ArtifactPackageEntrySchema's body half cannot describe the payload the load path actually registers

[hotlong]

Found while implementing #14162 (PR #14240). Filed rather than worked around there: widening a spec surface is not a loader's decision.

The fact

ArtifactPackageEntrySchema (packages/spec/src/stack.zod.ts, landed by #14161) wraps the package body as manifest: ManifestSchema. ManifestSchema.objects is z.array(z.string()) — glob patterns, the authoring-time shape.

What actually reaches the ADR-0130 load path is an assembled payload whose objects are object definitions: AppPlugin flattens an artifact into { ...bundle.manifest, ...bundle } before calling manifest.register(), and ObjectQL.registerApp iterates those bodies as definitions.

So a full parse of a real artifact entry refuses it. Measured on cfa99f3d against the landed schema:

ArtifactPackageEntrySchema.safeParse({ manifest: assembledPackageBody })
 → success: false
 → manifest.objects.0: Invalid input: expected string, received object

The entry schema therefore describes the authoring-time manifest, while the key it was added for (packages[] on ObjectStackDefinitionSchema) is read at load time, where the body has a different shape. One schema is being asked to describe two different lifecycle stages of the same noun.

Why it matters

Any consumer that reaches for the obvious ArtifactPackageEntrySchema.safeParse(entry) as its gate will refuse exactly the artifacts the path exists to register — and it will look correct while doing so, because the schema is the surface's one declaration and the refusal names a real key.

PR #14240 works within it deliberately and narrowly: it applies the schema as a gate on the wrapper only (issues at the entry root or on manifest itself), and hands registerApp the caller's original body rather than a parsed clone. That is written down in the module header with this measurement, so the next reader does not discover it again. But it is a local accommodation of a surface-level mismatch, and the accommodation is the kind of thing that quietly becomes convention if the surface is never revisited.

What the decision is

Not obvious, hence a card rather than a patch. Roughly:

• A. Leave the schema authoring-shaped and make "the wrapper is the only part this schema gates at load time" an explicit, documented property of the surface — the status quo PR ADR-0130 D5+D7: register an artifact's N packages in topological order, reusing resolvePluginOrder #14240 assumes, but stated on the schema instead of in one consumer.
• B. Give the assembled/registered form its own declaration, so both stages are describable and a load-time consumer has something correct to parse against.
• C. Widen ManifestSchema.objects to accept both spellings — cheapest to write, and the one that most deserves scrutiny: a union that accepts both stages makes neither stage checkable, which is the tolerance-at-the-consumer shape this repo generally refuses.

No packages[]-shaped artifact is produced anywhere yet, so nothing is broken in the field today and there is time to decide this properly rather than under a deadline.

[hotlong]

决定：取 B —— 装配 / 注册态的包体有自己的声明

维护者 2026-09-02 在 PM 会话（session_01UHvF5hyiZjnCyExFnfQB8m）对"立多包编译路径卡、顺带把 #14242 按 B 方向收口"答复「同意」。据此：

• B：为装配后的包体给独立 schema；发布物层 packages[] 引用它；装载门从 ADR-0130 D5+D7: register an artifact's N packages in topological order, reusing resolvePluginOrder #14240 的 wrapper-only 改为对装配态声明完整 parse。
• 否决 C：ManifestSchema.objects 不加宽——接受两个阶段的 union 会让两个阶段都不可校验，正是本卡自己指出的形状。
• A 作为过渡状态随实现终止。

实施与关闭在 #14439（Fixes #14242 与该卡同一 PR）。摘除 needs-user-decision，保留 priority:p2 / domain:spec。

---

Generated by Claude Code