If a measurer is faced with this issue, he has 3 choices:
1) Use a shorter course
2) Use a course that includes some hill
3) Use a remote course
And the question is, which poison is the one to pick? Personally, I will be wary of choosing #3.
In my experiment, the temperature increased 4F during the 90 minutes it took to do my 48 rides, although the sun was out so the pavement temp probably increased more (it was about half shaded). Here's a plot of the cal constants of all rides as a function of time.
The red line is a least square fit and shows that the cal constant changed about 4 counts (0.033%) every 30 minutes. Even if I live close by, I cannot drive to the course, drive the course (which I always do), warm my bike back up, and get started on my measurement in less than 30 minutes. If it ends up taking 60 minutes, the resulting error is 2/3 of the SCPF!
How much error does a cal course with a hill cause? I don't know. I plan to do another experiment to find out.