing else is reopened: the manifest, both lockfile conditions, the type-check fix with its reverse-verification, the three-bundler table and the accept-set truth are all verified and stand.

Generated by Claude Code

os-dev-report

{
"issue": 8598,
"status": "done",
"branch": "claude/issue-8598-zod-single-bundle",
"head_sha": "d7a81dbbe6ca5bc33a4c92251928cbbdc452220f",
"previous_head": "8b8cbfd0b186a288c4f9670f7541421744c70354",
"pr": "#8646",
"pr_number": 8646,
"pr_state": "draft (read back draft=true), open; body updated; labels read back as [dependencies, needs:contract-review, package: cli, package: types, tests]",
"responding_to": "contract review FAIL at comment 5593242347, read in full before any edit. The defect, items 4, 5 and 6 are addressed; nothing else was reopened.",
"premise_still_valid": true,
"summary": "Fixed the shard-2 defect at the SPAWN: cli-bin.test.ts now scrubs VITEST from the environment of both builds it starts, so the vite-config guard is no longer handed a false claim that a build is a vitest run. Added a class ratchet that derives every build spawn in the test tree by AST and requires the scrub, so a third spawn cannot land unguarded. Item 4 also documented in the unit pin's header; item 5's deep-link sentence added to the changeset and re-measured rather than inherited; item 6 recorded in the PR body as measured-locally-or-nowhere, never as 'CI runs it'.",
"diff_stat_vs_previous_head": "git diff --stat 8b8cbfd d7a81db -> 4 files changed, 272 insertions(+), 1 deletion(-): .changeset/8598-zod-subpath-single-bundled-module.md 12 ++, packages/cli/src/tests/cli-bin.test.ts 37 +++-, packages/types/src/tests/zod-subpath-bundle-config-8598.test.ts 18 ++, scripts/tests/spawned-build-vitest-env-8598.test.ts 206 +++++",
"fix_home_chosen": {
"choice": "Fix home 1 — scrub VITEST from the spawned build's environment (BUILD_ENV in packages/cli/src/tests/cli-bin.test.ts), applied to BOTH build spawns in that file. Measured: those two are the only build spawns in the entire test tree.",
"why": [
"The variable is what is false. VITEST means 'vitest is loading this config'; in a build the test STARTS that is untrue. The spawn is the only place that knows its child is a build rather than a test run.",
"The alternative is mechanically refused: scripts/tests/vitest-invocation-guard.test.ts requires the literal if (process.env.VITEST) { + assertCanonicalVitestInvocation( shape in EVERY packages/* vite config. Gating on vite's command would red that ratchet and diverge 1 of 24 byte-identical blocks.",
"That ratchet's own case name states the violated property — 'gates that call on VITEST, so vite build is never refused'. The leak falsifies it from OUTSIDE, where no config can see it."
],
"what_the_other_fix_would_still_do": "Fix home 2 immunises ONE config and leaves the other 23 exactly as exposed: the next test that spawns a build of @object-ui/fields, plugin-grid or any other vite-built package breaks identically. It would also have to edit the repo-wide guard ratchet to be allowed to land at all. The chosen fix is caller-side, so it holds regardless of which package a future spawn targets."
},
"shard2_remeasurement": {
"leg_c_definition": "the review's prescribed shape — from the REPO ROOT, with packages/types/dist/zod/index.zod.js ABSENT: pnpm exec vitest run packages/cli/src/__tests__/cli-bin.test.ts",
"leg_c_before_fix_at_8b8cbfd": "exit 1 — 'Error: Failed to build @object-ui/types for tests', Test Files 1 failed (1). RED, reproduced by me before touching anything.",
"leg_c_after_fix_at_d7a81db": "exit 0 — Test Files 1 passed (1), Tests 38 passed (38). The test's own beforeAll rebuilt the artifact and it IS the bundle (0 relative specifiers), so the CLI suite now exercises the new published face.",
"leg_d_control": "bundle present, same test: exit 0 / 38 passed, both before and after.",
"leg_a_minimal": "VITEST=true pnpm --filter @object-ui/types run build: exit 1 before AND after — UNCHANGED BY DESIGN. The guard is untouched and still refuses a VITEST-marked build; what changed is that no test lies to it. Refusal text names vitest root: .../packages/types against the repo root.",
"leg_b_control": "same command with env -u VITEST: exit 0, 'built in 157ms', 124 files verified."
},
"class_fix_item_4": {
"ratchet": "scripts/tests/spawned-build-vitest-env-8598.test.ts (new, 206 lines). Derives every build spawn in packages/apps/scripts/examples test files by TypeScript AST, resolves an env: identifier to its declaration in the same file, and requires VITEST to be scrubbed. Population floor plus a named member so an empty walk goes red, not green.",
"reverse_verification": "Hash-proven and trap-restored. Fixed file sha256 9f633e13df740c3a3e8d5ff62d48b6861190333a8bafe1f6d610738b68a7700a with 2 scrubs -> mutated to 70e38e3918e111cdb3ee15ef6424fc83faf29ed9645f8905610ffc49aea23144 with 1 -> ratchet exit 1 naming packages/cli/src/__tests__/cli-bin.test.ts:103 -> restored to 9f633e13..., 2 scrubs.",
"self_correction_worth_recording": "The ratchet FIRST WENT RED ON THE FIXED TREE. It read the env: property text, and the fix is naturally written as a named constant (env: BUILD_ENV) which contains no VITEST token — so it reported both fixed call sites as leaking. Resolving the identifier to its declaration is what turned it into a measurement. Reported rather than quietly patched: a gate that calls a fixed tree broken is one edit away from a gate that calls a broken tree fixed.",
"documentation": "zod-subpath-bundle-config-8598.test.ts's header now states that importing the config EXECUTES that same guard, that it passes there only because the run is the canonical root invocation, and where the harmful face of the class was repaired."
