I did a quick experiment to try to confirm this. In REW, I configured the generator for a 997 Hz -10 dBFS sine and used the REW functionality to save the generator output to a 32-bit float WAV file. I then opened that WAV file in the RTA - basically using REW to analyse its own test signal, to act as a control/reference. And then in the same RTA window I analysed the same signal but this time being played through WASAPI loopback. The two traces pretty much line up on top of each other, suggesting that the Windows audio engine pipeline is either bit-perfect or very, very close (at least up to the loopback mirroring point):
View attachment 205151
In both cases SINAD is about 136 dB which sounds about right for 32-bit float (23-bit mantissa * 6).
If I do the same at 0 dBFS then CAudioLimiter rears its ugly head as expected:
View attachment 205153
Volume doesn't really matter here, because Windows uses
hardware volume control anyway for devices that support it (which is
most of them). Meaning the master volume control for the device only affects the hardware volume control and has no effect on the bitstream flowing through the Windows audio engine. (Obviously this doesn't hold for per-app volume controls.)
Right. I suspect it might pass such a test though as long as the test signal doesn't get loud enough to trigger the limiter.