Ok, I think the code is not meant to support axioms so Salomé added a check that
it is not used on such tasks. If it is, it exits with an EXIT_UNSUPPORTED error.

I looked into this. It is true that lm_hm cannot handle axioms. But the reason
it crashes is not because it thinks a subset of the goal is unreachable, but
because it tries to get information from an operator that is actually an axiom:
Axiom-indices are at least in this part of the code saved together with
operator-indices (i.e. if we have 20 operators and 30 axioms, operator_index =
20 is the first axiom). The lm_hm code just uses operator_index for array index
of g_operators, therefore if the operator is actually an axiom, it accesses an
invalid element.

I also tried to made a quick test problem to see how the code can handle
unreachable goal-subsets. The goal is to move two packages at tow destinations,
but one of the destinations is not reachable. I am not sure if it is a
sufficient test case, but the output is the following:

$ ./downward-1-debug --search "eager_greedy(lmcount(lm_hm(m=1)))" < output
[...]
Subset of goal not reachable !!.
[...]
Landmarks generation time: 0.000142168s
Discovered 1 landmarks, of which 0 are disjunctive and 0 are conjunctive 
0 edges
Initializing landmarks count heuristic...
0 initial landmarks, 1 goal landmarks
Conducting best first search without reopening closed nodes, (real) bound =
2147483647
Initial state is a dead end.
Completely explored state space -- no solution!
[...]
Search stopped without finding a solution.
Peak memory: 5060 KB

So it seems to me that the code can cope with unreachable goal subsets.

$ ../translate/translate.py ../../benchmarks/psr-large/p01-s29-n2-l5-f30.pddl
$ ../preprocess/preprocess < output.sas
[...]
$ ./downward-1-debug --search "eager_greedy(lmcount(lm_hm(m=1)))" < output
[...]
Subset of goal not reachable !!. 
[...]
=> segfault

We suspect that:

- lm_hm cannot handle axioms
- therefore it thinks that the goal is unreachable
- the unreachable goal leads to problems in subsequent parts of the code (perhaps because we don't have landmarks, but are 
accessing the first one of them)

Two things have to be fixed:

1. even if the goal is not reachable, the planner shouldn't crash
2. lm_hm has to report that it cannot handle axioms by exiting with EXIT_UNSUPPORTED