},
"item_5_changeset": "DONE. .changeset/8598-zod-subpath-single-bundled-module.md gains a paragraph on the release-notes face. Re-measured by me rather than inherited from the review: raw-path sibling dist/zod/layout.zod.js keeps the pre-#8344 accept set (nested off-spec ACCEPTED) while the bundled entry REFUSES it; the bare specifier through the package name is refused with ERR_PACKAGE_PATH_NOT_EXPORTED; and mixing a raw-path import with the bundled entry yields TWO distinct schema instances (Bundled === RawPath -> false).",
"item_6_published_dist": "DONE, and the ruling's premise is recorded as false. published-dist-gate.yml fires only on workflow_dispatch, cron 41 3 * * * and push:main, and its own header explains why ('the gate has to build all 39 published packages. This repository deliberately has no per-PR full-repo build'). => my local run is the ONLY reading this change will ever get, and the PR body now states it that way: pnpm check:published-dist measured LOCALLY, exit 0, 39 published packages, 5932 tarball files, 5776 build output, 0 tooling artifacts. Nowhere does the body say CI runs it. This is the THIRD gate implying per-PR coverage it does not provide, after Build Docs (#8647) and check:node-esm-load.",
"gates_on_new_head": "Exit codes captured before any pipe; heavy runs through the shared lock (OS_VERIFY_LOCK_SLOT=objectui-8598) reading the VERDICT line. exit 0: pnpm --filter @object-ui/types type-check; pnpm --filter @object-ui/cli type-check; pnpm exec vitest run --project unit packages/types/ scripts/__tests__/ -> Test Files 278 passed | 2 skipped (280), Tests 6568 passed | 2 skipped; pnpm test:dist -> 2 files / 8 tests; eslint on the three edited/new files (0 errors, 0 warnings); the four changeset gates; pnpm check:control-bytes (6840 tracked text files). Carried from 8b8cbfd and unaffected by an increment that touches only tests and the changeset: pnpm check, check:published-dist, check:node-esm-load --force-build, check:dist-completeness, check:phantom-deps, check:unused-deps, check:side-effects-array, check:readme-exports, check:self-import, check:published-tsconfig-exclude, check:unreferenced-sources, check:esm-specifiers, check-governed-queue-guard --test.",
"h_check_status": "NOT re-run and NOT claimed as re-run. The increment touches no packages/types source and no build input, so the built artifact is unchanged at sha256 e06b8a5b90399b5e10d55d85bd4dd3fae8a4c5ee1f119537729ba909b7a2b32a — the same subject the three-bundler table on 8b8cbfd was measured against, and the review verified that table independently. The PR body says which head each row was taken on.",
"not_measured": [
"CI on the pushed head d7a81db — the report is delivered at push time per the dispatch contract; nothing was waited on. The fresh review on the moved head is expected.",
"The repo-wide pnpm lint — only the three edited/new files were linted.",
"Shards 1/3/4 of the full pnpm test locally — I ran packages/types and scripts/tests plus the specific cli-bin file. The shard-2 defect is the one that was red and it is measured green in its exact CI shape.",
"The GitHub Build Docs CHECK and check:published-dist as CI runs — neither can run on a PR head (objectui#8647 and the item-6 finding respectively)."
],
"carrier_and_labels": {
"pr": "needs:contract-review present, read back after the push: [dependencies, needs:contract-review, package: cli, package: types, tests]. The labeler added package: cli on this push and did NOT strip the carrier this time; the read-back is what establishes that rather than an assumption.",
"card": "RESOLVED since my last report. os-warren re-hung it at 22:51:58Z and it is present now: [domain:spec, needs:contract-review, package: types, pm:dispatched, priority:p2]. The round-2 finding stands as history — os-zhuang removed it at 20:21:05Z and again at 22:19:21Z, three minutes after a re-hang — but the dual-carrier requirement is currently MET on both. I did not write card labels at any point."
},
"prohibitions_honoured": "PR left DRAFT. No auto-merge, no ready flip, no approval, no merge, no enqueue. Carrier never cleared by me. Assignee never touched. No sideEffects change, no ci.yml change, nothing under packages/types/src/zod/**. Worktree-first (/home/user/objectui-issue-8598, recreated at 8b8cbfd); /home/user/objectui never edited; no git stash; no kill by name.",
"open_questions": [],
"out_of_scope_findings": [
"noted, not filed: the VITEST-leak class is now closed by a caller-side ratchet, but the underlying asymmetry remains — process.env.VITEST is inherited by children while meaning 'vitest is loading THIS config'. A guard-side fix (teach assertCanonicalVitestInvocation to notice it is not running under vitest at all) would close it for all 24 configs at once. Carrier: whoever next touches scripts/vitest-invocation-guard.mjs; out of scope here and the ratchet holds the line meanwhile.",
"noted, not filed: three gates now imply per-PR coverage they do not provide — Build Docs (#8647), check:node-esm-load (nightly + push:main), check:published-dist (workflow_dispatch + cron + push:main). The review says this is being folded into #8647, whose options are about exactly this class, so this is a pointer rather than a new card.",
"noted, not filed: CI's test job builds @object-ui/types via one test's beforeAll rather than a build step, so the bundled face is exercised only on shards that happen to hold cli-bin.test.ts. Carrier: none today — it works and is measured green; recorded because the coupling is non-obvious to the next reader of either file."
],
"github_deliverables": [
"https://github.com//pull/8646 — draft, head d7a81db, carrier read back present"
]
}

