@@ -0,0 +1,18 @@
+---
+"@objectstack/spec": minor
+---
+
+`composeStacks(…, { manifest: 'preserve' })` emits each definition **once**, in the body of the package that owns it: a multi-package release artifact no longer carries a flattened copy of its collections at the top level (#14512, ADR-0130 D4's 2026-09-22 addendum).
+
+Measured on `examples/app-multi-package` (two packages, three definitions): `dist/objectstack.json` goes from 8,223 to 5,328 bytes, −35%, and the ratio does not improve with size — it was one extra copy of everything a package owns. The artifact's own `manifest`, `packages`, `plugins`, `devPlugins`, `devLogins`, `api`, `server`, `i18n`, `runtimeModule`, `onEnable` and `devHint` are untouched at the top level; the 35 package-owned collections are what moves.
+
+**BREAKING** for anything that read a compiled multi-package artifact's top-level collections directly. Every reader the platform ships was converted first — the ruling's order was readers first, emitter last — and this lands only with #15004's acceptance probe green on both shapes, which boots a two-package collection zoo through all five load boundaries and fails if any subsystem sees an empty collection.
+
+- **No authored metadata changes, and no artifact on disk is invalidated.** The authoring shape is identical: N `defineStack` packages plus a project config composing them with `manifest: 'preserve'`. ADR-0130 D4's read-both rule is untouched — `packages` present ⇒ iterate it, `packages` absent ⇒ `manifest` as a single-element list — so an artifact built before this change, which carries both halves, loads exactly as it did.
+- **A single-package artifact keeps today's shape byte for byte** (ADR-0130 D7): the emitter strips nothing below two package entries.
+- **The flattened half is dropped only where it is a COPY — three conditions, and a composition failing any one of them keeps today's additive shape**: (1) two or more package entries; (2) every input's collections attributed to a body — an input declaring no `manifest` has no package that could own its collections, and an input already carrying `packages` contributes those entries untouched; (3) the bodies reproduce the flattened collections item for item. Condition 3 is what `objectConflict: 'merge'` / `'override'` and a standalone action bound onto a SIBLING package's object fail: composition RECONCILES those into a top-level definition no body carries, so the flattened half is not a copy of anything and stays. Nothing here is refused — a composition that was legal before stays legal.
+- **Why the copy goes rather than being compressed**: one definition was serialized twice with nothing keeping the copies equal. Where they differ today the flattened one is the reconciled copy and the platform's reader prefers it deterministically, which is exactly why this change removes the second copy only where the first one is redundant. One measured consequence on the compiled path: a seed dataset declared once reached the runtime's shared seed registry twice, and now reaches it once.
+
+Clause-②: yes (narrowing)
+
+<!-- adr-0087: not-required (no-migration-prescription) Nothing an author writes moves: no spec key, export or config field is removed or renamed, and no stored metadata representation changes, so `objectstack migrate meta` has nothing to rewrite. What changed is the SHAPE A COMPILER EMITS, and every reader of it in this repository was converted by the reader half of the same ruling (#15005, #15006, #15007, #15229, #15232) before this landed. -->
@@ -411,9 +411,10 @@ export default defineStack({
 
 `os build` compiles that project into **one** `dist/objectstack.json` carrying a
 `packages[]` list — one entry per package, each holding that package's own
-manifest fields and the metadata it owns. `os dev` boots the same shape straight
-from source, and `GET /api/v1/packages` on a running instance lists every
-package in the artifact.
+manifest fields and the metadata it owns. Each definition is written **once**,
+in the body of the package that owns it, beside the artifact's own `manifest`.
+`os dev` boots the same shape straight from source, and `GET /api/v1/packages`
+on a running instance lists every package in the artifact.
 
