This is fairly bare-bones in that there is still a couple of
problems to solve before this is actually useful.

1. Rename task: error signaling? Possible solutions are:
    1. Block thread on each error, fire a signal and wait for return
    value.
    2. Collect all errors, fire a signal, wait for return value.
    3. Pass errors to the “finished” signal.
2. Although unlikely, but, in theory, the changes queue could never get
   flushed, as the timer is reset with each push (a hardcoded limit can
   be used, which would force a flush - sound familiar? :p).
3. Rename task (again): the “finished” signal will usually arrive before
   changed files and their parents are notified, which presents a
   glorious race to work around in the test executable (quitting the
   main loop, namely).

There is probably something else that I forgot to mention.