VRMA phase 7: manual humanoid clips (avatarA_wave_hello.vrma, etc.) — Blender authoring + T-pose audit

[arkavo-com]

Scope

The VRMA design spec (`docs/superpowers/specs/2026-05-17-vrma-conformance-design.md` — Asset model section) calls for 4-6 hand-authored humanoid behavior clips paired with avatarA:

• `avatarA_wave_hello.vrma` — right arm raise + wave gesture
• `avatarA_nod_yes.vrma` — head yaw + pitch
• `avatarA_idle_breathing.vrma` — subtle hips + chest motion
• `avatarA_blink_sequence.vrma` — eye expression timing

These exercise the integration between humanoid bones + expressions + spring-bone physics + lookAt simultaneously — strong signal that the per-channel parametric sweeps (humanoid + expression + lookAt, 37 plans total) don't produce on their own.

Why deferred

Phase 6 (`docs/superpowers/plans/2026-05-17-vrma-phase6-bootstrap-findings.md`) scoped these out because they require Blender authoring + a one-time T-pose audit per methodology hazard #1 (the .vrma rest pose must be canonical T-pose; mismatches produce systematic bone-rotation diffs that look like renderer bugs but aren't).

Acceptance

• avatarA_1_0.vrm verified at canonical T-pose (one-time audit logged in findings.md)
• 4-6 hand-authored .vrma clips emitted via Blender → UniVRM VRMA exporter
• Plans live under `test-plans/manual/humanoid/` mirroring the existing `avatarA_bosom.test.yaml` shape
• Each plan tests both pose-vector and rendered-frame against the consortium reference (UniVRM)
• Findings.md entry documents cross-renderer agreement on realistic humanoid behavior (separate from the parametric sweep signal)

Methodology hazards to record

• T-pose drift in avatarA produces a systematic bias on every authored clip — audit before authoring
• Multi-channel interactions: a clip that animates both humanoid bones AND expressions both contributes to per-frame SSIM AND per-bone pose-diff; need clear test plan declaration on which channels to assert against
• Animation duration / sample-time choice: short clips with linear keyframes are most cross-renderer-stable; long clips with stepped/cubic keyframes risk interpolation divergence (per design spec methodology hazard VRMA generator: lookAt sweep .vrma files fail UniVRM load (TransferOwnership null reference) #8)

Dependencies

• Issue VRMA: UniVRM batch path applies head bone but leaves limb bones at identity #6 (UniVRM batch limb-bones bug) probably wants to resolve first — otherwise the manual clips will surface the same identity-bone signal we already have from the parametric humanoid sweep
• Issues VRMA generator: expression sweep .vrma files fail UniVRM load (NodeImporter.FixCoordinate index out of range) #7 + VRMA generator: lookAt sweep .vrma files fail UniVRM load (TransferOwnership null reference) #8 (.vrma emission bugs) wanted before the manual corpus to validate UniVRM's parser tolerance of authored-by-Blender .vrma; if Blender-emitted .vrma passes UniVRM cleanly but our generator-emitted .vrma fails, that's a generator-side bug separate from authoring