I looked at the "hockey-stick" lines and they seem arbitrarily drawn--without them you could easily fit a single negatively sloping line. Without the data I can't say which model of the data minimizes error more, however.
In the last graph, I had made the difference quite dramatic, using an earlier version of the code. But mainly it's an exercise to show that in any event, a linear trend is expected--since a line would exist any time there's a difference in party voting preferences. If Democrats are 5% more likely to vote Trump than downballot Republicans, while Republicans are 10% more likely to vote downballot Republicans than Trump, a line would exist in the data always. It will only be flat, as he asserts it should be, if those preferences are identical!
I will edit the graphs to agree with the earlier simulated model though to make it more clear, though, thanks for the attention to detail.