This is everything up to the list module... now it's time for sleep.

Including intuiting ccan depends based on headers...

In particular, we can assume "somefunc()" exists in our examples.

Try to glue consecutive examples together if all else fails (works well for
ccan/alloc/alloc.h).  This means we should put examples into list in order.

  
These are from hash: the entire example is indented, there is no main(), and
there are unused static functions.

Some of the compulsory tests don't have a total_score; treat that as "1".
And don't print out a score if the total is <= 1.

Don't assume that ../modulename.o already built!

And don't require the module.o to already exist: link against generated .o files.

Includes primitive attempt to place fragments into appropriate framework to
compile them.

We had example names like "example-opt.h-opt_parse", which got trimmed to
"example-opt".  By using a basename of example-opt.h-opt_parse.c, we only
strip the final ".c" not the middle ".h".

Don't allow NULL descs, and put them in macro.  Users can use "", but make
sure it's explicit, not an accidental omission.
Use errx() instead of assert() for API usage errors.
Rename opt_flags to opt_type.

I still don't like the duplication, but it's clearer and not much longer.

longidx is not updated in this case.  We should open-code option parsing.

This is more explicit than separate short and long (great for grep!) and
simpler.

This includes renaming ATTRIBUTE_UNLIKELY_FUNCTION to ATTRIBUTE_COLD,
and removing unlikely_func macro from ccan/likely.

Valgrind was giving a spurious error.

We were getting occasional failures: if all values are the same it's a 1-line
graph and there's no "|" anywhere.

Currently they sit in each module.

Macro packages do this.