Ok... This is totally weird. There must be a memory bug with the Win32 GDBM_File, but this is the correct FETCH to make my code work, in case anybody cares...

sub FETCH { my ($this, $key)=@_;
 return $this->{'l'}{$key} if exists $this->{'l'}{$key};
 my $v=eval($this->{f}{$key});
 return $this->{'l'}{$key}=$v;
}
[download]

For some reason, putting the eval'd structure in a my'd variable clears up the vivifacation issue. Go figure...