 The rules worth knowing before you split a product this way:
 
 
@@ -5,6 +5,10 @@
 ([#14487](https://github.com/objectstack-ai/objectstack/issues/14487)): the permission matrix
 §1.3(a) measures is **not** part of this boundary's payoff — permission sets stay whole in the
 `type: app` package.
+**D4's emitted shape amended by the 2026-09-22 addendum**
+([#14512](https://github.com/objectstack-ai/objectstack/issues/14512)): a multi-package artifact
+carries its metadata **once**, in `packages[]` — the flattened copy is no longer emitted beside
+it. D4's read-both rule and D7 are unchanged, so every artifact already on disk still loads.
 **Deciders**: ObjectStack maintainer, 2026-09-01, live PM chat, verbatim and untranslated:
 「立 ADR 起草卡派发」and 「14122 作为epic 任务集中跟踪」— approving the proposal in
 [#14122](https://github.com/objectstack-ai/objectstack/issues/14122) into the ADR-drafting lane
@@ -246,6 +250,11 @@ legibly.
 
 ### D4 — Artifact schema: `packages: [...]` is additive, and both shapes are read
 
+⚠️ **Amended 2026-09-22 (#14512) — see the addendum at the foot of this record.** "Additive" below
+describes the SCHEMA (the key is optional and `manifest` is retained) and the LOAD path (both
+shapes are read), and both still hold. What no longer holds is the EMITTED shape: a multi-package
+`composeStacks(…, { manifest: 'preserve' })` output carries its collections in `packages[]` only.
+
 `ObjectStackDefinitionSchema` gains an **optional** `packages` key carrying an array of
 manifests. `manifest` (singular, `packages/spec/src/stack.zod.ts`) is **retained**, and the load path reads
 both:
@@ -636,3 +645,101 @@ rediscovered:
   all-objects matrix at the environment-admin door. After a split the app package owns the sets
   but no objects, so where a *packaged* cross-module set's grants are authored is a UI question
   this record does not answer and #14488 inherits.
+
+---
+
+## Addendum (2026-09-22, #14512) — a multi-package artifact carries its metadata ONCE, in `packages[]`
+
+**Provenance.** Maintainer ruling, 2026-09-03, live PM chat with the director seat (decision batch
+#23), recorded on
+[#14512](https://github.com/objectstack-ai/objectstack/issues/14512#issuecomment-5528589044).
+Verbatim reply: 「同意」. It re-scopes the 2026-09-02 ruling of the same direction
+([#14512 comment 5518059994](https://github.com/objectstack-ai/objectstack/issues/14512#issuecomment-5518059994),
+also 「同意」) whose cost estimate — "an iteration change, not a shape adaptation" — a reader
+enumeration falsified: the reader surface was at least fourteen sites across three packages, and
+the one-fold option was measured dead. The ruling's own words for what stands:「B stands,
+re-scoped as a program: readers first, emitter last; the release is not held」.
+
+### What D4 said, and what this changes
+
+D4 above declares `packages` **additive**: the load path reads both shapes, and
+`composeStacks(…, { manifest: 'preserve' })` emitted the flattened collections **and** a package
+list carrying the same definitions a second time. That second half is what changes. A
+multi-package artifact now carries each object, view, flow, permission set and every other
+package-owned collection **once**, in the body of the package that owns it.
+
+**⛔ D4's read-both rule is untouched.** `packages` present → iterate it; `packages` absent →
+`manifest` as a single-element list. An artifact built before this change carries both halves and
+loads exactly as it did — which is why this is a producer change and not a format migration, and
+why no artifact on disk anywhere is invalidated by it.
+
+**D7 is untouched and is stated as a condition rather than trusted.** A single-package artifact
+keeps today's shape byte for byte: the emitter strips nothing unless the artifact carries **two or
+more** package entries.
+
+### Why the copy goes, rather than being compressed
+
+- **Size.** The reserved `{ ref, integrity }` external-segment position exists in D4 because
+  artifact size is a real constraint at 2.6 MB for a single package. Emitting every definition
+  twice moves in the opposite direction, and it did so in the same format that reserved the
+  position. Measured on `examples/app-multi-package` (two packages, three definitions):
+  8,223 → 5,328 bytes, −35%.
+- **One source of truth.** Two copies of one definition, and nothing that keeps them equal once
+  they are written. They are measured to differ whenever composition merges or overrides an object,
+  and which copy a consumer sees is decided by the reader it goes through rather than by the
+  artifact.
+
+  ⚠️ **Stated precisely, because the imprecise version is load-bearing in the wrong direction**
+  (PR #19666 review, 2026-09-22): where the two halves differ today, the flattened copy is the
+  RECONCILED one and the platform's own reader deterministically prefers it —
+  `resolveArtifactCollections` claims by name from the top level first, so a merged object is
+  answered once, merged, and its two unmerged halves are dropped. The divergence this record
+  objects to is therefore not "the reader picks arbitrarily"; it is that **one definition is
+  serialized twice with nothing reconciling the copies**, which is a standing invitation for a
+  producer, a hand-edit or a future reader to disagree with that preference. It also bounds what
+  the emitter may do: dropping the flattened half is only ever legitimate where it is a COPY, and
