I’d start with the depth of focus of the lens, and make the separation >2x this to start. 

Thanks for the insight, Mark. 

Is this the usual rule-of-thumb you would use for such a situation?

It’s a good starting point, but I’d follow up with some simulation. Try Geometric Image Analysis, and put a slide surface (a .bmp or .jpg) with a single pixel turned on for more detailed analysis

As long as the window is not placed at an intermediate image plane, any defects won’t be “imaged” onto the sensor.  

The more important thing to consider is ghost images and stray light.  For direct imaging of ghost reflections, you can use the Analyze > Stray Light > Ghost Focus Generator along with “Ghost” MFE operands (GPIM, GPRT, GPRX, GPRY). 

Any contamination, particles, or defects on the window will lead to scatter stray light, which you will need to model in Non-Sequential mode.  

Remember that the sensor itself typically has a cover glass and if this objective is designed to be swapped out with other objectives, then you might be limited from a mechanical standpoint as to where the window will be placed.  From a mechanical standpoint, the objective protective window is typically placed closer to the tube lens.