Structural math types, no three dependency

Vec2Like, Vec3Like, QuatLike and TransformLike are read-only structural interfaces. THREE.Vector3 and THREE.Quaternion satisfy them already, so 3D editors pass mesh.position and mesh.quaternion directly, while a 2D consumer of a button or a colour swatch never pulls a 3D library. THREE.Object3D is not itself TransformLike; a one-line adapter is expected at the call site.

Values are immutable snapshots, because Lit re-renders on assignment and mutating in place does not repaint. Properties declare hasChanged with a component-wise comparison so re-assigning an equal-valued object does not repaint either.

Vector fields apply Mixed per axis rather than whole-value, so their value is a union (Record<TAxis, number> | Record<TAxis, FieldValue<number>> | typeof Mixed). The plain record is the common direct-binding case; the per-axis form only appears under multi-selection.

Considered Options

• Depending on three for math types. A 2D colour swatch would transitively pull a 3D engine.
• Scrubbing a mixed field from a synthesised start. The first pointer move collapses the whole selection onto a value the user never saw, in one undo step. Mixed disables gestures that need a starting value; typing stays available.