AI v2 B2: retrain the vehicle mesh segmentation on generated (TRELLIS/TripoSG) vehicles + rig-prior labels from the vehicle template

[fernandotonon]

Part of #818 (Track B2). Follow-up to #1015 (vehicle plumbing shipped, model trained on synthetic blobs only) and #1050 (vehicle skeleton template + rigid one-bone binding).

Why now

meshseg_vehicle.onnx (vehicle_body / wheel / window / wing / rotor) is trained exclusively on the synthetic make_vehicle generator in scripts/export-meshseg-onnx.py — boxes with blob wheels. Real cars (fused low-poly bodies like ugly_car.obj, detailed multi-part models like the Buick) look nothing like that, and the model has never seen one. Two things changed that make a real retrain cheap:

1. We can generate vehicles on demand. qtmesh generate3d --prompt "…" (TRELLIS.2 / TripoSG / TripoSR) produces cars, trucks, buses, planes, helicopters, motorbikes, wheels-only, wing-only meshes — unlimited, varied, and license-clean (our own outputs of MIT-licensed models; no ShapeNet/PartNet non-commercial data).
2. The vehicle rig gives EXACT body/wheel labels for free. --skeleton vehicle binds every vertex to exactly one of Chassis / FrontAxle / RearAxle / 4 wheel bones (feat(#1013): vehicle skeleton template + rigid one-bone binding #1050). That is a per-vertex label. The existing rig-prior miner (qtmesh segment --dump-training-data, AutoRig::rigPriorPartLabels → MeshSegmenter::partForBoneName) already turns bone weights into exact training samples — it only knows body-part bone names today.

Plan

• Miner: extend partForBoneName with the vehicle vocabulary — Chassis/*Axle → vehicle_body, *Wheel → wheel (also common artist names: wheel_fl, tire, rim, hub; wing, rotor, propeller, blade). Make --dump-training-data write the sample's category (schema bump qtmesh-meshseg-training-v2, v1 = body) so the export script routes it to the right model.
• Pipeline script scripts/mine-vehicle-segdata.py: prompt list × seeds → generate3d (TRELLIS.2 when available, else TripoSG; --target-tris 25000) → rig --skeleton vehicle --rigid → segment --dump-training-data. Reject samples whose wheel snap failed (a wheel bone with < N vertices, or wheels not at the four corners — reuse the fitVehicle quadrant check). Also mine the CC0 vehicle packs already used for the body miner (Quaternius) and the two local test cars.
• Parts as standalone meshes: run PartOps split (segment --split-parts) on the rigged cars and export each part alone (wheel, body) — the Auto category classifier (meshseg_category.onnx) must still say vehicle for a lone wheel or a wing, which is the verified failure mode noted in the export script (detached parts at odd offsets).
• Window / wing / rotor labels: the rig gives body vs wheel only. Options, in order of preference: (a) keep the synthetic generator for these three classes and mix; (b) label windows on generated cars via the baked texture (dark/low-roughness glass texels → window) — cheap but noisy, gate it; (c) SAM 2 multiview (AI v2 B2: sam2seg — zero-shot mesh segmentation via multiview SAM 2 #1014) if it lands first. Do NOT block the retrain on this — body/wheel is where the real-mesh gap is.
• Retrain meshseg_vehicle.onnx with --real-data ./mined_vehicle/ mixed into the synthetic set (yaw/tilt/jitter aug as for body) and retrain the category classifier with the generated vehicles + lone parts. Upload via the existing HF flow (segment/ on QtMeshEditor-models), update THIRD_PARTY_AI_MODELS.md provenance (generated-data statement).
• Evaluation (held out, never trained on): ugly_car.obj, the Buick, 3–5 Quaternius vehicles, 3 generated vehicles from prompts not in the training list. Metrics: per-vertex accuracy + wheel IoU vs the rig-prior labels; category classifier accuracy on whole vehicles and on lone parts. Record the numbers in docs/MESH_SEGMENTATION_STRATEGY.md. Acceptance: wheel IoU ≥ 0.8 on the two real cars (today the geometric fallback's "lowest 30 %" rule is the bar to beat).
• Verify in-app: Edit Mode → "Select by Part (AI)" on an unrigged generated car selects a wheel cleanly; qtmesh segment car.glb --category vehicle per-part counts look right.

Notes / risks

• Generated meshes are single fused shells: wheel arches merge into the body, exactly like ugly_car.obj — good, that IS the hard case the synthetic set lacks. Discard generations where the rig's wheel snap fails rather than training on bad labels.
• Training samples are point clouds + labels, so TRELLIS textures are irrelevant except for the optional window heuristic.
• Keep the one-model-per-category design (decision record in docs/MESH_SEGMENTATION_STRATEGY.md); this retrains two files, adds none.
• Vehicle-rig-derived labels only cover 4-wheel ground vehicles; planes/helicopters/bikes need the synthetic classes or SAM 2 until a matching rig template exists.