MIG-5: nothing left behind — the mechanical check and the reduced contract

CHANGELOG.md

@@ -38,7 +38,12 @@ Current schema contract version: **1**.
   the plain lane. It checks the build graph and needs no build. It fails on:
   - a source tree of another repository (`add_subdirectory` out of the repo,
     `FetchContent`, `ExternalProject`);
-  - a `usd-motion-plugins` identity built here;
+  - an identity that left for `usd-motion-plugins` or `motion-connectors`
+    coming back, under the name it had here or the one it has there. That
+    covers a member directory, a target or target prefix
+    (`motionCore_tests`), any file under `adapters/` or `profiles/motion/`,
+    and a source that opens one of their namespaces (`openstrata`, `motion`,
+    `vrmAdapter*`, …). This is MIG-5's mechanical check;
   - the root resolving a consumed package;
   - a member reaching a package that WORKSPACE.md §2 does not allow it. The
     VRM importer may reach no motion package, and the `.vrma` importer may

docs/architecture/PACKAGE_CONTRACT.md

@@ -76,16 +76,18 @@ true.
    descriptor does not declare would put WORKSPACE.md §2 and this file in
    disagreement, and §2 wins. Both left with MIG-3; the rule did not.
 4. **A platform dependency belongs to the library that uses it, not to its
-   consumers.** `liveTransport` links `ws2_32` (Windows) or `Threads::Threads`
-   (elsewhere) `PUBLIC` and its config resolves `Threads` itself; the three
-   adapters that link it declare neither. `ws2_32` needs no `find_dependency`
+   consumers.** `liveTransport` linked `ws2_32` (Windows) or `Threads::Threads`
+   (elsewhere) `PUBLIC` and its config resolved `Threads` itself; the three
+   adapters that linked it declared neither. `ws2_32` needs no `find_dependency`
    because it is a raw library name rather than an imported target — which is
-   also why it is the half of [#113](https://github.com/animu-sphere/usd-vrm-plugins/issues/113)
-   that a Windows run cannot check.
+   also why it was the half of [#113](https://github.com/animu-sphere/usd-vrm-plugins/issues/113)
+   that a Windows run cannot check. All four left with MIG-4; the rule did not,
+   and no package here has a platform dependency today.
 5. **Two libraries with no edge between them bring nothing for each other.**
-   WORKSPACE.md §2 forbids an edge between `liveTransport` and `osc` in both
-   directions, so a consumer of both resolves both. `vrmAdapterVmc`'s config
-   does exactly that, in two separate guarded blocks.
+   WORKSPACE.md §2 forbade an edge between `liveTransport` and `osc` in both
+   directions, so a consumer of both resolved both. `vrmAdapterVmc`'s config
+   did exactly that, in two separate guarded blocks. All three left with MIG-4;
+   the rule did not.
 
 ### 3.1 A `requires` range between two releases
 

docs/roadmap/README.md

@@ -173,15 +173,16 @@ Where things stand, as of 2026-09-19:
   bake agreeing on 414 598 values, every one `==`
   ([release record](../releases/v0.9.0.md)).
 - **Current priority: the [motion migration](motion-foundation-split.md).**
-  MIG-0 started on 2026-09-19. Both destinations are being scaffolded, and
-  MIG-1 moves `motionCore` once `usd-motion-plugins` has a tree that can
-  receive it. Carried beside it: operator evidence for both input halves, the
+  MIG-0..MIG-4 are done on this side (2026-09-19..24): every generic motion
+  identity and every live input left, and this repository consumes what it
+  still uses by digest. What is left is MIG-5's release-artifact proof and
+  its cross-repository test ([current.md](current.md)). Carried beside it: operator evidence for both input halves, the
   Workspace Phase 5 packaging P0, and the build machine's RPATH in packaged
   POSIX binaries ([current.md](current.md#carried-out-of-v090)).
-- **The recorded half's second format family waits behind three tracks**, and
-  its boundary is already built: NPZ / AMASS enters through `motionSource`
-  exactly as BVH does. Whether that is one identity (`motionNpz`) or two
-  (`motionNpz` + `motionAmass`) is a **measurement, not a preference**
+- **The recorded half's second format family left with the layer it extends**:
+  it is no longer this repository's. NPZ / AMASS enters through `motionSource`,
+  which is `usd-motion-plugins`' since MIG-3, and whether it is one identity
+  (`motionNpz`) or two (`motionNpz` + `motionAmass`) is measured there
   ([the recorded track](recorded-motion-sources.md) §13).
 - The display slice is **re-scoped** (2026-07-29). OpenUSD 26.08 resolves exec
   prim adapters from a hard-coded list, so a skinned VRM avatar cannot be

