The softness effect is definitely noticeable between 720p and 1366x768 panels even on 3d and video content, though.
Yes, when the resolutions are almost identical, scaling is going to result in a similar-but-less-good image, whether it be ringing or blur or blocking that does you in. This is an argument for staying native, not for using nearest neighbor.
You simply can't add detail that isn't there
I'm not suggesting that we try to. I'm suggesting that we try to preserve the details that
are there as accurately as possible.
Nearest-neighbor may emulate the look of some 720p or 1080p TVs, but this is only because said TVs use poor image reconstruction. Square pixel fill can occasionally accurately reproduce an edge, but the vast majority of the time it will take the form of a quantization artifact; most edges don't align with the pixel fill edges, and even those that do will usually not lie very close to the pixel boundaries.
Who knows, maybe you're right in this case. In that maybe weird complications make it fairly difficult to
strictly out-do the quality of a square kernel when you're only going up a factor of two on each axis. After all, you don't have a whole lot of options for filter shape unless you're willing to use a really wide function (though I wish I had MATLAB on hand, it would make it super easy to try out some 4-by-4 patterns I have in mind).
But on something like a factor of three or four (or more), where you can more easily implement a filter shape that's tight enough to not cause blurring but also doesn't have the directional preference and quantization issues of nearest-neightbor? No way the square approach still looks attractive.