Its strange. Shortly after the fix was ready the first generated Core had no more such noise.
Or I could not hear/detect it.
It seems more, the Sigma/Delta converter is still instable.
Yeah, I just wanted to be sure it wasn't just a quirk of Gods before chasing the problem, but I can't reproduce it using WinUAE, whereas I *can* reproduce in WinUAE the effect I thought might be causing it, by fiddling with the sample repeats in ProTracker. So I agree, it's something to do with the Sigma/Delta converter.
I wonder whether the multiplexer that feeds the sigma/delta converter might be the culprit? Currently, if I'm reading the code correctly, the input of the sigma delta converter is swapped rapidly between two different sound channels, and that happens at the same rate as the sigma-delta modulation itself. [edit: that's how it used to work - the latest version with the audio fix has a nicer implementation which still multiplexes the volume multipliers but which sums their outputs on alternate cycles, so the input to the SD modulator should be stable.]
The first thing I plan to try is changing this multiplexer to an addition, so that the input to the modulator is stable from cycle to cycle. I think the point of the multiplexing is to share the volume multipliers between pairs of channels, but we have multipliers to spare, yes? [edit: Tobiflex has already taken care of this, in such a way that extra multpliers aren't needed! Nice!]
Even if doing that doesn't fix the problem, it'll make trying out Frenchshark's 3rd order converter easier, and will be necessary for box-filtering, too.
The other thing I shall try, just to prove the source of the issue, is to set up Paula so that the output is always zero if the sample repeat length is set to minimum. It'll probably break some chiptune or other, but it'll at least help identify what it is about Gods in particular that causes the issue.
Whole synthesis changed by altering the source somewhere else, so this error noise could reappear by adding any other feature or fix.
Oh great - a build-to-build instability problem! That's always fun!