This is how I output 512-width lines and blend them to create the top-left image. It averages the next pixel to output with the previous pixel, then saves the true current pixel for the next loop iteration. Feel free to use something similar, or not.
A side effect of applying a lowpass filter to the entire screen (which is what you're doing) is that the parts of the screen that are fake-blended using pseudo-hires appear sharper than the unblended parts. Look at the fence posts near the left and right edges--the part of the post that's "behind" the text window has pixel-sharp edges, while the part that isn't is blurred.
ETA: the difference between the flames in the blended and non-blended parts of the upper left screenshot is even more obvious.
Also, it would be nice if you posted a screenshot of SD3 under your renderer for comparison.
