There is no conventional choice, the Jones vector formalism must be attached to
- A local coordinate system dependent on the $\bf{k}$-vector and its direction. The Jones vector is always chosen so that the $\bf{k}$-vector corresponds to the positive $z$-axis. This is the current way that optical professionals and particularly professionals in polarization optics define it (as far as I know), see below for further discussion.
- A right-handed coordinate system.
You do have a choice between the increasing and decreasing phase conventions, which will change signs all over the place, and can make calculations incorrect if you mix them up. Be very careful of the wikipedia article on Jones vectors and matrices, they do not use a consistent phase convention (even though they state that they do at the top).
- Decreasing phase: $({\bf{k}}\cdot {\bf{r}}-\omega t)$
- Increasing phase: $(\omega t-{\bf{k}}\cdot {\bf{r}})$
Since the Jones vector is attached to a (possible changing) local coordinate frame, a good way to do a polarization ray trace is to use a global three dimensional coordinate system. This has been accomplished by Yun, Crabtree, McClain and Chipman in their papers "Three-dimensional polarization ray-tracing calculus": "definition and diattenuation" (Appl. Optics 50 no. 18, pp. 2855-2865 (2011), doi:10.1364/AO.50.002855), and "retardance" (Appl. Optics 50 no. 18, pp. 2866-2874 (2011), doi:10.1364/AO.50.002866).
Take note that because of the inherent rotation of the coordinate system through/from interfaces, the Jones matrices will appear to have retardances that are not physically created, but are just a geometrical artifact of the change of local coordinate system.
Poynting vector vs $\bf{k}$-vector definitions
As user17581 pointed out, some people have defined the Poynting vector as the direction of the $+z$-axis. This makes sense because ${\bf{S}}={\bf{E}}\times{\bf{H}}$ so the electric field is always perpendicular to ${\bf{S}}$ and ${\bf{H}}$. The question is then, why isn't it defined that way?
I believe it is because in a material, like a crystal, there are two or three indices of refraction, in two or three particular directions. If a Jones matrix is described for a particular input angle via the Poynting vector in a crystal then a particular Jones matrix is obtained (i.e., it really doesn't describe the entire space of possible inputs to the crystal). This matrix, however, doesn't really make sense because the underlying material properties are anisotropic. If you, however, define the Jones matrices for each index of refraction (and the associated $\bf{k}$-vectors), and model the propagation as two (or more) rays through the crystal, then recombine them at the end, then you basically have a "basis" of Jones matrices for the crystal, which you do not have to recompute each time.
