This is a great thread (all your threads are awesome, btw).
I draw the opposite conclusions from your results here. Your results show that, at worst, the amount of IMD caused by aliases in the passband with the "slow" filters is more than about -125 dB down -- i.e., definitely inaudible. You don't show the time domain measurements, but the steep filter necessarily has about double the amount of pre-ringing. There is debate about how much pre-ringing is audible, but to the extent it may be audible, it's more likely to be audible than IMD at -125 dB down. (When you hear increased presence with the sharp filter, I'd argue what you're hearing is a psychoacoustic effect of the pre-ringing.)
A lot of people on this forum (not you) don't seem to understand that you can't have good attenuation at Fs/2 on 44.1kHz sampled data while preserving 20kHz bandwidth without also having considerably more pre-ringing. You can't have your cake and eat it too because of the way the math works. The reason DAC designers from all of the major IC vendors have generally settled on filters with a 0.54Fs stopband isn't because they're incompetent -- it's because they realize the inherent tradeoff in time-domain performance. For telecom applications or higher sample rates, Fs/2 stopband is obviously ideal, but at 44.1Khz for audio applications there is no ideal solution, and I'd even go so far as to say that Fs/2 is more "wrong".