So I made a little test circuit of the 'Resistive divider' schematic I shared in post #1.
It was powered like this: SMPS > LM317 > resistive divider circuit > AC coupled audio line input of computer.
In this way I can compare the noise level of the shorted line input and one fed from the bias supply. The audio input has a 2V rms input sensitivity.
First is the shorted input: -111dB noise floor.
(Click to enlarge)
Resistive divider with 10uF cap and op-amp buffer: -110dB noise floor. Slight mains harmonics.
I then tried connecting directly to the resistive divider without 10uF cap or op-amp buffer (i.e. LM317 simply followed by a divider): -96dB Noise floor. Wow, that is a lot worse!
This shows me the op-amp with the 10uF capacitor helps greatly to reduce audio band AC muck. It also shows me the cleaned up bias is at least good enough not to affect the noise floor of my 2V rms audio interface line input.
FFT of course averages away random noise, so it is difficult to say if the slight increase in noise floor with the bias signal is due to those 50Hz mains harmonics or random AC noise.
It seems to me I should focus on the filtering at the op-amp input to get something super clean, rather than worry about the source of that voltage (i.e. voltage reference, high performance regulator or just resistive divider). Perhaps I'll do a 2nd order filter to be confident about it.