Supposedly Forth programs can be "compiled" but I don't see how that is true if they have words that are only evaluated at runtime. For example, there is the word DOES> which stores words for evaluation at runtime. If those words include an EVALUATE or INTERPRET word then there will be a runtime need for the dictionary.
To support such statements it would mean the entire word list (dictionary) would have to be embedded inside the program, essentially what interpreted programs (not compiled programs) do.
This would seem to prevent you from compiling small programs using Forth because the entire dictionary would have to be embedded in the program, even if you used only a fraction of the words in the dictionary.
Is this correct, or is there some way to compile Forth programs without embedding the dictionary? (maybe by not using runtime words at all ??)
Forth programs can be compiled with or without word headers. The headers include the word names (called "name space").
In the scenario you describe, where the program may include run-time evalutation calls such as EVALUATE, the headers will be needed.
The dictionary can be divided into three logically distinct parts: name space, code space, and data space. Code and data are needed for program execution, names are usually not.
A normal Forth program will usually not do any runtime evaluation. So in most cases, the names aren't needed in a compiled program.
The code after DOES> is compiled, so it's not evaluated at run time. It's executed at run time.
Even though names are included, they usually don't add much to program size.
Many Forths do have a way to leave out the names from a program. Some have a switch to remove word headers (the names). Other have cross compilers which keep the names in the host system during compile time, but generate target code without names.