This is an interesting discussion and the sample you presented is definitely suitable for most photographic analysis purposes I feel. I am not an expert on the intricacies of the JPEG compander // I am a user like most folks. However, the way I use the imaging is different from most folks, because I go down to the pixel level at times in the superfine JPGs when needed on the S50, resorting to RAW only under special conditions. Because the JPEG philosophy is in essence "compress as much as possible and keep the image looking as close as possible to the original scene when viewed by eye at a normal viewing distance," then your conclusion of the small difference between fine and superfine does indeed apply.
In this context, you asked me to make a judgement on what I would call "a floating comparison." What I mean is this: you are presenting me an incremental difference without a "scene reference" or "perfect baseline." When viewed from afar, the images look the same indeed even when reasonably zoomed (I did not examine your dark difference image not to bias my opinion). Here in my lab however, I do the comparisons differently: they are based on shooting a RAW with the JPEG and then comparing each JPEG quality image to the RAW (or a super-superfine version of the RAW converted outside the camera), rather than to each other. Then one can account for CCD pixel-to-pixel gain nonlinearities as well, especially in the low-light regions in the image. Also, the (new to me) CHDK feature allows me to turn off bad pixel subtraction, so I can identify where the bad pixels are and I was planning to work with it on the S90. But because JPG compression is very poor for my application on the S90, I am only left with the one choice unfortunately and it would not be worth doing those investigations, unless a way could be found to enable superfine if it exists in the camera.
Having said all of the above, the evidence I base my judgement on are the high-frequency component differences I detected by eye only. Also, I don't know the intensity nonlinearities of your specific CCD as I do my S50. Keep in mind that yours are two different shots as well, so there could very well be a (very small, on unit pixel order even if on a tripod) displacement component that unfairly changes the pixel-to-pixel conditions between the two images, and is one of the reasons I feel a comparison to a RAW would be more effective. Considering everything, I note the top image has the more high-frequency definition in (my assumed) contrast transitions than the bottom.
