I started activating new program_state.saving and discovered that
saving of ball and chain could access freed memory. The change
for the former and fix for the latter are mixed together here (but
easily distinguishable).
The saving flag inhibits status updating and perm_invent updating,
also map updating that goes through flush_screen(). That should
fix the exception triggered after an impossible warning was issued
during a save operation. impossible() goes through pline() which
tries to bring the screen up to date before issuing a message.
During save, data for that update can be in an inconsistent state.
The code to save ball and/or chain when not on floor or in invent
(I think swallowed is the only expected case) was examining the
memory pointed to by uball and uchain even if saving the level had
just freed floor objects and saving invent had just freed carried
objects. So for the usual cases, stale pointer values for uball
and uchain would be present and checking their obj->where field
was not reliable.
Kicking a container that had gold in it took the gold amount
away from hero's credit or added to hero's debt, then didn't
give a refund if the container and its gold landed within the
shop. Throwing behaved likewise, just less verbosely.
The problem is caused by addtobill() treating gold specially
and then subfrombill() not being able to perform a reverse
operation. Actually, it may be possible for subfrombill() to
do that, but verifying all its uses is too much work. This
moves the gold handling for drop+selling into its own routine
and adds calls to that for the throwing and kicking refunds.
The other calls to subfrombill() outside of shk.c appear to be
ok as-is. (The calls inside that file are the ones that still
need evaluation if the gold handling is to move to there.)
bill_dummy_object() now uses the same o_id assignment for its
dummy object as split_object() does for its new partial stack.
I don't know whether the old code led to any price glitches.
Adds sanity checks for mtrapped and mundetected states.
Fixes cases where those were left in wrong state.
1. Trapped monster (eg. a nymph) teleported out of a trap
2. Monster was hiding under ball or chain, which then got removed
3. While restoring a level, a zombie corpse revived while monster
was hiding under it
4. A general case where the only object was deleted off floor and
a monster was hiding under it
Monsters hiding under ball or chain will now get revealed when
the b or c are moved.
When SCORE_ON_BOTL is enabled, you could tell how much gold is
inside a container with unknown contents by having 'showsore' On
and watching how much the score changed on the status line when
picking the container up.
I was looking into adding a confirmation prompt for '!' and it
isn't very promising due to sequencing issues. (The check for
whether '!' is allowed should happen before the prompt about
running it but the latter should take place in the core rather
than in the port code.) In the mean time, I noticed that VMS was
ignoring the SHELLERS value from SYSCF.
Untested implementation of a SHELLERS check on VMS. Even if it
works, it should not be using $USER as the user name to verify.
Tweaks the Unix implementation of check_user_string() but doesn't
switch the testing loop to the simpler version used by VMS which
is derived from the generic users test used by Qt.
