Alright, thanks to spades's willingness to do the mindnumbing work, I think we can confirm some of how this works. Check out the graph from his combat log (much cleaner than mine in terms of lag):
So, what we have here is actually really close to the original statement regarding how parry works. In other words, my understanding of it was pretty darn wrong. I thought I was going to find something new, but the data doesn't bear that out:
When parry gets a % reduction, the swing timer is reduced by 40% of the orginal weapon speed, except in the event that it would take the weapon speed faster than 20% of the total weapon speed. That's why the slope of the line before the flat part is 45 degrees. Each one of those points is 3.6 (weapon speed) * 40% = 1.44 faster than the swing otherwise would be.
We see the cap at .72 seconds (20% of the weapon speed).
The interesting parts, which were what threw me to my initial suspicion of the parry hypothesis being wrong is that stuff that you see when the non-parry swing timer is already lower than 20% of the original weapon speed (which is why i assumed the 20% cap hypothesis had to be incorrect).
Apparently at that point the game realizes that you can't cap a swing at a longer time than it should take, and a new rule of parry mechanics takes over which simply says "hey, parry can't reduce you below 20%, so just let the swing land like it's supposed to".
So, there we have it, in a neat little bundle - The original theory was right, and I feel a bit dumb. But at least now I'm convinced of the original theory. I suppose someone could try to figure out why there are trace bands of extra haste in little bits, but the overall picture is pretty solid now imo.
The next step is to verify what happens parry wise when dual wielding - is it next swing haste, or always the main hand?
Again, super big thanks to Spades and his sacrifice at the hands of a warrior who thought it would be cool to gank someone doing testing :-p It made the graphs extremely easy to read.