MIG-5: the reference pipeline — three sources, one downstream half

[snkmcb]

MIG-5's cross-repository test ("VRMA → MotionClip → a target VRM runs somewhere by default"), which is also boundary consolidation's BND-1 ("one reference pipeline, proved once for every category").

What it runs

workspace_reference_pipeline bakes three source categories onto Seed-san.vrm:

Source	First arrow	Input
.vrma	this repo's usdVrmaFileFormat	canonical_walk.vrma
BVH export	usd-motion-plugins' motion_convert	tests/motion/fixtures/mocopi-mobile-arm-raise-turn.usda (committed output)
recorded live session	usd-motion-plugins' motion_record	tools/motionRetarget/tests/fixtures/recorded_session_clip.usda (committed output)

All three go through one motion_retarget argument list that differs only in --animation.

What it asserts: the downstream half is the same code

• One set of loaded code. --load-report's plugins and modules differ across sources by the first arrow's plugin alone. Measured: a .vrma adds exactly UsdVrmaFileFormat / libUsdVrmaFileFormat.dll, and the other two load identical sets.
• One output shape. Every authored spec's path, kind and type, and the animation's joint order, are identical across sources.
• One set of motion checks, one tolerance. For each source:
  • every bone the clip carries and the avatar binds reproduces the clip's rotation away from rest;
  • the hips reproduce the clip's travel;
  • bones the avatar cannot bind are named on stderr;
  • nothing outside the humanoid moves;
  • scales are authored (usdVrmaFileFormat: BodyAnimation omits scales, so an imported .vrma resolves no joint transforms #64).

Mutations (each fails where it should; unmutated passes)

Mutation	Caught by
--root-motion ignore for the trace source only	hips travel
left/right upper arms swapped via --humanoid-map for the BVH source only	rotation (18 failures)
no first-arrow plugin allowed for .vrma	loaded-code check
--animation-name Other for the trace source only	output shape

Local, non-redistributable data

workspace_reference_pipeline_local runs the same checks over an avatar and .vrma files that cannot be committed. It reads them from USDVRM_LOCAL_AVATAR and USDVRM_LOCAL_VRMA_DIR, and USDVRM_LOCAL_OUTPUT_DIR optionally keeps each bake. Without them it reports Skipped (SKIP_RETURN_CODE 77). On 2026-09-24 it passed over AliciaSolid.vrm (VRM 0.x) and the seven VRM Animation MotionPack clips, with the BVH export and the recorded session beside them (9/9). Neither dataset may be redistributed: AliciaSolid is redistributionAllowed: false in the corpus manifest, and the MotionPack terms forbid redistribution in extractable form.

Verification (Windows, local)

• ost test: 41 passed, 1 skipped (the local test, without its variables).
• ost test --filter workspace_reference_pipeline with the local variables set: both pass.
• check_docs.py OK; check_cmake_boundaries.py holds.

Docs: CHANGELOG [Unreleased], MIG-5 item 4 ✅ (track and current.md), BND-1 ✅. MIG-5's last item, the release-artifact proof, is the next release's dry run.

🤖 Generated with Claude Code