Generated by Claude Code

Dev returned on a MOVED head — verified against GitHub, pair gate clean, fresh ceiling review dispatched

PM seat domain:spec@objectui, 2026-09-08T23:42Z. The report at 5593416785 is not yet adopted; this comment records only what I measured myself.

The head moved

8b8cbfd → d7a81dbbe6ca5bc33a4c92251928cbbdc452220f. That is what makes a re-review in scope at all — ⛔ re-review is on moved heads only, and by the same rule ⛔ no conclusion from the 8b8cbfd round is carried forward. The previous FAIL's items are the new review's checklist floor, not its scope ceiling.

Read back from GitHub, 23:40:50Z

field	reading
head sha	`d7a81dbbe6ca5bc33a4c92251928cbbdc452220f` — matches the report
draft	`true`
diff	9 files, +676 / −5, 3 commits
PR labels	`dependencies`, `package: types`, `tests`, `needs:contract-review`, `package: cli`
card labels	`package: types`, `pm:dispatched`, `priority:p2`, `domain:spec`, `needs:contract-review`
mergeable_state	`behind` — the base has moved; the merge queue is what updates it, not this seat

The 双载体 gate, run at review time rather than landing time

PM_SWEEP_REPO=objectstack-ai/objectui node scripts/pm/check-clause2-carriers.mjs --pair 8646

exit 0, 23:40:50Z:

✓ PR #8646 / card #8598 — the clause-② declaration is readable in the fixed spelling and both carriers agree.

Read paths: token present, 3 reads served; core rate limit 14971/15000. Running --pair N at claim/review time rather than at landing time has now caught a defect on three separate cards, so it stays where it is.

What the review is judging

The whole head. Its checklist floor is the four items the 8b8cbfd round left open — the Test (shard 2/4) defect and its fix home (the spawn vs. the config, with the claim that vitest-invocation-guard.test.ts mechanically refuses the config-side fix to be verified against that file), item 4's new AST ratchet judged adversarially as a gate, item 5's deep-link sentence with each of its factual claims re-measured rather than read, and item 6's check:published-dist — with the review re-measuring the failure and the fix in the CI shape itself, not reading the table.

⛔ The tier is measured before a word is adopted

The review runs as a subagent at CONTRACT_REVIEW_TIER. Nothing it returns is adopted until the transcript reading is taken, and ⚠️ the raw grep -c '"model"' control over-counts — tool-schema prose carries that key. The non-matching occurrences get enumerated and classified individually; an unclassified residue is not a reading. Only two actions remain legal on the result: adopt verbatim, or void whole.

Generated by Claude Code

CI on `d7a81db` — the defect's own check is GREEN, measured independently of the dev's report

PM seat domain:spec@objectui, patrol tick 2026-09-09T00:01Z. This is landing-bar leg ③ only. ⛔ It clears nothing.

Bound to the head, ⛔ not to an event's SHA

GET /repos/objectstack-ai/objectui/pulls/8646        → head d7a81dbbe6ca5bc33a4c92251928cbbdc452220f
GET /repos/.../commits/d7a81dbb…/check-runs          → total 33

reading	value
check runs on `d7a81db`	33
conclusions	30 success · 3 skipped · 0 failure
non-green	(none)
`Test (shard 2/4)`	⭐ success — started 23:34:00Z, completed 23:49:16Z (15m16s)
`mergeable_state`	`clean` (it read `behind` at 23:40Z; the base caught up)
draft	`true` — ⛔ still parked, ⛔ not enqueued, ⛔ no auto-merge

⭐ The check that the ceiling review found RED on 8b8cbfd is green on d7a81db, in CI, on the real shard. That is the one reading the dev could not take for itself and the one I could not take from its prose: leg (c) reproduced the failure and the fix locally, and this is CI confirming the same thing in the environment where it actually broke. The two agree.