docs/roadmap/current.md

@@ -182,46 +182,39 @@ steps **Migration Phase A–F**.
 
 It starts after v0.9.0, on purpose: the OpenExec foundation's findings are the
 API defects the move fixes on arrival, and they only exist once its nodes do.
-**MIG-0 is done (2026-09-19)** ([the track §2](motion-foundation-split.md#2-mig-0--preparation-)).
-`motionCore` (MIG-1) and `motionRuntime` (MIG-2's first item, as
-`motionSampling` and `motionRecording`) have arrived in `usd-motion-plugins`
-with their history, and so have `vrmRetarget`'s generic half, `motionUsd`'s
-authoring half and the recorded sources (MIG-3). **MIG-1 and MIG-2's library half is done too (2026-09-21).** `motionCore` and
-`motionRuntime` are consumed packages: `motionCore`, `motionSampling` and
-`motionRecording` from `usd-motion-plugins` v0.5.0, pinned by digest per target
-in five descriptors, with the whole suite green against them.
-**So is the retarget (2026-09-23)**: `vrmRetarget`'s generic half is the
-consumed `motionRetarget`, both builder copies are calls to it, every parity
-row came out identical, and what stayed is `vrmRig`
-([the track §4](motion-foundation-split.md#4-mig-2--sampling-retarget-usd-bridge-)).
-**And so is `motionUsd`'s reading half (2026-09-23)**: `motion_retarget`
-reads the clip through the consumed `ReadMotionStage` and keeps only the
-`vrm:` tracks, the first edge here that only a tool declares — what `ost`
-0.23.3 made materializable
-([report 44](../reports/ost/44-2026-09-23-v0.23.2-a-tool-edge-reaches-nothing-and-a-tree-keeps-its-runtime.md)).
-**MIG-3 is done too (2026-09-23)**: `motionSource`, `motionBvh`, the BVH tools
-and the profiles are deleted here, their non-ASCII path cases having moved
-first, and the suites that baked a real capture read a clip the published
-converter wrote. What is left of the consuming side is one thing: `execMotion`,
-which that repository published in v0.5.0 and which no descriptor here can
-name — `requires.bundles` takes no artifact pin in `ost` 0.23.3 (report 45) —
-so the parity rows cannot yet be re-run against the consumed package before
-the copy here is deleted.
-
-**MIG-4 is done on both sides (2026-09-21).** All six connector-bound
-identities arrived in `motion-connectors` — the two leaves, the tracker layer
-and the three adapters with their recorders — and this repository deleted its
-copies in one change, under [WORKSPACE.md §9.2](../architecture/WORKSPACE.md#92-moving-rules)
-rule 7. `ost` 0.23.2 is what made the cross-repository edge declarable
-([report 43](../reports/ost/43-2026-09-20-v0.23.1-the-root-build-cannot-see-an-external-library.md)),
-so MIG-1..MIG-3's consuming change — this repository resolving `motionCore` and
-the rest as installed packages — is unblocked and is the next thing the track
-owes. MIG-4's other half, `motion_capture`, followed on 2026-09-23. It is
-`usd-motion-plugins`' `motion_record`, and `motion_retarget`'s suite bakes a
-clip the published recorder wrote
-([the migration track](motion-foundation-split.md#6-mig-4--recording-and-live-input-) §6).
-Until an identity moves, it takes fixes and the work v0.9.0 owes, and **no new
-generic capability** ([WORKSPACE.md §9.1](../architecture/WORKSPACE.md#91-destination-of-every-identity)).
+
+**MIG-0..MIG-4 are done (2026-09-19..24), and this repository builds no
+generic motion code.** Every identity
+[WORKSPACE.md §9.1](../architecture/WORKSPACE.md#91-destination-of-every-identity)
+sends elsewhere arrived there with its history, and this repository either
+consumes it by digest or no longer uses it:
+
+- `motionCore`, `motionSampling`, `motionRetarget` and `motionUsd`'s reading
+  half are consumed packages. What stayed of `vrmRetarget` is `vrmRig`.
+- The recorded-file layer and `motion_capture` are deleted here. Suites that
+  baked a real capture read a clip the published converter or recorder wrote.
+- The six connector-bound identities left in one change (MIG-4).
+- `execMotion` is the published bundle, pinned in `execVrm`'s
+  `requires.bundles`. The parity rows were re-run against it before the copy
+  was deleted, with divergence 0 (2026-09-24).
+
+Each step's record is in [the track](motion-foundation-split.md).
+
+**What is left is MIG-5, "nothing left behind"** ([the track §7](motion-foundation-split.md#7-mig-5--nothing-left-behind-)):
+
+- ✅ No generic motion source file remains, checked mechanically by
+  `workspace_cmake_boundaries` (2026-09-24).
+- ✅ WORKSPACE.md §1 and §2 describe the reduced tree and §9 is a record;
+  `PACKAGE_CONTRACT.md` lists only what this workspace installs (2026-09-24).
+- ⬜ The aggregate product installs and opens a `.vrm` and a `.vrma` from
+  release artifacts, with the shared core resolved as a dependency. That is
+  the next release's dry run.
+- ⬜ The cross-repository test — VRMA → `MotionClip` → a target VRM — runs by
+  default somewhere, and here until an integration repository exists.
+
+A new generic motion feature is proposed in `usd-motion-plugins`, and a new
+device or protocol input in `motion-connectors`
+([WORKSPACE.md §9.1](../architecture/WORKSPACE.md#91-destination-of-every-identity)).
 
