Socket.connect_nonblock blocks

JRuby 1.7-only bug. It's actually fine on 1.6.7.

Repro:

Clone celluloid-io: https://github.com/celluloid/celluloid-io
Bundle
run "rake"

The tests hang around here:
Celluloid::IO::TCPSocket
inside Celluloid::IO
...

It's hanging on a follow-up call to Socket.connect_nonblock. By "follow up", I mean that it's called Socket.connect_nonblock once, gotten Errno::EINPROGRESS, has waited for the given socket to become writable (via nio4r), and is now calling Socket.connect_nonblock again to ensure that the socket is now connected. The code in question looks like this:

        begin
          @socket.connect_nonblock Socket.sockaddr_in(remote_port, @addr.to_s)
        rescue Errno::EINPROGRESS
          wait_writable
          retry
        rescue Errno::EISCONN
          # We're now connected! Yay exceptions for flow control
          # NOTE: This is the approach the Ruby stdlib docs suggest ;_;
        end

Okay, that comment is a bit of a lie, here's the actual example given by the stdlib documentation:

  begin # emulate blocking connect
    socket.connect_nonblock(sockaddr)
  rescue IO::WaitWritable
    IO.select(nil, [socket]) # wait 3-way handshake completion
    begin
      socket.connect_nonblock(sockaddr) # check connection failure
    rescue Errno::EISCONN
    end
  end

My version works fine on both Ruby 1.9.3 and JRuby 1.6.7. However, on JRuby 1.7, connect_nonblock seems to block.