I have a slide_tile problem that is solvable by lazy_greedy([hff]) and 
astar(blind()), but not by lazy_greedy([hcea]). I believe the heuristic is non-
admissable, but it should still find _a_ solution, as [hff] does. I have also 
tried with reopen_closed=true but no change.

I'm not familiar enough with the codebase to be able to figure out exactly what 
the issue is, but I thought you might like to take a look into it.

I have attached the domain/problem and the generated sas_file, as well as output 
from each of the above runs. Hope it helps!

Ahh, I didn't see that page, thanks for pointing it out!

Ye it would be great if it would mention something, maybe even just print that 
table when starting up to show whether it is admissible/safe etc. Maybe with an 
option to quiet so those who know don't get spammed, but new comers are 
informed...

Thanks for the detailed response!

This can happen with the cea heuristic (and also with its relative, the cg heuristic). These are both known to be "unsafe" heuristics, i.e., they can assign infinite heuristic values to solvable states. Such states will then be pruned by the search, and therefore greedy search with these heuristics is not a complete algorithm and can fail on some solvable instances.

There is some brief mentioning of the safety property near the top of https://www.fast-downward.org/Doc/Evaluator, and section https://www.fast-downward.org/Doc/Evaluator#Context-enhanced_additive_heuristic mentions that cea is unsafe.

I understand that this is very sparse documentation, and hence this behavior can be surprising. (It is mentioned somewhat more prominently in the papers that introduce cg() and cea().)

In the future, it would be great if we could make the planner more introspective, so that it could tell explicitly in the output in cases like these that an incomplete algorithm was used and "no solution found" does not imply "no solution exists".