spec(types): redirect the node recursion point from BaseSchemaCore to AnyComponentSchema — measured at 9 newly-refused documents, and it drops 118 phantom strict refusals

[claude]

Priced by the domain:spec @ objectui seat on #7935 (the pricing card the #5250 ruling names as its step 2). ⛔ Filed unassigned, not claiming. Appetite is in this body and is binding: the maintainer approved the direction, not a blank cheque on scope (#5250, director comment 5534418546).

This is the FIRST card of the strict-authoring-face programme, and it is not the strict face. It is the structural precondition #7581's measurement identified: until the node recursion point resolves per-type, every document-level number the programme quotes is measuring the recursion point instead of the components.

What is true today

packages/types/src/zod/base.zod.ts:53:

export const SchemaNodeSchema = z.lazy(() =>
  z.union([BaseSchemaCore, z.string(), z.number(), z.boolean(), z.null(), z.undefined()])
);

(The export additionally carries an explicit z.ZodType annotation taking an any type argument — the recursion-breaking annotation objectui#7760 is in flight to replace. Elided above only because a type-argument bracket does not survive GitHub's body sanitizer; the union arms are the part this card is about.)

BaseSchemaCore is the ~21 base keys and nothing type-specific. Every child slot (body, children, and every per-component redeclaration of them) is z.union([SchemaNodeSchema, z.array(SchemaNodeSchema)]). ⇒ per-type enforcement is root-only, at every depth, for every component type. That is the finding already recorded as #7869, measured independently there (an off-spec size on a nested icon node is accepted; the same node alone is refused).

The measurement this card is priced from

Measured on origin/main 5505aec1a07c26976d160a39f152a48b088cfbb4, zod 4.4.3, face read from packages/types/dist/zod/index.zod.js, over the catalog + docs corpora — 553 node documents (431 catalog JSON files, 204 ```json fence openings). Four twins, one corpus, one commit:

reading	recursion point	strict?	documents refused
R1 baseline — the face as shipped	BaseSchemaCore	no	45 / 553
R2 naive strict — what a .strict() flip really does today	BaseSchemaCore	yes	294 / 553
R3 this card — redirect only	AnyComponentSchema	no	54 / 553
R4 redirect + strict — where the programme is going	AnyComponentSchema	yes	176 / 553

Controls: R1 and R2 reproduce scripts/measure-strict-authoring-face.mjs at the same commit (46 / 555 and 296 / 555) — the 2-document and 1-refusal difference is corpus 3, the 2 authored documents this reading omits.

⭐ The two results that decide the ordering

1. The redirect costs 9 documents. R3 minus R1 = 54 minus 45. Every one is enumerated below.
2. The redirect removes 118 documents of phantom redness. R2 minus R4 = 294 minus 176. Those 118 are refused today only because the recursion point unrecognises every child's own declared props — no per-component declaration budget can ever close them. ⇒ pricing the strict face on 294 rather than 176 prices the wrong programme.

The 9 documents, and why each refuses — both classes are pre-existing debt

Class A — a child node whose type resolves in no arm of AnyComponentSchema (4 documents, 7 spellings):

• examples/schema-catalog/src/schemas/components-form-calendar/form-integration.json — child ui:calendar
• examples/schema-catalog/src/schemas/components-layout-semantic/complete-layout.json — children header, nav, main, footer
• content/docs/utilities/runner.mdx#L269 — child my-component
• content/docs/utilities/runner.mdx#L323 — child h1

Class B — a child node already red under its OWN schema today, shielded until now by the recursion point (5 documents):

• examples/schema-catalog/src/schemas/components-complex-filter-builder/search-interface.json — filter-builder: field invalid_type, operator / logic invalid_value
• examples/schema-catalog/src/schemas/components-feedback-toaster/with-toast-trigger.json — button: onClick custom
• content/docs/api/schema-reference.md#L173 — badge: variant invalid_value
• content/docs/api/schema-reference.md#L1267 — dashboard: type invalid_value + unrecognized_keys
• content/docs/guide/schema-playground.md#L98 — stack: direction invalid_value, gap invalid_type

⇒ The redirect creates no new defect class. It surfaces two existing ones, at a size that fits in one card.