 What becomes of [boundary consolidation](boundary-consolidation.md):
 
@@ -294,12 +287,10 @@ the workspace layout, the output structure, and the import/runtime boundary.*
 (design policy §15, §17-P0)
 
 - 🚧 Describe `vrmSchema`, `usdVrmFileFormat`, `usdVrmPackageResolver`,
-  `usdVrmaFileFormat`, `execMotion` and `execVrm` as separate bundles;
-  `vrmContainer`, `motionCore`, `motionRuntime`, `vrmRetarget`, `motionSource`,
-  `motionBvh`, `motionTracking`, `liveTransport`, `osc` and the three
-  `vrmAdapter*` leaves as plain libraries; `motion_retarget`, `motion_capture`,
-  `motion_bvh_convert` and the `*_record` tools as CLIs; and `usdVrm` as the
-  aggregate product name only.
+  `usdVrmaFileFormat` and `execVrm` as separate bundles; `vrmContainer` and
+  `vrmRig` as plain libraries; `motion_retarget` as a CLI; the
+  `usd-motion-plugins` packages and the `execMotion` bundle as consumed, not
+  built; and `usdVrm` as the aggregate product name only.
 - 🚧 Unify phase notation to **Product P0–P6**, **Workspace Phase 0–8**, and
   **Motion Phase A–H** — three sequences, never a bare "Phase N".
 - 🚧 Align build / test / install examples with what CI actually runs.

docs/roadmap/motion-foundation-split.md

@@ -1,6 +1,6 @@
 # Motion migration — generic motion to `usd-motion-plugins`, input to `motion-connectors`
 