+  where composition reconciled something it is not.
+
+### The order this landed in, which is the decision's substance
+
+**Readers first, emitter last.** Every reader of a top-level collection gained a `packages[]` path
+while the artifact stayed additive (#15004's acceptance probe, #15005 `@objectstack/runtime`,
+#15006 `@objectstack/cli`, #15007 `@objectstack/plugin-security`, #15229 `@objectstack/verify`,
+#15232 `@objectstack/plugin-dev`, and the `plugins` / `devPlugins` contract ruling #15219). The
+emitter is the last step, and the acceptance probe — which boots a two-package collection zoo
+through every load boundary in **both** shapes and fails if any subsystem sees an empty collection
+— is the mechanical criterion for calling the program done, ⛔ never a card-state reading of the
+blocker list.
+
+The failure mode that order exists to contain is a reader nobody enumerated, and it is SILENT:
+nothing throws, the collection is simply absent. ⛔ A missed reader is a new reader card, never a
+fallback bolted onto the emitter.
+
+### What is NOT decided here
+
+- ⛔ **Not a partly flattened artifact.** The emitter flips whole-artifact. A shape carrying some
+  collections flattened and others only in bodies is a new permanent shape and was refused by name
+  in the same ruling (option D).
+- ⛔ **Not the four envelope keys.** `packages`, `plugins`, `devPlugins` and `devLogins` are not
+  package-owned collections (#15219, #17556): they stay at the artifact's top level, where
+  `composeStacks` still concatenates them and the host reads them.
+- **Three conditions hold the emitter to a copy-removal, and a composition failing any one of them
+  keeps the additive shape** — by construction rather than by choice, since stripping there would
+  delete metadata rather than a copy of it:
+
+  1. **two or more package entries** (D7 above);
+  2. **every input's collections attributed to a body** — an input declaring no `manifest` has no
+     package that could own its collections (`manifest` is optional on the stack schema), and an
+     input that already carries `packages` contributes those entries untouched, so its own
+     collections are attributed to no body either;
+  3. **the bodies reproduce the flattened collections, item for item.** Composition is not always a
+     concatenation: `objectConflict: 'merge'` / `'override'` RECONCILE two packages' same-named
+     objects into one, and `mergeActionsIntoObjects` binds a standalone action onto an object a
+     SIBLING package owns. In both cases the flattened half carries a definition no body carries,
+     so it is not a copy. Condition 2 cannot see this — it reads the INPUT stacks, while what the
+     strip deletes is the COMPOSED result. Measured on a two-package `merge` composition: the
+     registration-path reader answers one reconciled object with the flattened half present, and
+     two conflicting partial objects with it gone (PR #19666 review, 2026-09-22).
+
+  Every one of those shapes stays readable by D4's read-both rule, and none is refused: a
+  composition that was legal before this change stays legal.
@@ -20,14 +20,31 @@ import ordersStack from './src/packages/orders/index.js';
  *
  * ## What `preserve` produces
  *
- * The composed stack the platform has always produced — every collection
- * flattened to the top level — PLUS `packages[]`, one entry per input stack,
- * each carrying that package ASSEMBLED (its manifest fields with the
- * collections it owns written over them). The flattened top level is what the
- * metadata service reads; `packages[]` is what `ObjectQL.registerApp`
- * registers, package by package, in dependency-topological order — which is
- * where per-package ownership comes from. Without the list, a two-package
- * artifact would install two package records owning nothing at all.
+ * `packages[]`, one entry per input stack, each carrying that package ASSEMBLED
+ * (its manifest fields with the collections it owns written over them) — and
+ * the artifact's own `manifest`, which is its identity. Each definition is
+ * serialized ONCE, under the package that owns it: a multi-package artifact
+ * carries no flattened copy of its collections at the top level (#14512,
+ * ADR-0130 D4's 2026-09-22 addendum). This project is that shape — its two
+ * packages own different objects, so the package bodies carry everything the
+ * flattened half would have.
+ *
+ * ⚠️ A composition that RECONCILES something keeps the flattened half instead,
+ * because there it is not a copy: `objectConflict: 'merge'` / `'override'` fold
+ * two packages' same-named objects into one, and a standalone action binds onto
+ * an object a sibling package owns. Neither is refused and both stay readable —
+ * the emitter drops the flattened collections only when the bodies reproduce
+ * them item for item.
+ *
+ * `packages[]` is what the metadata service's artifact door reads and what
+ * `ObjectQL.registerApp` registers, package by package, in
+ * dependency-topological order — which is where per-package ownership comes
+ * from. Without the list, a two-package artifact would install two package
+ * records owning nothing at all.
+ *
+ * An artifact built BEFORE that change carries both halves, and every reader
+ * still resolves either (D4's read-both rule), so upgrading the compiler does
+ * not invalidate an artifact already on disk.
  *
  * ## Why the module is listed FIRST
  *
 
