Upon startup, the server collects from all configured packages the -proc.tcl files and builds from these files a so-called blueprint (consider this as a single large file containing all the definitions). Aolserver uses threads for various purposes. For the handling of incoming HTTP requests, the "connection threads" (sometimes called "request threads") are used.
When a connection thread starts up, its interpreter is initialized with the blueprint. Therefore, every request has all procs/objects/global variables of the configured packages available without any loading.
Aolserver handles a pool of connection threads, every
connection thread serves typically several http requests (the number is configurable in the config file). After a connection thread has processed the configured number of requests (or an idle timeout) it shuts down. New connection threads are created upon needs.
After every request, the connection thread performs a cleanup that deletes e.g. the global variables (the variables in the namespace "::" of the tcl interpreter of the connection thread).
Connection threads are an important part of the picture, but there is much more infrastructure available: In addition to the connection threads, there are driver and spooler threads (naviserver), there are threads for scheduled procs, it is possible to define job-queues, one can define persistent threads, etc., there are options for lazy-loading, etc.
From your terminology, "global global" refers to the blueprint, and "local global" refers to the interpreter of the connection threads.
Hope this helps to understand the basic model.