The key point is that even with finite bit depth (as long as you dither properly), the effect of finite bit depth is easily controlled noise of program chosen spectrum. i.e. as long as your sampling isn't doing anything really dumb, the noise introduced by sampling is well below noise floor.
I've been trying to figure out whether this is true or not, and I've only been able to find that it doesn't compound: the noise from the original quantization and the noise from the second (post-resample) quantisation aren't independent, both arise from rounding the data to the same 16-bit grid. Somehow, this seems to mean the noise doesn't compound.
I wish I understood this better and at least knew whether it's true or false. I have to do more reading on it.
