fix(animation): Convert animation rest pose for VRM 0.0 retargeting

[arkavo-com]

Screenshots

Before / after on the same VRM 0.0 + VRMA case (AliciaSolid + Idle.vrma, frame 0):

Before — upside-down	After — upright

Regression checks — VRM 1.0 unchanged, and the fix composes with the silhouette feature in #137:

VRM 1.0 + VRMA (no regression)	VRM 0.0 + VRMA + --silhouette

Summary

Applies the existing convertForVRM0 coordinate transform to animationRestRotation and animationRestTranslation in VRMAnimationLoader.swift so the delta-retargeting computation delta = inverse(rest) * rotation no longer mixes coordinate frames.

The bug

For VRM 0.0 models with a VRMA animation, the avatar rendered upside-down once playback started — but rendered correctly when no animation played, and VRM 1.0 + VRMA was unaffected. Triangulating the symptoms pointed to the retargeting layer:

In makeRotationSampler, the per-frame rotation was being converted into the model's coordinate frame via convertRotationForVRM0, but the rest rotation it was being subtracted against (rotationRest) was not converted. The delta was therefore expressed in a hybrid frame. For pure Y-axis rotations the conversion is a no-op (Y stays positive), so the bug was invisible there; X- and Z-axis rotations (which is what an idle-hips track contains) accumulated a residual frame rotation that, combined with the renderer's vrmVersionRotation at draw time, manifested as a 180° X flip.

The fix

Symmetry: convert the rest rotation/translation with the same convertForVRM0 transform before computing deltas. Translation gets the same treatment (convertTranslationForVRM0 applied to animationRestTranslation).

19 lines added / 2 removed across one file. convertForVRM0=false (VRM 1.0) is bit-identical.

Verification

Case	Before	After
AliciaSolid (VRM 0.0) + Idle.vrma	upside-down	✅ upright in idle pose
AliciaSolid (VRM 0.0) + Idle.vrma + --silhouette	upside-down	✅ upright black silhouette
VRM1_Constraint_Twist_Sample (VRM 1.0) + Idle.vrma	upright	✅ unchanged, upright
swift test --parallel	n/a	✅ exit 0, 1188 tests

Background — and what NOT to pull forward

There is a stale local branch (now archive/animation-coordinate-fix-jan2026, ex feature/animation-coordinate-fix, Jan 19) that addressed a related symptom with a different mental model: it negated quaternion X+Y at the parser level on every glTF animation, framed as "RH→LH conversion." That premise doesn't match this codebase — the renderer doesn't flip handedness, and applying that conversion universally would break VRM 1.0 + VRMA renders. The salvageable part of that branch is the bone-name heuristic improvements (Mixamo, Blender, namespace prefixes), filed as #135.

Test plan

• swift build clean
• swift test --parallel --num-workers 14 -j 16 --disable-sandbox — 1188 tests, exit 0
• Visually verified VRM 0.0 + VRMA renders upright (still + animated frames)
• Visually verified VRM 1.0 + VRMA unchanged
• If you have a regression model that exercised the original "lying down" symptom, please verify it stays fixed

Notes

Branch is named spike/ because it was developed as a focused experimental fix; it is now verified and ready to land. If you'd prefer a renamed branch for tidiness happy to push to a fix/ name and update this PR.

🤖 Generated with Claude Code