I wonder how this'd work: Take the XYZ coordinates of a vector on an unit sphere with North Pole being at Z=1 and lat/long 0,0 at X=1. Drop the smallest of XYZ,

I wonder how this'd work: Take the XYZ coordinates of a vector on an unit sphere with North Pole being at Z=1 and lat/long 0,0 at X=1. Drop the smallest of XYZ, keep the other two. Use 3 bits to store which of XYZ was dropped and whether it was positive or negative. The point can then be reconstructed since the vector must always have length 1, and the encoding scheme will always drop that coordinate which would contribute most to loss of precision when going from 3 coordinates to 2.