To prove that modern DSOs do in fact alias on long timebases:
Attached images show Rigol DS1074Z aliasing with 50MHz input signal on long timebase.
The "Anti-Aliasing" setting does not prevent aliasing. (Not entirely sure what it does other than slow down the scope and make the render look prettier: they probably should have called it "high-quality" or something similar.) Hi-res mode also shows an aliased signal: whatever hi-res averaging Rigol is using appears to be post-decimation. Peak detect does prevent aliasing.
OK, if you use the memory limitation to force the scope to reduce its sampling rate you should expect aliasing. The fact that the scope reduces its sampling rate by decimation rather then slowing the ADC clock is just the way it works--the results wouldn't change except maybe to the extent that a slower ADC might have a longer sample window.
The OP has proposed:
Data for time domain display is downsampled to sample rate appropriate to timebase setting using a folded LP filter
which I had interpreted as meaning that the LP/downsampling operation would be applied between the memory and the screen (post acquisition) not between the ADC and memory. However, if the folded filter can be applied in real-time, what would that look like?
Which of the three options posted here is the most 'correct' response--the one you want to see--to a signal at twice Nyquist of the posted sample rate?
If you apply the appropriate LP filter for a 20mS/div display to a 50MHz signal, what is the appropriate result? A flat line?
In this case, an LP filter might be the appropriate default. Aliasing has to be prevented by limiting BW
before you impose the Nyquist limit by sampling or downsampling as the case may be, so if it can be done in real-time, great--especially if the user can have explicit control over the filter parameters. However, I think it will add some complexity that will be hard to justify in an entry level product. The FFT will need its own data channel and memory.
I cannot see how this can be claimed to eliminate the need for peak detection. I also don't have any high-end DSOs so I have no idea if any of this has already been implemented somewhere. Anyone?