Your documentation at http://velocity.apache.org/engine/releases/velocity-1.5/user-guide.html#velocimacro_miscellany has the following snippet -
Currently, Velocimacros must be defined before they are first used in a template. This means that your #macro() declarations should come before using the Velocimacros.
This is important to remember if you try to #parse() a template containing inline #macro() directives. Because the #parse() happens at runtime, and the parser decides if a VM-looking element in the template is a VM at parsetime, #parse()-ing a set of VM declarations won't work as expected. To get around this, simply use the velocimacro.library facility to have Velocity load your VMs at startup.
Is this likely to change in any forthcoming versions?
The use case is theme development for Confluence. It would be nice to be able to share some common macro definitions between a number of templates within a theme (i.e. not inlining). And using the velocimacro.library property isn't really an option because themes are dynamically loaded or unloaded.