I figured out the sine waves noises don't merge into one another because they don't affect the past. And at 1kHz amplifier setting. The noises of 50Hz vs 900Hz is identical because making it pulse faster (higher frequency below 1kHz) doesn't produce more noise. And I want to test this.
Sorry, I can't follow your thoughts. What do you mean with "noises of 50HZ vs. 900Hz"? You have noise, and you have your wanted signal. Just consider them as two independent signals. At the end, noise and wanted signal simply add up. The noise does not change if your wanted signal changes. You don't get a "different noise" if your wanted signal is 50Hz or 900Hz.
Nevertheless keep in mind that your noise is not white, but your amplifier also suffers from 1/f noise at low frequencies, and that you deal with bandwidth-limited noise (~1kHz). Both imply that your noise is not independent and identically distributed, but it is autocorrelated. So the deformation of the waveform due to noise (if you zoom-in) will definitively look different for a 50Hz sine wave signal and for a 900Hz sine wave signal. I have attached example plots for 1kHz-bandlimited (2nd order Butterworth) white noise and 12dB SNR. Note that for the 900Hz signal, the noise mostly affects the envelope. Keep in mind that these plots are still
not represenative for the noise of
your amplifier.
To get a more representative picture, can't you record the noise floor of your USBamp (e.g. with Audacity, as you already did it with your other amplifier) and then use Audacity to add an artificial sine wave signal to the recorded noise in order to see what you would get?
Supposed my signal is 1V and I want to convert it to 10uV.
That's a factor of 100,000 or in other words 20*log10(100000) = 100dB.
What button or number should I press?
The enabled stages simply add up. Example: If you turn on 20+16+2, then you get 38 dB.
That is an option but the step attenuator allows to make many different levels quickly which is handy while experimenting.
The OP was very much focused on accuracy. The chosen step attenuator is likely not a precision device. I wonder what total accumulated uncertainty can be achieved for the involved components (generator's no-load voltage, generator's output impedance, 8 attenuator stages, and the terminator impedance). I have doubts that we are still within 1% at the end.