@@ -226,15 +226,28 @@ describe('#18024 — `os build` reports a permission-set name another package in
     expect(await findPermissionSetNameCollisions(parsed)).toEqual([]);
   });
 
-  it('reports the collision ONCE, not once per surface the composed artifact carries it on', async () => {
-    // `manifest: 'preserve'` is ADDITIVE: `permissions` composes by `concat`,
-    // so on an artifact the top-level collection is the union of every
-    // package's sets with their per-package provenance flattened away. A walk
-    // that read both levels would report each collision twice — the second time
-    // attributed to whichever manifest composition picked.
+  it('reports the collision ONCE, not once per surface an artifact carries it on', async () => {
+    // A multi-package artifact built before #14512's emitter half carries BOTH
+    // surfaces: `permissions` composes by `concat`, so its top-level collection
+    // is the union of every package's sets with their per-package provenance
+    // flattened away, beside the bodies that still carry them. Such artifacts
+    // are on disk and D4's read-both rule still loads them, so the hazard this
+    // case exists for is live: a walk reading both levels reports each
+    // collision twice — the second time attributed to whichever manifest
+    // composition picked.
     const parsed = parsedArtifact([ordersStack([ordersSet(SHARED_NAME)]), coreStack([coreSet(SHARED_NAME)])]);
-    expect((parsed.permissions as unknown[]).length, 'the top level really does carry both sets').toBe(2);
+    expect(parsed.permissions, 'today the composer emits ONE surface (#14512)').toBeUndefined();
     expect(await findPermissionSetNameCollisions(parsed)).toHaveLength(1);
+
+    // The legacy shape, synthesized: the flattened union written back beside
+    // the bodies, which is exactly what this walk must not double-count.
+    const legacy = {
+      ...parsed,
+      permissions: (parsed.packages as Array<{ manifest: { permissions?: unknown[] } }>)
+        .flatMap((entry) => entry.manifest.permissions ?? []),
+    };
+    expect((legacy.permissions as unknown[]).length, 'the legacy top level really does carry both sets').toBe(2);
+    expect(await findPermissionSetNameCollisions(legacy)).toHaveLength(1);
   });
 
   it('a set whose name no OTHER package here declares is not a finding — the cross-artifact case stays supported', async () => {