That is a bit strange in so far as LAMA shouldn't be affected -- it never used
0-cost operators in the heuristics. But in any case, looks good to me. Feel free
to close this if you're also happy with it.

Here is the evaluation data.
BJOLP is not different at all.
LAMA has some differences, most notably in parcprinter.

(Doesn't have to be a huge problem suite, but I'd like a large selection of
domains -- say, the first five instances of all IPC-2008 domains with
seq-opt-bjolp and with seq-sat-lama-2011.)

Merged. I'd prefer if this were given a proper test on a larger set of
benchmarks to verify that it doesn't misbehave. Can you do that?

Fix is in downward-fixes, branch issue252

Thanks, Erez!

For what it's worth, maybe the whole LM code needs a good check that 0-cost
landmarks are handled properly. I found another thing in the same file:

    // BUG/TODO/FIXME: This first test likely does the wrong thing in
    // the presence of zero-cost landmarks.

I can take care of this

The following code in LandmarkCountHeuristic::compute_heuristic is wrong:

    if (!use_preferred_operators || h == 0) { // no (need for) helpful actions,
return
        return h;
    }

The intent of the "h == 0" test is to skip computing preferred operators if
we're already at the goal, but in domains with zero-cost operators, h == 0
doesn't imply that we're at the goal.

One way to fix this would be to test if we're at the goal; maybe a cleaner way
would be to adapt the following code so that it works whether or not h == 0.

Any takers?