Sounds good. :-) Closing this.

The short version:
After a bit of debugging, I was able to come to this conclusion: the uniform 
cost partitioning in the new version is better, even without action landmarks. 
To be more exact - my old uniform cost partitioning had a bug, and everything is 
fine with the new version :-)

The long version:
I noticed that on blocks-9-1, even when disabling action landmarks, there was a 
difference in the initial state heuristic (14 vs. 16). I tracked this down to 
the following problem with the old uniform cost partitioning:

If X is a landmark, and A is an achiever of X, then X is assigned a cost of 
cost(A) / number_of_landmark_effects(A).
number_of_landmark_effects(A) considers the status of landmarks (only considers 
landmarks that are not reached), but does not consider if A is a first achiever 
or a possible achiever. So that at the initial state, if A is a first achiever 
of X, and a possible (not first) achiever of Y, then X would be assigned a cost 
of 0.5 instead of 1.

Done as part of the merge of the emil-new branch in issue122.

I made a very informal test of this by comparing the current head of default
(369dc6a57d3f) to the current head of issue122 (c11630203a5c) with:

translate.py ../../benchmarks/blocks/probBLOCKS-9-0.pddl && preprocess <
output.sas && ./downward --search 'astar(lmcount(admissible=true),mpd=true)' <
output

For the default code, I get:
    Evaluated 550175 state(s).
    Peak memory: 893020 KB

For the issue122 code, I get:
    Evaluated 368913 state(s).
    Peak memory: 53428 KB

So even if we take into account that fewer states of course require less memory,
there is a very noticeable decrease in memory usage.

I don't quite know where the reduction in the number of evaluated states comes
from, but I guess that is more related to issue122, so this one can be closed.

The landmark information could be stored much more compactly. See changes in
r3918,r3919 and r3920 in the branch emil-new: this led to a speedup and reduced
memory usage on an example Blocksworld task from 2617 MB to 280 MB.

(This is with conjunctive landmarks though -- if we only have fact landmarks,
the improvement will be less dramatic since landmark info doesn't overall space
usage to the same extent as with conjunctive landmarks.)