Say we are in Minkowski. Choose a set $S$ of null vectors $k^\mu$ with $(k^x)^2+(k^y)^2 + (k^z)^2 = (k^0)^2=1$. In the spatial sense, the vectors of $S$ span a "cellestial sphere". For every of these initial null vectors, there is a unique light-ray coming from the origin -- a vector not in $S$ would generate a light-ray already in the set, just with a linearly rescaled affine parametrization. We can thus define a projective equivalence $k^\mu \sim l^\mu$ if $k^\mu = \lambda l ^\mu, \lambda >0$ and work with null vectors modulo this equivalence. This set is isomorphic to $S$ and can be called "the true cellestial sphere".
The question now is: Is there a Lorentz-invariant measure over this set of light-ray directions/ the cellestial sphere?
We could integrate with a euclidean measure over the sphere in $(k^x)^2+(k^y)^2 + (k^z)^2 =1
$ but this would trivially fail under boosts. I.e., the problem seems to be somehow analogous to the problem of gauge fixing in el-mag field quantization because a part of the solution is to covariantly constrain an unphysical degree of freedom of a massless particle (at least in the sense of classical light-rays, not waves), and $k^0=1$ will not do.
I thought about the geometrical construction of spinors and how 2-spinors can be understood exactly as directions of light-rays but the $\xi^\alpha, \alpha=1,2$ representation is plagued with the very same $\lambda$-redundancy as the $k^\mu$s. The only representation not plagued by the $\lambda$-redundancy is the one obtained by a stereographic projection from $S$ into the complex plane, i.e. through a single c-number on the Riemann sphere $z=\xi^1/\xi^2$. Lorentz transformations are then represented by restricted Möbius transformations but Möbius-invariant measure is impossible, at least in a canonical sense.
Another approach would be to try to "factor out" the $\lambda$-infinity or regularize it in a renormalization-like manner but approaching from time-like or space-like vectors really doesn't help. Any ideas?