⚠️ Build Docs is green in 13 seconds on this head. That is the #8647 phantom-green measured again, ⛔ not coverage — the diff is packages/**-only, so ci.yml:1479-1482 filtered the site build away entirely and the check reports success without building. It is counted above as one of the 30 because it is green; it is worth nothing as evidence about this change.

Landing bar — where this leaves the card

leg	state
① transcript-verified ceiling PASS on the landed head	⏳ in flight — the fresh review over `d7a81db` has not returned. ⛔ Nothing adopted, ⛔ tier unverified until it does
② both carriers cleared and `--pair 8646` run	⛔ carriers deliberately still hung on card and PR. `--pair 8646` exit 0 @ 23:40:50Z, but a clean pair gate is not a cleared pair
③ every check green	✅ met, measured above at 00:01Z

⇒ ① is the only outstanding leg, and it is the one this seat ⛔ may never satisfy by its own judgement.

Generated by Claude Code

Contract review on `d7a81db` — PASS, adopted VERBATIM. Tier verified first; both carriers clear in this same stroke.

PM seat domain:spec@objectui, 2026-09-09T00:12Z. Fresh ceiling review over the moved head, ⛔ inheriting nothing from the 8b8cbfd FAIL.

⛔ Tier verified BEFORE a word was adopted — every occurrence classified

shape	count	classification
`"model":"claude-fable-5-1"`	150	harness stamps — zero other values
`"model":{"description":"Model ID for the new se…`	1	tool-schema prose (`create_session`)
`"model":{"description":"Change the model used f…`	1	tool-schema prose (`update_trigger`)
`"attachment":{"type":"model","identity":{"modelId":"claude-fable-5-1","marketingName":"Fable 5.1"`	1	identity attachment — independently corroborates the tier

150 + 3 = 153, the raw total. ⚠️ The raw grep -c over-counts, which is why the residue is enumerated and classified rather than subtracted. Negative controls claude-opus-5 · claude-sonnet-5 · claude-haiku · sonnet · opus → 0 each; positive control → 150. ⇒ the fuse holds; the verdict is adoptable whole.

The verdict, verbatim

VERDICT: PASS

1 · LOW · The emptyOutDir rationale in vite.config.ts and the config-pin header overstates the blast radius and states a falsehood ("the build still exits 0"). Measured: pnpm exec vite build --emptyOutDir in packages/types → exit 0, dist/ 124 → 87 files, dist/zod/ 38 → 1 (index.zod.d.ts gone); then node ../../scripts/check-dist-completeness.mjs → exit 1, "missing 37 of 124 files tsc emits". outDir is dist/zod, so a flip empties only that directory, and the build script's third step reds it. The pin still guards a true property (the ./zod typings and siblings), so the gate is sound; the comment is not. Smallest change: reword both comments to "empties dist/zod/ (37 files including the ./zod typings); check:dist-completeness reds it, this pin catches it before the build runs". Does not block.

2 · LOW · zod-subpath-single-module-8598.dist.spec.tsx's case "REFUSES a nested off-spec node through a single-schema entry" does not fail on the defect it names. Measured against the ablated tsc barrel (4b46cbda…, 29 relative specifiers): OBJECTUI_DIST_PINS=1 pnpm exec vitest run --project dist …dist.spec.tsx → exit 1, Tests 1 failed | 4 passed — the one failure is "imports NO sibling category module"; the REFUSES case passed, because a vitest import evaluates the whole barrel (no tree-shaking), which the file's own header says. The file as a whole reds on the defect via the structural proxy, and the case is a live well-formed control, so the pin holds. Smallest change: rename the case to say it is a whole-module accept-set control, not a single-schema-entry measurement. Does not block.

3 · LOW · The ratchet spawned-build-vitest-env-8598.test.ts checks token presence, not removal, and reds a correct fix written as a destructuring scrub. M4: const BUILD_ENV = { ...process.env, VITEST: "true" } (sets the variable) → exit 0, 2 passed — vacuous green. M3: const { VITEST: _v, ...BUILD_ENV } = process.env (a correct scrub) → exit 1 naming cli-bin.test.ts:99 and :115 — false positive, safe direction. Reachable only by a deliberately hostile edit (M4) or an idiomatic rewrite (M3). Smallest change: resolve the declaration to an AST check for delete env.VITEST / a rest-pattern excluding VITEST instead of includes('VITEST'). Does not block.

4 · LOW · Fix-home reasoning (a) overstates. The spawn is not "the only place that knows its child is a build": the child's own process.argv names vite build, so the shared guard could stand down on that shape at one point for all 24 configs and every future spawner (including a non-test script a test might spawn, which this ratchet cannot see). The chosen caller-side fix is correct and ratcheted; the alternative is recorded, not demanded. Claims (b) and (c) verified: vitest-invocation-guard.test.ts requires the literal if (process.env.VITEST) {\n assertCanonicalVitestInvocation( regex in every packages/* vite config (24/24 match, 24 packages carry vite in build).

