Wasn't Array.length 32-bit on Java when the article was written? In fact, isn't it 32-bit even now?
Moreover I don't see how you deny that using signed would lose functionality in this case—it's pretty undeniable that it gives the wrong answer in cases where unsigned would give the correct answer; the code is right in the article and you can test it out. This is true irrespective of any other cases that neither might handle correctly (assuming you believe any exist, but see my previous paragraph).
