So use %define instead of %global.

Adding an RFE in bugzilla to state an expectation that %global should work like
%define so that Fedora packaging policy recommendations can be accurately
stated as
    Always use %global.
makes very little sense to me.

Have fun!

That is fine but are the expected behaviours of the two documented somewhere?

There's not much to document.

Macro expansion proceeds through levels of recursion.

%define is wrto whatever recursion level is active, with
automatic garbage collection and %undefine upon return.

%global is just a macro defined at the top level, i.e. level 0.

The problem is that lusers are surprised when %define
not at level 0 disappears. This is exactly the behavior
implemented and intended, just not what lusers expect.

Are luser expected behaviors of RPM documented anywhere? ;-)

Thanks for that explanation.

Given that my testcase is for level 0 %define, what is the difference in this base between %define and %global?

(Docs, muahaha;)

The difference is that %define "works" and %global does not.

The difference is that %global expands the body at definition time whereas %define doesn't. Fedora recommendation of %global over %define completely misses this point.

You can force %global to be lazily expanded with %%-escapes though:

%global globalrequires()\
Requires: %%1, %%{?2:%%2-devel}