Flipping the switch triggers metering. Since evaluative metering invokes code that looks at multiple metering points and tries to intelligently guess what to do, it's somewhat reasonable to assume that it might a different decision on what to do each time it's invoked if the light entering the lens in non-uniform across the image.
A way to test if this is happening is to point the camera so the image is completely uniform in brightness across the image — at the sky, or a flatly coloured wall, etc.
Please bear in mind that I don't own this lens or camera, but I do have experience in programming. In algorithms like this, it's often the case that there's some guesswork involved — hell, I once implemented an algorithm that would just randomly choose one of tho values if it couldn't determine one value over another with any degree of certainty, and it worked just fine.
Since I doubt the camera saves the reasoning behind its metering decisions between each metering, I can easily see this sort of thing happening in something that's supposed to be "smart" like evaluative metering. If you require absolutely consistent metering, I guess evaluative isn't the right mode to choose.