I think that it is better to do this with "binmode" but not with switches. One reason for this is that using "binmode" you could run the same cgi-app module under mod-perl.

I use the decode_utf8() in prerun stage approach in production code. Initially I had some troubles caused by interaction of utf8 and MD5, but I escape this case. Pay attention that "auth_login/auth_passwword" is what I use for authentication, the default in CGI::Application::Plugin::Authentication is different.

Now I see why you does not have problems with CGI::Fast - because you create new object for every request. My initial experience was with CGI::Application::FastCGI where the run handler is :

sub run {
    my $self = shift;
    my $request = FCGI::Request();
    $self->fastcgi($request);
    while ($request->Accept >= 0) {
        $self->reset_query;
        $self->SUPER::run;
    }
}
[download]

In your approach you have "new,run,new,run,new,run...". In my case I have "new,run,run,run...". Actually the difference in performance is not quite big but usually I have some heavy initialization in init stage that is run only once (on new object creation), so I see some benefit in using this approach.