Ok, but dither would still be needed, right?
Not by user intervention. A small amount of dithering may be added on the final stage of DSP processing without of user knowledge. Even if a math is done with a high precision, rounding errors are cumulative and depends on the algorithm designer makes a choice to add a final touch or not. Remember also that internal math operations are made in a floating point 32-bit or double precision, so conversion to the integer format always take place. When changing format or reducing resolution, adding dithering can take place. It depends on the final bit depth, by example 24-bit output do not require adding dithering at all, for 20-bit is optional, but for 16-bit mandatory, otherwise sound will be harsh.
I made number of tests with oversampling on the PC with my Audio GD R2-R11 upgraded to femto second Crystek oscilators using PGGB offline resampler which is better quality than real-time resampling in the HQP. My DAC is pure NOS, there is no DSP hardware that would alter sound, no way of cheating. I found that in the best case I am unable to recognise a difference, in all other cases sound is slightly losing natural character. With a high quality of PGGB, adding a dither to 24-bit file only deteriorate sound, while noise shaping option removes a large portion of microdetails during fast transients. If your music is computer generated, it may be unnoticed, but on the good recording of acoustic instruments I can't live without these microdetails.
In result, I play in the original file format, I don't need Hqp, foobar 2000 does everything I need in bit-perfect mode.
If upsampling to the maximum frequency helps SQ, it is most probably because the internal DSP with filters, dithering, noise shaping, or ultrasonic data scrambling for hiding ladder linearity errors (all this stuff) is bypassed feeding a ladder with pure data. It also means that a ladder is most likely not operating at the optimal frequency due to the physical transient restrictions. It would be better to have an option to defeat these features at a lower sampling rate too. It means strict requirement for a pure NOS operation.
If I were upgrading my R2R-11 to a balanced version, it would be the an upper Audio GD DAC like R8/R7HE, I am an AGD fan. Also considering Sonnet Morpheus which is offspring of Metrum at the affordable price.