Don’t know why we’re bothering talking about IMD when the track is clipped. Of course you’ll hear things.

Shaped dither is meant to move more of the energy of the noise into the upper frequency bands, and thus be less potentially audible to most folks. Triangular dither has a noise distribution (so how the random values of the dither are spread around the true sample value) that is more often close to zero than a uniform spread, though I don’t know why it would be superior to Gaussian.

Second the vote to just keep it at 24 bit unless space is a real concern. If it is, shaped 16-bit should do fine.

You can think about it in terms of dB. 0.15% is about -56dB, so that means the distortion products they measured were 56dB below the signal. That isn’t likely to be audible, especially if the distortion is in the harmonics we find “pleasing” like the 2 and 3.

92dB RMS for two years‽‽ I hope that was all bass.