One of the main design goals of the converter was to minimize memory consumption as much as possible. Because of this, it does not construct an abstract syntax tree prior to generating the HTML output, and this means that it has to perform all operations on a single pass, including deciding whether it's looking at a field declaration, an inner class or whatever.
The most obvious shortcoming is that the javadoc linking features rely only on the name of the identifier and not on the context. For example, if you do this:
int String = 3;
Code2Web will think that "String" is a class reference and will link it to the appropriate javadoc page.
This may be fixed in some future version, but I must confess that I'm not in a hurry to do it. If you progarm like this, you deserve what you get :-)