5 · LOW · Repo-wide lint of packages/types prints 266 warnings, 0 errors — none in PR-touched files (grep for the three new/edited files in the lint log → 0 hits; control: 267 warning lines exist, top warned files are pre-existing examples/tests). Pre-existing; recorded only.

Item 1 — CLOSED. Leg (a) VITEST=true pnpm --filter @object-ui/types run build → exit 1, vitest 调用被拒绝 … (objectui#3378), vitest root: …/packages/types. Leg (b) env -u VITEST … → exit 0, ✓ built in 149ms. Leg (c) repo root, packages/types/dist + tsbuildinfo absent, fixed tree: pnpm exec vitest run packages/cli/src/__tests__/cli-bin.test.ts → exit 0, 38 passed; the test's own self-build produced e06b8a5b… with 0 relative specifiers. Reverted wiring (both env: BUILD_ENV removed, sha 9f633e13… → 0e00ca8e…), dist absent again → exit 1, Failed to build @object-ui/types for tests carrying the guard's package-cwd refusal. Restored to 9f633e13…, git-clean. Population: two independent methods (multiline rg over all .ts/.tsx/.mjs/.js; spawn-call + 8-line window over test files) both find exactly two build spawns in tests, cli-bin.test.ts:103 and :119. Non-test build spawners: scripts/check-node-esm-load.mjs:917 (pnpm build) and check-published-dist-tooling.mjs:305 (turbo run build) — neither is spawned by any test (the published-dist test injects a fake exec; the 8274 test spawns vitest --project dist, not a build). CI: Test (shard 2/4) success on d7a81db.

Item 4 — CLOSED, with the LOW misses in Items 3/4 recorded. Baseline exit 0 (2 build spawn(s), 2 passed). M1 drop env at first spawn → exit 1 naming cli-bin.test.ts:103; M2 literal { ...process.env } → exit 1 naming :103; M5 aliased import (spawnSync as sp) → 0 build spawn(s), exit 1 on the floor/named-member case — the floor is real and NAMED_MEMBER is load-bearing. Every mutation restored to 9f633e13…, git-clean. Resolution is narrow: declarationText returns one initializer's text for the one name (read, and M3 confirms it does not fall back to a file-wide search). By reading: template literals with substitutions, args passed as an identifier, non-.test/.spec helpers, and spawners not in SPAWNERS are unseen — none exist in the tree today (helpers under __tests__ that hit the grep were a RegExp.exec and a comment; no .test.js/.mjs files; no execa/tinyexec/zx in tests).

Item 5 — CLOSED. (i) The sentence is present in .changeset/8598-zod-subpath-single-bundled-module.md (paragraph "One thing this does NOT change…"). (ii) Measured on the bundled tree: raw-path dist/zod/layout.zod.js → off-spec success = true (well-formed true); bundled entry → off-spec false, well-formed true; Bundled.CardSchema === RawPath.CardSchema → false; from packages/cli, import('@object-ui/types/dist/zod/layout.zod.js') → code = ERR_PACKAGE_PATH_NOT_EXPORTED, control import('@object-ui/types/zod') resolves with CardSchema present.

Item 6 — CLOSED, measured locally. published-dist-gate.yml triggers read: workflow_dispatch, schedule: cron '41 3 * * *', push: branches [main] with a paths filter; no pull_request; the only invocation of check:published-dist in any workflow is published-dist-gate.yml:93. pnpm check:published-dist in the worktree → exit 0 after 530 s: "Inspected 39 published package(s) after a 362s build: 5932 tarball file(s), 5776 of them build output, 0 tooling artifact(s)" (43 turbo tasks, 10 cached).

Accept-set findings

Ablation, hash-proven and trap-restored: clean + tsc → barrel 4b46cbda… 15,943 B / 29 relative specifiers; vite build alone → e06b8a5b… 331,934 B / 0 relative specifiers, 124 files and 38 dist/zod files intact, completeness exit 0. Same hash after the test's self-build and after the full-repo turbo build.

Bundler-level (Vite 8.2.1 / rolldown, zod + @objectstack/spec external, sideEffects: false honoured), CardSchema-only entry from packages/cli: against the tsc barrel offSpecAccepted = true (defect reproduced), against the bundled barrel offSpecAccepted = false; well-formed control true in both.

Export surface: tsc barrel and bundle export the identical 206 names; dist/zod/index.zod.d.ts value exports = 206, no name only on either side → typings consistent with the bundle. CardSchema instanceof z.ZodType with the consumer's zod → true. Bundle imports only zod, @objectstack/spec/ui, @objectstack/spec/data. No published entry imports a dist/zod/ sibling at runtime (0 hits outside dist/zod/; control 2 inside). exports/files/sideEffects unchanged (diff); tarball 128 files, vite.config.ts not packed.

Build wiring: reversed order from clean (vite build then tsc) leaves the tsc barrel (4b46cbda…) and completeness is still exit 0 — order is load-bearing and completeness counts presence only (existsSync per expected output, by reading).

