my question is not very clear ... I need to render a mesh(sort of quad) between 2 joints ...
No, it isn't. Once again, you're describing HOW you want to do it, rather than WHAT you want the results to be. You get full credit for trying different things to achieve the results you want. However, at this point you haven't described what you want the appearance of something (mesh, quad, ...?) to be.
It appears part of the confusion is whether you want to render a mesh or a quad. It's not clear what "sort of quad" means in that context. A quad is 4 vertices. A mesh is either comprised of 4 vertices, or it's not. Further, just so we're working with the same assumptions, the term "quad" does not imply billboarding, 2D rendering, orientation, etc. It's just 4 positions in some space (which space that may be is not defined by the term).
Because you mention scale, this is just a guess - do you want to draw a mesh representing a bone that appears the same size on the screen, regardless of it's distance from the viewpoint? I.e., do you want the circles representing the joint positions to always appear to be the same size in pixels (i.e., screen-size rather than world-size)?
how would you render a quad between those two joints knowing that its position, orientation and scale need to be computed in projection space ?
Because it's still not clear WHAT you want the results to be, my answer: I wouldn't. Seriously, if the user has chosen perspective projection (rather than orthographic), I would render a bone mesh in perspective - the farther it is from the eye, the smaller it appears. If, in contrast, the joint circles are all rendered with the same radius in screen space, when the user zooms out, it might become a confusing mess of overlapping circles. Your second pix above would be what I would do, because the circle sizes provide visual cues to the user with regard to which joint is closer, and what the relative orientation of the bone is.
Still guessing what you're trying to achieve, You may want to consider an orthographic projection. In that case, all objects, near and far, are rendered to the same scale. All the joint circles would be the same size, etc.
EDIT: It's important for you to decide what you want the results to be - and to know what the advantages and disadvantages of that are.
For my own purposes, because my old eyes need a decent size target for picking, I use an adjustable screen-size bill-boarded quad for points (vertices) in a mesh editor. In the pix below, the screen-size of the quads is identical (~8-12 pixels I think). However, note that, when zoomed out, the fixed screen size of those points gets messy. Note, even with an orthographic projection, using a fixed-size quad for the vertices in the model on the right below, would still result in those quads overlapping when zoomed out.
[attachment=27071:fixedsize_closeup.png][attachment=27072:fixedsize_zoomedout.png]