Model calibration

Measure each unique GLB once with inspect-glb, only trusting bounds when boundsSource is vertices, and keep one tuning record per asset:

const ASSET_TUNING = {
    model: {
        boundsSource: 'vertices', dims: [8, 7, 18], center: [0, 2.5, 0], groundOffset: 1,
        intended: { dimension: 'length', size: 18 }, scale: 1, y: 1, yaw: 180
    }
} as const;

1. Pick the dimension that defines the model's real size, in metres or relative to an already calibrated model.
2. Set scale = intendedSize / measuredDimension and, for floor-resting models, y = groundOffset * scale. Record a deliberate offset for models that should not rest on the floor.
3. Confirm facing once in the running app and record yaw.

Nest three levels so a root position means the same for every asset: the semantic root, a yaw wrapper, then the visual child with the scale and local position (-center.x * scale, y, -center.z * scale). Apply calibration once, never per instance.

Place the root on a measured support point, such as a mount point or raycast, not a global AABB maximum. Skinned bounds are bind-pose estimates; confirm contact in animated poses.

Read only the reference matching the code: direct Engine, React, or Web Components.