The cost driver that is NOT in the document count — the module cycle

SchemaNodeSchema is declared in base.zod.ts, which today imports only tombstone.zod.js and expression.zod.js. AnyComponentSchema is built in index.zod.ts from all 13 category modules, and 14 modules import base.zod.js. ⇒ pointing the recursion point at the union makes base.zod.ts depend on index.zod.ts, which depends transitively back on base.zod.ts. z.lazy defers the evaluation, not the module graph, so this needs a deliberate break — a late-binding holder that index.zod.ts fills at load, or relocating the union. ⛔ Do not discover this at implementation time; it is the reason this card is not a one-line diff.

And one hard constraint, measured, that a naive implementation will get wrong

The recursion point must be recognised by identity on the exported wrapper, ⛔ never through .unwrap(). Measured at the same commit:

exported wrapper identity stable   : true    (two module reads return the same object)
CardSchema.body union contains S   : true    (identity survives through a declared slot)
S.unwrap() === S.unwrap()          : false
getter() === getter()              : false
_zod.innerType stable              : true    (zod 4.4.3 caches the resolved inner)

This is #7918's consequence ①, relocated by os-sam on #7935 (comment 5563938784): the public handle is unstable for eight of the ten lazy exports, and SchemaNodeSchema is one of them. Identity on the export works today and is what scripts/measure-strict-authoring-face.mjs already relies on.

Appetite — one dispatched card, one PR

In scope: the redirect, the module-cycle break, the 9 documents above, and a pin that a nested off-spec node is now refused (the #7869 reproduction, both directions).

⛔ Falls off the back rather than growing this card:

• Any of the 7 Class-A spellings that turns out to need a new component arm — that is a public-surface widening, and it splits into its own card rather than riding this one.
• Everything strict. This card does not add .strict() anywhere; R3, not R4, is its acceptance number.
• The declaration-repair worklist (the 184-node / 2099 per-component reading). Not this card.

If the 9 do not fit, the card is wrong and comes back for re-pricing — ⛔ do not widen it in flight.

Grading notes

• Clause-② yes. This changes contract accept/reject behaviour at every depth below the root. finding(types): safeValidateSchema enforces per-type component schemas on the ROOT node only — every nested child is matched against the passthrough BaseSchema, so a DECLARED key with a wrong-typed value passes at any depth #7869's triage (comment 5555946183) already reached that verdict and asked for a census before the tightening could be priced: "先立一张 report-only 普查卡,量出「若逐类型 union 下沉到每一层,今天会新增多少个拒绝、分布在哪些 fixture 与文档语料里」。有了那个数字,收紧才是一次可定价的改动而不是一次赌博。" ⇒ That census is the table above; this card is the priced change it asked for.
• ⚠️ File-surface collision: finding(types): the recursion-breaking annotation on 10 zod mirrors erases their published input face to unknown — the single-or-list producer objectui#7069 named systematic is unauditable at the type level because of it #7760 (dispatched, in flight) touches the same ten recursive mirrors to give them an explicit INPUT type argument. Sequence after it or coordinate; ⛔ do not dispatch both onto base.zod.ts at once.
• Discharges the finding finding(types): safeValidateSchema enforces per-type component schemas on the ROOT node only — every nested child is matched against the passthrough BaseSchema, so a DECLARED key with a wrong-typed value passes at any depth #7869 (recording-only; no mechanism proposed there).

