While I only looked at the performance implications back when, here are my 2¢:
As far as I can tell, all the problems you describe boil down to "In order to make things look good at arbitrary resolutions, we want to round to device pixels."
As TFA shows, with iPhone 6+, you can't. Round to device pixels, that is. I wrote about this when the 6+ came out[1], it's good to see it confirmed empirically (awesome job by Ole, incidentally!).
So there no longer is a choice between your options 1 and 2, because #2 is out the window, no integral scaling available.
Of course, to make #1 work properly, what they need to do is remove all the places where snapping to device pixels is done, because otherwise you get pixel cracks, and those are noticeable.
The anti-aliasing effects are apparently theoretical at this point: yes you can create test cases to show them, but in practice the resolution is high enough that it no longer matters. From TFA:
"I am actually surprised how little of an issue the automatic downsampling is in practice. As I mentioned, I simply don’t notice anything of the effects I have illustrated here in real life."
Or as Jon Gruber put it: "Its 401 PPI display is the first display I’ve ever used on which, no matter how close I hold it to my eyes, I can’t perceive the pixels. "[2]
I do think it makes more sense to do the scaling directly in CoreGraphics, disabling all the features that make snapping to the pixel grid possible and letting the pixels fall where they may. It would certainly be less work overall and use less memory. I can only speculate as to why it wasn't implemented this way, my guess would be that getting all the pixel-snapping out is not that easy, and possibly it was a rush job (as suggested elsewhere) due to last minute unavailability of the higher resolution display.
In the future, I would expect either (a) an actual 3x display or (b) the display stack adapted to your (1) choice.
As far as I can tell, all the problems you describe boil down to "In order to make things look good at arbitrary resolutions, we want to round to device pixels."
As TFA shows, with iPhone 6+, you can't. Round to device pixels, that is. I wrote about this when the 6+ came out[1], it's good to see it confirmed empirically (awesome job by Ole, incidentally!).
So there no longer is a choice between your options 1 and 2, because #2 is out the window, no integral scaling available.
Of course, to make #1 work properly, what they need to do is remove all the places where snapping to device pixels is done, because otherwise you get pixel cracks, and those are noticeable.
The anti-aliasing effects are apparently theoretical at this point: yes you can create test cases to show them, but in practice the resolution is high enough that it no longer matters. From TFA:
"I am actually surprised how little of an issue the automatic downsampling is in practice. As I mentioned, I simply don’t notice anything of the effects I have illustrated here in real life."
Or as Jon Gruber put it: "Its 401 PPI display is the first display I’ve ever used on which, no matter how close I hold it to my eyes, I can’t perceive the pixels. "[2]
I do think it makes more sense to do the scaling directly in CoreGraphics, disabling all the features that make snapping to the pixel grid possible and letting the pixels fall where they may. It would certainly be less work overall and use less memory. I can only speculate as to why it wasn't implemented this way, my guess would be that getting all the pixel-snapping out is not that easy, and possibly it was a rush job (as suggested elsewhere) due to last minute unavailability of the higher resolution display.
In the future, I would expect either (a) an actual 3x display or (b) the display stack adapted to your (1) choice.
[1] http://blog.metaobject.com/2014/09/iphone-6-plus-and-end-of-...
[2] http://daringfireball.net/2014/09/the_iphones_6