Related to #2025.

With input from nigoroll.

These versions implement previous drafts that will fail the upgrade.

platforms

cost by correctly holding the mtx while examining it.

It's amazing how long time it can take to spot a bug like this :-(

See also: #2061

VRT_CTX used for VCL events.

Acquiring a read lock on the CLI thread is unnecessary since a write
lock may only be acquired on that very thread.

The locking order when both are required is:

    (struct vcl).temp_rwl => vcl_mtx

Fixes #2008

There is currently no locking around temperature changes despite the
fact that it may be read outside of the CLI thread to make decisions.

In the case of a panic, instead of locking we simply make sure not to
perform an incomplete read of the pointer.

Pass childs status on failure

Make atexit more robust

thanks to @dridi for the dictionary lookup

Maybe it's just my brain that is incapable of considering the
fact that a bit of macro-fillin may in fact be #if 0 - disabled
here, so this should hopefully help me (and others?) to not
spend more life time than absolutely necessary on feeling stupid.

if they are signed or unsigned.

Range-check enums is sound programming practice, but that concept
seems to be beyond the imagination of certain compiler people:

   vhp_decode.c:96:2: error: comparison of unsigned expression >= 0 is always true [-Werror=type-limits]

Add a dummy "MIN" value to the enum, give it value -1 to force
the compiler to use signed ints for the enum, and then check
that they're never negative.

Fixes #2084

To avoid another regression, a test case now passes a probe to the debug
VMOD. This problem was discovered while working on #2068.

Also try to clarify CLIS_PARAM (= nonsensical parameter) from
CLIS_CANT (= params make sense, but you can't do that.)

Spotted by: Dridi Boukelmoune

Thanks @nigoroll for the suggestion.

This allows to test vgc_inistep only once and make it easier to work for
the compiler.