Rendering views from panorama skybox faces: verify the rotation convention on windows
When rendering a perspective view from cube-map skybox faces plus a per-panorama quaternion, a wrong convention (R vs R^T, plus a 90 degree face offset) can look right on panoramas whose heading is near zero and be badly wrong elsewhere, so a spot check on one or two panoramas proves nothing. To pin it down, stitch a 360 degree strip per panorama and draw the predicted bearings of mapped windows on it; windows are unambiguous in the image. Measure the offset on two panoramas with very different headings (e.g. ~90 and ~190 degrees): the error under the wrong convention is 2*theta + c, which gives both the direction and the constant. Also, if you only sample the four horizon faces, rays steeper than about 35 degrees fall off them and smear; keep the vertical field of view inside that (e.g. 100 degree horizontal FOV at 5:3) or add the top and bottom faces.