-**Status:** ✅ MIG-0; ✅ MIG-3 — `motionSource`, `motionBvh`, the BVH tools and the profiles deleted here on 2026-09-23; 🚧 MIG-1 and MIG-2 — **consumed here since 2026-09-21: `motionCore`, `motionSampling`, `motionRecording`, and since 2026-09-23 `motionRetarget` and `motionUsd`**, with what stayed of `vrmRetarget` renamed `vrmRig`; and since 2026-09-24 `execMotion`, the last generic motion code this repository built; **every sending half of MIG-2 has arrived** — `motionRetarget` 2026-09-19, `execMotion` and `motionUsd`'s reading half 2026-09-20; **✅ MIG-4 on the connector side** — all six identities arrived in `motion-connectors` 2026-09-19..21 and left here in one change on 2026-09-21; `motion_capture` arrived in `usd-motion-plugins` as `motion_record` (2026-09-20) and was deleted here on 2026-09-23 · **Target:** after the OpenExec foundation ·
+**Status:** ✅ MIG-0..MIG-4 on this side — every identity [WORKSPACE.md §9.1](../architecture/WORKSPACE.md#91-destination-of-every-identity) sends elsewhere arrived there with its history (2026-09-19..21) and is consumed or deleted here (2026-09-21..24), `execMotion` last; what MIG-4 still owes is `motion-connectors`' (recorded evidence, ARDY); 🚧 MIG-5 — the mechanical check and the reduced contract are done, the release-artifact proof and the cross-repository test are not · **Target:** after the OpenExec foundation ·
 **Structure:** [architecture/WORKSPACE.md §9](../architecture/WORKSPACE.md#9-destinations-under-the-motion-architecture) ·
 **Policy:** the `usd-motion-plugins` design policy §37, and
 [design/INTEGRATION_SCOPE_POLICY.md](../design/INTEGRATION_SCOPE_POLICY.md) §13 ·
@@ -117,7 +117,7 @@ repository's, and needs nothing from this one.
   Boundary suites (`*_boundaries`) travel too, and they are rewritten there
   for that repository's edges rather than reproduced.
 
-## 3. MIG-1 — the core 🚧
+## 3. MIG-1 — the core ✅
 
 - ✅ `motionCore` arrives in `usd-motion-plugins` under the same identity, with its
   history, renamed to the shared names, under `openstrata::motion`
@@ -140,13 +140,15 @@ repository's, and needs nothing from this one.
   ([WORKSPACE.md §9.2](../architecture/WORKSPACE.md#92-moving-rules), rule 1).
   Adapting code here to the renamed types was acceptable during migration
   (motion-plugins policy §37); keeping two cores was not.
-- ⬜ The `.vrma` stage does not change: `/Animation`, `HumanoidSkeleton`,
+- ✅ The `.vrma` stage does not change: `/Animation`, `HumanoidSkeleton`,
   `BodyAnimation`, the `vrma` custom data. A standalone motion stage in
   `usd-motion-plugins`' shape is a separate decision, not a side effect.
+  Held across the whole move: no `usdVrmaFileFormat` golden changed from
+  2026-09-18 to MIG-2's last deletion on 2026-09-24.
 
-## 4. MIG-2 — sampling, retarget, USD bridge 🚧
+## 4. MIG-2 — sampling, retarget, USD bridge ✅
 