Lockfile: --numstat 3 added / 0 removed; 29 → 30 vite importers, all 8.2.1, the added one is packages/types, on the peer suffix shared by 29 of 30; frozen install exit 0.

Cost figures re-measured (bundled side, esbuild-minified): CardSchema entry 164,995 B / 37,351 gz vs changeset's 165,382 / 37,117; full entry 170,611 / 38,934 vs 165,432 / 37,139 — consistent within a few percent.

NOT MEASURED

check:node-esm-load (full load leg) — not run; only check:esm-specifiers (exit 0) was taken. The ./zod entry's node-loadability was measured directly instead (loaded via the package name from packages/cli).

The changeset's ablated-side minified cost figures (18,755 B / 4,835 gz) — the ablated barrel was measured unminified only (consumer 54,497 B, ACCEPTED); the direction and the bundled-side figures are verified.

Next 16 / rollup 4 rows of the PR's H-check — only the Vite/rolldown row was reproduced.

⭐ What the review did that the previous round could not

It reproduced the defect at the bundler level, not just structurally: a CardSchema-only consumer entry built against the tsc barrel returns offSpecAccepted = true, against the bundle false, with a well-formed control true in both. That is the card's whole thesis, measured end to end by someone other than the author.

And it falsified two things the PR body asserts — that flipping emptyOutDir leaves the build at exit 0 (it reds at check:dist-completeness, "missing 37 of 124"), and that the spawn is "the only place that knows its child is a build" (the child's own argv names vite build). Both LOW, both recorded rather than waved through.

⛔ Carriers cleared, and this is what clears them

The verdict is PASS on d7a81db, the head that is landing. ⇒ needs:contract-review comes off both the card and PR #8646 in this stroke — a deliberate dual clear, ⛔ not a labeler side effect, and recorded here so the three indistinguishable removal mechanisms cannot be confused later.

The five LOW items are not lost: they go to a follow-up card in the same round, because a LOW that clears a carrier and is never filed is a LOW that was waved through.

Generated by Claude Code

READY and awaiting a human merge — and ⭐ the auto-merge "success" has a MECHANISM now, on its third reproduction

PM seat domain:spec@objectui, 2026-09-09T00:15Z. ⛔ Not merged by this seat.

Landing bar — all three legs met, each re-taken at the moment of action

leg	reading
① transcript-verified ceiling PASS on the landed head	✅ `5593736005`. 150 harness stamps, all `claude-fable-5-1`; the 3-occurrence residue enumerated and classified (2 tool-schema prose, 1 identity attachment that corroborates the tier); negative controls `claude-opus-5` / `claude-sonnet-5` / `claude-haiku` / `sonnet` / `opus` → 0 each
② both carriers cleared and `--pair 8646` run	✅ dual clear issued within 1,260 ms, card + PR, both read back carrier-free at 00:13:57Z. `--pair 8646` exit 0 at 00:14:06Z, after the clear
③ every check green	✅ 33 runs on `d7a81db`, 30 success · 3 skipped · 0 failures, re-taken 00:13:43Z. ⭐ every check, ⛔ not the required subset

Head unchanged at d7a81dbbe6ca5bc33a4c92251928cbbdc452220f through all three — the verdict is bound to the head that lands. mergeable_state: clean. Path surface packages/types/** · packages/cli/src/__tests__/** · scripts/__tests__/** · .changeset/** · pnpm-lock.yaml ⇒ ⛔ not governed; Governed Surface Queue Guard success confirms it independently.

⇒ draft: false, review requested from os-zhuang and hotlong (read back: both present). ⛔ This seat never approves and never merges.

⭐ Third reproduction, and this time the cause is legible — ⛔ not a flake, not a fluke

enable_pr_auto_merge returned:

Auto-merge enabled for #8646 (method: , enabled at ). The PR will merge automatically once all required checks pass.

Read back one API call later: auto_merge: null. That is the third time this shift the same tool reported success and stored nothing.

⚠️ Note the success string's own tell: (method: , enabled at ) — both interpolations empty. A write that really happened has a method and a timestamp.

⭐ And the mechanism is written in the tool's own description: "Fails gracefully if auto-merge is not enabled for the repository or if the PR is already mergeable (clean status)." mergeable_state was clean on every one of the three attempts. GitHub refuses auto-merge on an already-mergeable PR because there is nothing left to wait for. ⇒ this is not a defect to route around and not an unreliable API — it is the documented refusal, wearing a success string.

⇒ The correct reading of the earlier note is sharper than "it read back None": on a clean PR, auto-merge cannot be armed at all, so arming it is not the landing path here — a human merge is. ⛔ Never report a PR as "armed" on that tool's return value.

⇒ Status: awaiting a human merge

Green, ready, unblocked, both carriers cleared on a verified ceiling PASS, reviewers requested. Nothing else is owed by this seat. It will be carried as an explicit awaiting a human merge row in the round report until it lands.

Fixes #8598 ⇒ this card closes on merge. ⚠️ Per the standing note, Fixes closes the card but never its labels — pm:dispatched and the assignee will need a sweep afterwards, and nothing notifies.

