AI: RigNet auto-rigging (ONNX, ML)

[fernandotonon]

Parent epic: #397

Goal

Higher-quality auto-rigging than Pinocchio using RigNet (2020, ONNX-exportable). Better handling of arbitrary topology, non-humanoid shapes, and unusual proportions.

Scope

• Reuse the ONNX Runtime infrastructure landed by AI: PBR map synthesis from albedo (DeepBump-style, ONNX) #404 (PBR synth).
• Vendor a working RigNet ONNX export (will need to verify availability — fallback: do the export from the published PyTorch checkpoint as a one-time tooling step).
• Model size: ~50 MB; download via ModelDownloader on first use.
• CLI: qtmesh rig model.fbx --algo rignet -o rigged.fbx (default: pinocchio for offline reliability).
• GUI: skeleton-prediction algorithm dropdown alongside AI: Pinocchio auto-rigging (classical, native C++) #407.
• MCP: extend the auto_rig tool with an algo param.
• Pipe into AI: libigl bounded biharmonic skinning weights #402 (libigl skin weights) the same way as AI: Pinocchio auto-rigging (classical, native C++) #407.

Acceptance Criteria

• Works on humanoid, quadruped, and at least one non-standard shape (e.g. chair-with-legs).
• Output skeleton hierarchy matches what a reasonable artist would author.
• First-run model download flow with progress and offline failure messaging.
• CLI/MCP/GUI parity.
• Sentry breadcrumb ai.assist.auto_rig with algo: rignet tag.

Dependencies

• ONNX Runtime infra (AI: PBR map synthesis from albedo (DeepBump-style, ONNX) #404).
• Pinocchio task (AI: Pinocchio auto-rigging (classical, native C++) #407) for the rigging plumbing.

Effort

~10–12 days.