-- 🚧 `motionRuntime` arrives as `motionSampling` and `motionRecording`, with
+- ✅ `motionRuntime` arrives as `motionSampling` and `motionRecording`, with
   the exec findings fixed on arrival: a status-carrying `SampleClip`, a
   stateless `PoseFilter` step, `ConditionRootMotion` as a free function, an
   N-way blend that can answer *nothing to blend*.
@@ -178,7 +180,7 @@ repository's, and needs nothing from this one.
     The two behaviours that changed with the package — `BlendPoses`'
     `std::nullopt` and a NaN weight counting as none — are read, not
     predicted, by `execMotion` here.
-- 🚧 The generic retarget arrives as `motionRetarget`, with a
+- ✅ The generic retarget arrives as `motionRetarget`, with a
   `SkeletonDescriptor` built from joint tokens and rest matrices — the
   finding `execVrm` and `motion_retarget` both carry a copy of today.
   - ✅ Arrived with its history (2026-09-19,
@@ -223,7 +225,7 @@ repository's, and needs nothing from this one.
     [WORKSPACE.md §9.5](../architecture/WORKSPACE.md#95-the-line-through-vrmretarget)'s
     line is held after the cut rather than only drawn before it. MIG-0's
     vocabulary check retired in the same change, its last headers gone.
-- 🚧 `motionUsd`. The authoring half arrived on 2026-09-19
+- ✅ `motionUsd`. The authoring half arrived on 2026-09-19
   ([usd-motion-plugins #7](https://github.com/animu-sphere/usd-motion-plugins/pull/7)). Its source was
   `motion_capture`'s `ClipWriter`, not `StageIo`. `StageIo` reads a clip and
   bakes it onto a VRM, and that writing half is VRM-specific and stays here.
@@ -363,7 +365,7 @@ repository's, and needs nothing from this one.
       because `execVrm`'s boundary check had imported them from the bundle's
       tree. The schema partition is now read from the consumed bundle's
       `plugInfo.json`.
-- 🚧 What stays is re-read as a consumer: the VRM humanoid map,
+- ✅ What stays is re-read as a consumer: the VRM humanoid map,
   `ExpressionResolver`, `LookAtEvaluator`, `motion_retarget` as a VRM CLI,
   `execVrm`. The OpenExec parity values are re-run against the consumed
   packages before anything here is deleted — done for the retarget on
@@ -537,11 +539,37 @@ repository's, and needs nothing from this one.
 - ⬜ The ARDY adapter is created there, behind the generator interface
   `usd-motion-plugins` specifies; Motion Phase F does not start here.
 
-## 7. MIG-5 — nothing left behind ⬜
+## 7. MIG-5 — nothing left behind 🚧
 
-- ⬜ No generic motion source file remains here, checked mechanically.
-- ⬜ WORKSPACE.md §1 and §2 describe the reduced tree, §9 becomes a record,
-  and `PACKAGE_CONTRACT.md` drops the packages that left.
+- ✅ No generic motion source file remains here, checked mechanically
+  (2026-09-24). `scripts/check_cmake_boundaries.py`
+  (`workspace_cmake_boundaries`, and first in the plain-CMake lane) fails
+  when an identity [WORKSPACE.md §9.1](../architecture/WORKSPACE.md#91-destination-of-every-identity)
+  sent elsewhere comes back. It checks every name the identity has had, here
+  and in its destination. A member directory fails, and so does a target or
+  target prefix (`execMotion_blend_tests`), any file under `adapters/` or
+  `profiles/motion/`, and a source that *opens* one of their namespaces.
+  Using them (`openstrata::motion::MotionPose`) is the point of consuming
+  them, and passes. The tree passes today. Before this, the check knew only
+  `usd-motion-plugins`' current names, so none of the connector identities,
+  `motionRuntime` or `vrmRetarget` could have failed it.
+- ✅ WORKSPACE.md §1 and §2 describe the reduced tree, §9 becomes a record,
+  and `PACKAGE_CONTRACT.md` drops the packages that left (2026-09-24).
+  - WORKSPACE.md lost about 700 lines. §1 lists what this workspace builds,
+    and names what it consumes rather than tabulating it. §2's edges are the
+    ones the descriptors declare, and the graph gate reports
+    `5 bundle(s), 3 bundle edge(s), 2 libraries, 13 library edge(s), 1 tool(s)`
+    on `ost` 0.23.6. §5 keeps the product-membership rule — no product name,
+    and nothing that opens a transport — with the adapters and the BVH layer
+    it was argued over gone. §9 opens with the moves done and dates every
+    row.
+  - The reasoning each removed section held about identities that left is
+    the v0.9.0 copy of the document, the last release that built them, and
+    the document says so rather than keeping it.
+  - `PACKAGE_CONTRACT.md` §4 had already dropped each package in the change
+    that moved it. What was left was two §3 rules stated over `liveTransport`,
+    `osc` and the adapters in the present tense, now in the past with the
+    rule kept.
 - ⬜ The aggregate product still installs and opens a `.vrm` and a `.vrma`,
   with the shared core resolved as a dependency, from release artifacts.
 - ⬜ The cross-repository test — VRMA → `MotionClip` → a target VRM — runs