Oh I see what you're saying. You're right, since he defines the x-axis as % straight-ticket Republican voters, at 0% it should be impossible to have a negative value, since you can't have negative votes for Trump. In my notebook I use the "registered republican voters" quantity instead, which is a little different.
If you edit the code to plot the downballot votes vs. the difference, you get the same pattern, though. Check out the Gist here:
I don't think Ayyadurai assumes 100% of Republicans voted for Trump though--because in that case no negative values could be possible on his graph, but he shows a few "acceptable" examples with negative values like at 45:15.
Thanks for engaging / helping sharpen up this analysis