The real answer is that there is an industry of designers
accustomed to doing per-pixel tuning of their art assets and
this work is worthless if the assets are going to be scaled
anyway.
That's an effect, but not the cause – the roots of the raster-based paradigm of modern GUIs are far more technical. I would point to the tradeoffs between speed and memory limitations in early GUI systems: If you look at the original Macintosh UI, you'll see a system more constrained by rendering speeds than by RAM, and thus a high emphasis on bitmaps versus vector rendering.
Look at this Folklore.org story on the origin of the roundrect in System 1: http://www.folklore.org/StoryView.py?story=Round_Rects_Are_E... It was only through a combination of Jobs' perfectionism and Atkinson's ingenuity that such a simple thing as rounded rectangles could be accomplished on the original Macintosh; can you imagine how long it would have taken to render a Finder window full of vector icons?
And yet Linux machines have been drawing SVG icons and SVG widgets for years. This limitation doesn't exist any more.
Look at Display Postscript, which was created by NeXT in the late 80s and carries through to much of Apple's Quartz rendering system today. Why are we still limiting ourselves to pixel graphics?
Look at this Folklore.org story on the origin of the roundrect in System 1: http://www.folklore.org/StoryView.py?story=Round_Rects_Are_E... It was only through a combination of Jobs' perfectionism and Atkinson's ingenuity that such a simple thing as rounded rectangles could be accomplished on the original Macintosh; can you imagine how long it would have taken to render a Finder window full of vector icons?