Refs: #5250 (the ruling) · #7935 (this pricing) · #7581 / PR #7916 (the measurement) · #7869 (the finding) · #7760 (in flight, same files) · #7918 (the unstable public handle) · #6318

---

Generated by Claude Code

[os-sales]

Claim: PM loop round 1
Session: `session_01CZY49skxUBYyJcdnTcYPrE`
Branch: `claude/issue-8344-node-recursion-point-redirect`
Worktree: `objectui-issue-8344`
Domain: `domain:spec`
File surface: `packages/types/src/zod/base.zod.ts` (the `SchemaNodeSchema` union arms + whatever the module-cycle break requires there) · `packages/types/src/zod/index.zod.ts` (`AnyComponentSchema` and the late-binding fill) · the 9 enumerated corpus documents under `examples/schema-catalog/src/schemas/**` and `content/docs/**` · a new pin under `packages/types/src/__tests__/` · `.changeset/`. ⛔ `packages/types/src/__tests__/zod-mirror-parity.test.ts` is OFF LIMITS — held this round by #7762 (see below). ⛔ `content/docs/releases/**` never. (stop on breach; explain in the report)
Container & model: `M`, `mode:subagent`, `model: TIER_DEFAULT (opus)` — 主动预降 from `CONTRACT_REVIEW_TIER` under the standing maintainer instruction 「fable不多了，后续推 opus 任务」; compensating control = `needs:contract-review` hung on this card in this same stroke and on the draft PR when it exists, cleared only by a `CONTRACT_REVIEW_TIER` reviewer.
`Clause-②`: yes
Thread-read: none (the card carries zero comments at claim time)
Serial constraints cleared: ⭐ **the card's own stated collision is DISCHARGED.** Its grading note reads 「#7760 (dispatched, in flight) touches the same ten recursive mirrors … ⛔ do not dispatch both onto `base.zod.ts` at once」. **#7760 is CLOSED `completed`, by PR #8354 MERGED 2026-09-08T00:48:57Z** — verified on the card, not on a merge notice. Lane re-read 01:0xZ: 0 parked, 0 enqueued, 0 other in flight from this seat; the only other `pm:dispatched` card in `domain:spec` is **#6939** (`os-project-manager`, a `domain:ui`-seat dispatch), ⛔ untouched. ⚠️ **This seat's own #7762 went out minutes earlier** and holds `packages/types/src/zod/objectql.zod.ts` + `packages/types/src/__tests__/zod-mirror-parity.test.ts` — **file-disjoint from this card**, and an explicit stop-and-report is written into both briefs if either needs the other's file. Heavy-verify lock read at 01:1xZ: `state: lock is free`, `queue: empty` ⇒ arriving depth **1**, under `LOCK_DEPTH_HOLD`.

---

Why this is dispatchable now, and what the dispatch does NOT change

The card was priced complete on #7935 and its appetite is binding, stated in its own body. ⛔ This claim does not re-price it, does not widen it, and does not add anything strict — R3 (54 / 553), not R4, is the acceptance number.

The single thing that held it was the file collision with #7760 on base.zod.ts. That card closed and its PR merged at 00:48:57Z, so the ten recursive mirrors are settled on main and this card can take base.zod.ts cleanly.

⚠️ One consequence of #8354 the card could not have known, flagged as a PM mechanism assumption for the dev to measure rather than trust: #8354 filled seven of the ten z.lazy mirrors with explicit type arguments and moved the parity ledger's SCHEMA-NODE reading substantially (its verdict recorded WiderThanDeclared 34 → 23 entries, 52 → 36 keys, SCHEMA-NODE keys 24 → 6). SchemaNodeSchema itself is one of the seven that was filled. ⇒ the card's quoted base.zod.ts:53 shape and its parenthetical about the recursion-breaking annotation are pre-#8354 readings. Re-derive the current declaration by symbol before touching it; if the redirect now interacts with the filled type argument, that is a finding to report, ⛔ not a licence to widen.

⚠️ If the redirect moves a row in zod-mirror-parity.test.ts, stop and report — that file belongs to #7762 this round. Do not edit it, and do not work around it.

The card's two hard technical constraints stand and should not be rediscovered: the module cycle (base.zod.ts would come to depend on index.zod.ts; z.lazy defers evaluation, not the module graph, so this needs a deliberate late-binding break) and identity on the exported wrapper (⛔ never .unwrap(), whose result is measured non-stable).

---

Generated by Claude Code