> Turns out rounding errors in floating point operations can propagate to a point where they produce this distinct, "metallic" sound.
Reminds me of a Karplus-Strong synthesis implementation that produced a gorgeous guitar/mandolin sound, but only for delay durations that weren't simple ratios with the given sample rate. The simple-ratio durations would end up sounding like crude, attenuated periods of noise-- metallic sounds like you'd expect from a pitch produced in a KSS demo. Everything else had some kind of subtle interpolation error that ended up shaping the noise just enough to make it sound like a million bucks.
The problem with most KSS is that the filter used will typically saturate the timbre. So rather than hearing a guitar string, you're hearing a guitar-adjacent interpolation scheme whose prominence makes you wonder just how un-guitarlike the original unfiltered sound must have been.
That's really hellish, considering 44.1k is divisible by numbers 1-10 and more.
Now that I think of it, the problem may have been independent of the sample rate. It may have just been all periods that were integers. So A440 sounded like a guitar string because it's a period of about 2.72ms. But at 1000Hz you're suddenly transported to a sound-effect from a 1970s episode of Dr. Who.