The point is that the high resolution mode doesn't seem to work as you would expect. The DS1K doesn't seem to be adding the 1LSB of Gaussian noise needed in order to achieve higher than 8bit resolution. All it seems to achieve is a slightly less noisy 8bit signal, negating any benefit of having >256px vertical display resolution....
...but Rigol are being cheeky to claim that box car averaging alone is 'high resolution mode'.
I think you're confused between how you expect High Res mode should work, how it's actually supposed to work, and how well the DS1000Z achieves it's results.
Rigol is neither the source of claims about averaging - nor the company that first named it High Res (I believe that honor goes to Tektronix - see image below from 2003 Tek paper on acquisition modes).
There are papers detailing the effects of successive sample averaging (also called boxcar filtering or moving average filtering), such as:
http://www.daysalive.com/share/EffectsofAveragingtoRejectUnwanted Signals.pdf
...but the main points are that it removes noise by decreasing a DSO’s bandwidth (it applies an LPF function with a -3dB point approximated by 0.433*sample rate/averaged samples), achieving 1 bit of effective resolution improvement for each factor of four of averaged samples (e.g. 4 bits improvement = 256 averaged samples).
The DS2000 does a good job of implementing 256 sample averaging in real-time (functioning almost identically to the Agilent X-Series), but the DS1000Z does not - I'm guessing because the engine is not fast enough to support it (except perhaps at very slow timebases). The difference between the two DSOs is quite apparent, so the problem isn't with the technique but the implementation on the DS1000Z.
I know how 'high-res'/oversampling works. The issue is that the DS1000Z is doing something wrong because there is an apparent lack of increased resolution when the high res mode is turned on.
You can only achieve increased resolution if there exists 1LSB (or more) of Gaussian noise on the signal. If the signal is not sufficiently 'noisy' within the window of the averaging filter, it won't work properly. For example, you feed in a DC signal that falls somewhere between quantization levels '232' and '233' of the 8-bit ADC, but the ADC outputs '233' consecutively since the signal is closer to 233 than 232. You apply box car averaging to a long string of 233s and you still get 233, the result is no increased resolution.
Now if you add 1LSB of analog noise before the ADC, the ADC will output out a series of samples which alternate between 232 and 233, with a distribution that when averaged will give a value that is in between 232 and 233, thus increasing the resolution beyond the number of quantization levels of the DAC.
If the DS1000Z isn't adding 1LSB of noise internally, then it is obviously not going to guarantee increased resolution in every circumstance. I haven't been successful in getting significantly increased resolution out of my DS1000Z, even when attempting to add external noise there are still obvious 8bit quantization levels present (it jumps up/down by 2 pixels in places instead of 1). I'm not 100% sure why this is. I suspect the result might be rounded back to 8bit somewhere before it is rendered on the screen. The high res mode is obviously doing some averaging, it just doesn't seem to increase the resolution. The intensity grading also effectively does some averaging (but not in a way that would increase resolution), which makes it even harder to observe what is going on.
On some vertical ranges there is even less than 8bits of resolution displayed on the screen. Every time you hear a relay click, you are changing the real analog range being fed into the ADC, so certain ranges have 8bit and some are effectively 7 or maybe even 6 bit. It wouldn't be such an issue if the high res mode worked properly!
Tektronix scopes implement high-res correctly and consistently achieve increased resolution that pretty much matches the display resolution, even for noiseless DC. I'm not sure how/if the DS2000 does it because I don't recall seeing anyone demonstrate it and haven't had the chance to get my hands on one.
You do not need 256 sample averaging to increase the resolution. Even if the 1000Z can only do 16sample wide windows, it should still achieve 12-bit resolution. 256 sample window would achieve 16bit. It really only needs 9bit since the display isn't even 512pixels tall...
That said, complaining that a $400 scope can't do high res properly would be a bit rich, i just think that 'high-res' is an inaccurate description of what it is doing. 'Noise reduction' or 'HF reject' would be a better description.