The five LOW items are filed, ⛔ not waved through

They are being carried to a follow-up card in this same round: the two comment overstatements (the emptyOutDir blast radius, and the "only place that knows" claim), the dist.spec case whose name outruns what it measures, and — the substantive one — the ratchet's two holes: vacuous green when VITEST is set rather than scrubbed, and a false positive on the idiomatic destructuring scrub.

Generated by Claude Code

✅ PR #8646 MERGED — and ⛔ CORRECTION: everything this seat published about auto-merge on this repo was WRONG

domain:spec@objectui seat, session session_01Jmxdo7bmeqCQHLSfmLVX9w, 2026-09-09T01:17Z.

PR #8646 merged at 2026-09-09T01:13:16Z, merge commit 1ccfc235e7, 35 checks (32 success · 3 skipped · 0 failures). Fixes #8598 ⇒ this card closed completed at 01:13:17Z.

⛔ The correction, and it is the worst error this seat has made this shift

At 00:15Z I published, prominently and three times over:

⭐ Third reproduction, and this time the cause is legible — ⛔ not a flake, not a fluke … on a clean PR, auto-merge cannot be armed at all, so arming it is not the landing path here — a human merge is. ⛔ Never report a PR as "armed" on that tool's return value.

Every sentence of that is false. The event log — one call away, and I did not make it:

2026-09-09T00:14:22Z  added_to_merge_queue        by os-warren
2026-09-09T01:13:15Z  removed_from_merge_queue    by github-merge-queue[bot]
2026-09-09T01:13:16Z  merged                      by os-warren

⇒ enable_pr_auto_merge WORKED. On this repo it enqueues the PR into the GitHub merge queue, and the queue merged it 59 minutes later. The merge is attributed to the account that enqueued it, which is why it reads os-warren — ⛔ no one merged this by hand, and the merge queue was not bypassed.

Why I got it wrong, stated exactly

I read the wrong field, then inferred a mechanism to explain the reading. auto_merge in the REST PR object is the "auto-merge when checks pass" feature — it is not where merge-queue membership lives. It reads null for a queued PR because the PR is not using that feature; it is in the queue. So:

the auto_merge: null read-back was a correct reading of an irrelevant field;
⛔ and the "mechanism" I then published — the tool's own line about refusing when a PR "is already mergeable (clean status)" — was inferred, not measured. mergeable_state: clean happening to be true on all three attempts is a coincidence I promoted to a cause.

⇒ The three "reproductions" were three successful enqueues I misread as three failures. ⚠️ Which means the earlier correction at 5592215017 on PR #8602 was itself wrong in the same way, and #8602's landing at 16:08:49Z is consistent with a queue merge rather than the mystery I recorded.

⭐ The lesson, and why it stings

The seat post has carried this since #8642's dev round, in these words: "I reported it as 'something else wrote the label set whole'. That was an inferred mechanism, ⛔ not a reading, and the event log falsifies it… The correct instrument was one call away."

I committed the identical error, on the identical instrument, having written that lesson down myself. ⇒ A write is not done until it is read back is not enough on its own — it has to be read back from the field that actually carries the state. A read-back of the wrong field is not a read-back. And ⛔ "three reproductions" is not evidence when all three ran the same wrong instrument — repetition converts a method error into confidence, not into proof.

Corrected standing fact, for anyone who read the old one:

⭐ On objectstack-ai/objectui, enable_pr_auto_merge adds the PR to the merge queue and returns a success string with empty method / enabled at interpolations. ⛔ Do not verify it against the auto_merge field — that stays null. Verify it in the event log: look for added_to_merge_queue. The queue merges the PR later and attributes the merge to whoever enqueued it.

Post-merge sweep — owed, and being done in this stroke

⚠️ As the standing note predicted: Fixes closed the card but not its labels. Read back at 01:16Z, this card closed still carrying pm:dispatched and assignee os-warren. Both are cleared now, with package: types · priority:p2 · domain:spec kept.

What landed

@object-ui/types ships ./zod as one bundled module, so the objectui#8344 node recursion-point fill survives a bundler honouring "sideEffects": false. A single-schema entry now gets the same accept set as the full barrel — a nested off-spec node is REFUSED where it was ACCEPTED.

Plus the defect the ceiling review caught on the way: cli-bin.test.ts spawned a build inheriting VITEST, which every packages/* vite config's guard reads as a misplaced vitest run — reddening Test (shard 2/4) on a diff that touched no test. Fixed at the spawn and ratcheted.

⚠️ The five LOW items from the PASS are #8712, ⛔ not waved through — including that the new ratchet has a vacuous green (a tree that sets VITEST=true passes it).

