Yes, thanks! I closed that one. The discussion there also said that the
discussion of issue214 was relevant to this. I didn't go through it, but wanted
to copy this information over here now that issue423 is closed.

Is this a duplicate of issue423?

Perhaps just a small (five-minute, or even only one minute) experiment with
LM-Cut, and one of the main merge-and-shrink configurations to be sure? I also
often run blind search because there runs tend to be not too noisy.

The first draft of the implementation is finished and experiments for
inadmissible heuristics (especially cea and cg) are in the queue. 

What optimal configurations should I test? None of our good heuristics supports
derived variables, so the main concern will be to verify that I did not break
anything for axiom-free tasks. Since my implementation does not change the
variable order, the only expected effect is a 4KB larger memory usage.

To save memory, we do not want to store the values of derived variables anymore
but instead recompute them each time a state is unpacked. As a side effect, the
preprocessor will no longer fix a total variable order.