I assume you have a system to sync the cams?

I would keep the cameras parallel and fix them ~6.5 cm apart. You can always slide the L/R images later to fudge the convergence. You might need to do that if you see annoying artifacts when objects that you are focusing on break the frame.

If the inter-occular distance is wider than ~6.5, objects will appear smaller than they really are, and vice-versa.

Try to avoid objects that break the frame at different depths if the audience is likely to be focusing on those parts of the image. I did some fireworks in stereo and the director said its fine for stuff to break the frame so long as you let off more fireworks in the middle to take the audience's attention away from those parts.

People's eyes can't follow objects coming towards them too fast. The muscles aren't that quick. They can handle objects moving away quickly better, the muscles can relax (back to parallel) faster. So, for "3d moments" its better to bring the object up to the nose relatively gently, and then it can shoot off into the distance if you like.

If you are adding CG you, you should be able to track just one eye and position the other cam relative. Its going to be very important to scale everything reclaimed from the track (camera feature points and matched geometry) to world units. If you do make left and right tracks it might be easier to align the features from both in order to get the scene scale right.

If you are putting CG in with maya, you can use maya's stereo cams with an image plane for each, and view in stereo in the viewport with red/green glasses. Its ok for getting the depth right but can be confusing if you have colorful or highly specular elements, or if color changes with facing ratio. Make sure the convergence is set to parallel in the maya cam. Then you can render and comp.

Its a nightmare to do paint fixes on stereo images in comp. Try to avoid that. Nuke handles stereo channels nicely. Not sure about other comp systems like AE or shake.

There are a number of stereo cheats you can do in comp, usually by moving corresponding pixels further or closer. For example, if you want to drop the ground lower or bring it up, make a B/W ramp from horizon to near and displace the pixels of each eye left and right.

We have a realD TV at work. There's a bit of software from nvidia on a pc that loads the L/R frames in and sends stereo to the TV. Not sure what its called, but that part of the workflow is straightforward and similar to setting up playback on other stereo systems, like framecycler with dual projectors.

There's plenty more stereo-specific stuff that will be up for discussion I'm sure - DOF, transparency color grading, stereo subject matter...