I'm not 100% sure about the patch - can't be as I did only test my error case. Here it did remove the crash and, furthermore, did not add any other. As I see it the following work flow is done:

srv_running()
- begin_turn()
-- new initialisation to ai/adv phase
- loop (phases)
-- begin_phases()
--- initialisation to ai/adv phase; if initialised, do nothing
-- (some code)
-- end_phase()
--- end of ai/adv phase
- end (loop phases)
- end_turn()
-- in this function the ai/adv phase can be closed to have the correct count of begin/end; these functions will do nothing at the moment but as you said, would be helpful to read the code

I did read bug #18333 and at least the reason I found is similar. The signal 'turn_started' is called in begin_turn. At that time ai/adv phase is not open for any player. It is initialised in begin_phase() in srv_running(). It crashs on contact to a player which exists before (the new player has the phases initialised).

I did not test any other but the default phase setting. I I remember correctly I never did test the different possibilities of this setting.

I did some tests and I do use a unique setting: civil_war() is called via the lua signal 'turn_started'. At the time this signal is raised, neither the advisor nor the AI are initialized. Thus the core dump. The attached patch fixes the error by doing the initialization before ...

Please test ...

(file #13791)

civil_war() was designed to be called during turn change (ai phase for all players). Now, it can be executed via a lua function at any time. Thus, the ai phase does not have to be initialized. One possible solution would be to check for this if a civil war occurs. Thus, the ai phase would be initialized for all players, the civil war executed, and after that the ai phase would be closed. Thoughts?