Thank you for the reviews. I merged the code and pushed.

Yes, looks good to me, too.

The code looks good to me.

The experiment is done. It compared the generated output files (for technical
reasons I used the sha512 hashes of output.sas.xz as a proxy). There were only
four cases with a difference and they were all from the trucks domain where the
discovered invariants depend on the time limit.

I think this is ready for review.

I created a pull request on bitbucket:
https://bitbucket.org/FlorianPommerening/downward-issues/pull-requests/37

Experiments to check if this influences any existing benchmark are running.

I attached the pddl files where the error occurred and a simplified version of
them with only the toggle-light action.

In the simplified domain there is only one predicate (lights on/off, initially
off) and one action (toggle the light). The goal is to switch the lights on and
blind search finds no solution.

The translator simplifies conditional effects for binary variables in the method
prune_stupid_effect_conditions in the following way:

IF v=0 THEN v:=1 is simplified to v:=1
and
IF v=1 THEN v:=0 is simplified to v:=0

This is problematic if both effects occur (for example in a
"toggle-light-switch" action) because in this case
  (IF v=0 THEN v:=1) and (IF v=1 THEN v:=0)
is simplified to
  (v:=0 and v:=1).