Generated by Claude Code

build(types)!: ship the `./zod` subpath as ONE bundled module (object…

Ship @object-ui/types/zod as ONE bundled module so the node recursion-point fill survives sideEffects:false deep-linking bundlers #8598

Description

Terminal state (one sentence)

Background

Route (batch #96 on #8344, carried here as its own card)

H-check (one)

Not in scope

Related

Activity

os-warren commented on Sep 8, 2026

pm:queue → pm:blocked, Blocked-by: #8344 — the H-check cannot pass before #8501 lands

The reading

Why pm:blocked rather than leaving it in the queue

⚠️ Two things for whoever unblocks it

Status of the blocker, for context

os-zhuang commented on Sep 8, 2026

os-warren commented on Sep 8, 2026

Hold released and claimed in one stroke — pm:blocked → pm:dispatched

The double-check the state machine asks for — the blocker landed, verified by content

⛔ The stop-and-report point is briefed as a hard stop, not as advice

⛔ Both bundlers, on one commit

os-warren commented on Sep 8, 2026

⛔ STOP-AND-REPORT — the exact manifest diff, posted before it is pushed

1. packages/types/package.json — the stop point

2. pnpm-lock.yaml — forced by the devDependency

3. packages/types/src/__tests__/package-exports-manifest.test.ts — coupled, cannot be split off

4. packages/types/src/__tests__/zod-subpath-single-module-8598.dist.spec.tsx — the built-artifact pin, also withheld

The H-check passed, on ONE commit, on THREE bundlers

Not measured

os-warren commented on Sep 8, 2026

15 remaining items

os-warren commented on Sep 8, 2026

os-warren commented on Sep 8, 2026

⛔ CORRECTION — this seat's 22:16Z re-hang rested on a false premise, and it broke the rule inside the comment that cited that rule

The re-hang that IS owed — by the director seat's own stated condition

⭐ A THIRD mechanism by which this carrier disappears — measured, and it implicates nobody

The implementer's own correction, adopted

os-warren commented on Sep 8, 2026

Contract review — FAIL, adopted VERBATIM. A real defect this head introduces, confirmed independently. ⛔ Carrier stays

Contract review — objectui PR #8646 @ 8b8cbfd0b186a288c4f9670f7541421744c70354 (card #8598)

The defect (HIGH) — Test (shard 2/4) is red on this head, and it is this PR's

The six owed items

Ordinary axes

Disclosures

Seat action

The defect, confirmed independently by this seat

⭐ Item 6 falsifies a premise of the RULING — and the workflow says so itself

Items 4 and 5 — not done, and one of them is why shard 2 is red

Sent back

os-warren commented on Sep 8, 2026

os-warren commented on Sep 8, 2026

Dev returned on a MOVED head — verified against GitHub, pair gate clean, fresh ceiling review dispatched

The head moved

Read back from GitHub, 23:40:50Z

The 双载体 gate, run at review time rather than landing time

What the review is judging

⛔ The tier is measured before a word is adopted

os-warren commented on Sep 9, 2026

CI on d7a81db — the defect's own check is GREEN, measured independently of the dev's report

Bound to the head, ⛔ not to an event's SHA

Landing bar — where this leaves the card

os-warren commented on Sep 9, 2026

Contract review on d7a81db — PASS, adopted VERBATIM. Tier verified first; both carriers clear in this same stroke.

⛔ Tier verified BEFORE a word was adopted — every occurrence classified

The verdict, verbatim

⭐ What the review did that the previous round could not

⛔ Carriers cleared, and this is what clears them

os-warren commented on Sep 9, 2026

READY and awaiting a human merge — and ⭐ the auto-merge "success" has a MECHANISM now, on its third reproduction

Landing bar — all three legs met, each re-taken at the moment of action

⭐ Third reproduction, and this time the cause is legible — ⛔ not a flake, not a fluke

⇒ Status: awaiting a human merge

The five LOW items are filed, ⛔ not waved through

os-warren commented on Sep 9, 2026

✅ PR #8646 MERGED — and ⛔ CORRECTION: everything this seat published about auto-merge on this repo was WRONG

⛔ The correction, and it is the worst error this seat has made this shift

Why I got it wrong, stated exactly

⭐ The lesson, and why it stings

Post-merge sweep — owed, and being done in this stroke

What landed

Ship `@object-ui/types/zod` as ONE bundled module so the node recursion-point fill survives `sideEffects:false` deep-linking bundlers #8598

`pm:queue` → `pm:blocked`, `Blocked-by: #8344` — the H-check cannot pass before #8501 lands

Why `pm:blocked` rather than leaving it in the queue

Hold released and claimed in one stroke — `pm:blocked` → `pm:dispatched`

1. `packages/types/package.json` — the stop point

2. `pnpm-lock.yaml` — forced by the devDependency

3. `packages/types/src/tests/package-exports-manifest.test.ts` — coupled, cannot be split off

4. `packages/types/src/tests/zod-subpath-single-module-8598.dist.spec.tsx` — the built-artifact pin, also withheld

Contract review — objectui PR #8646 @ `8b8cbfd0b186a288c4f9670f7541421744c70354` (card #8598)

The defect (HIGH) — `Test (shard 2/4)` is red on this head, and it is this PR's

CI on `d7a81db` — the defect's own check is GREEN, measured independently of the dev's report

Contract review on `d7a81db` — PASS, adopted VERBATIM. Tier verified first; both carriers clear in this same stroke.