There are two issues here; first, if you are viewing your image at less than 100% zoom in Photoshop, any change in layers (flattening, opacity adjustments, etc.) are not truly accurate at those non-1:1 viewing zooms. In Photoshop CS4 with OPenGL GPU viewing if dithered using bilinear but if you don' have the GPU (or are using an earlier version) the dither is done using nearest neighbor for the dithering. What that means is that in particularly high frequency textual detail will look a lot different at 100% zoom and downsampled to smaller screen views. This applies to flattening in particular...
The other issue is that when you merge visible layers, if you take the capture or output sharpening layers and simply merge just those layers you will not get an accurate merge because of the nature of the way Photoshop merges layers. In order to get an accurate merge you actually need a complete copy of all visible layers inside of the merged layer group otherwise you'll get an inaccurate merged result. This has been a Photoshop "issue" since layers have been around. The problem being that the result of full stacks of layers depend upon the total layer count and blending modes and opacities.
So, you'll need to explain very carefully what it is that you think you are seeing.
It's difficult to explain
Let me try another example and this is where it also stems from. I discovered this while using one of your techniques prior to printing while soft proofing. Everything looks wonderful just how I've prepared things to look.
Lets say my top layer is a curves or layers adjustment. And while soft proofing, I notice that my blacks are not where I want them to be anymore. So I employ your technique where I select a color range of black with a fuzziness of 25. In order to properly select the color range and create the new layer I have to ctrl-alt-shift-e all visible layers to a new layer so that I can select from this layer before I ctrl-j my selection to a new layer where I change the display mode of this layer to multiply.
So what I am noticing is that when I ctrl-alt-shft-e to create my new layer of all visible layers, something is changing.
Take this example.
Notice the highlights in the in the tips of the trees, these highlights were exemplified through the capture sharpening process and whatever modifications I may have made, when I merged all visible it was as though some of these steps were not being captured into the newly merged layer (or so it seemed when viewing at say 25%). If I enabled/disabled the visibility of the newly merged layer what I am seeing will actually change. It's as described above, some of my edits are not being captured into this copy of all visible layer.
If you need me to clarify on any of these steps just let me know.
Dave