Merged and pushed. issue635 deals with using the Fact struct in the remaining 
places.

Here is the plot: http://ai.cs.unibas.ch/_tmp_files/seipp/issue621-v1-total_time-cegar-10K-original.png

It seems that the CEGAR code has very nondeterministic runtimes.

Fine with me. One thing I noticed was that the total times for the CEGAR configs
can be very noisy (e.g., airport/p09: 15s vs. 19s, airport/p17: 398s vs. 364s).
This seems to be in both directions and cancels out more or less overall. I
don't think this is caused by this issue, so this should not be an argument
against merging. I would be interested in a relative scatter plot, though, to
compare against the noise we saw in issue627.

Thanks for the review! I ran an experiment. Here are the results:

http://ai.cs.unibas.ch/_tmp_files/seipp/issue621-v1-issue621-base-issue621-v1-compare.html

Nothing surprising as far as I can see. Do you think we can merge this, Florian?

I made a pull request at 
https://bitbucket.org/jendrikseipp/downward/pull-requests/42 

Florian, do you think this can be merged?

The POR code will introduce the Fact struct (issue499). Therefore, this issue 
could be concerned with using the Fact struct in search/tasks/, task_proxy.h and 
abstract_task.h.

We want to add a Fact struct with two int members "var" and "value" to 
abstract_task.h to produce more self-documenting code.