It's using a Vec right now; it should be simplified. - Use “small vec” optimization; no allocation for small dimensions - Cap maximum number of axes to some limit (u16?)