The problem is that when the main thread ends, the RTS doesn't stop for another 6 or so seconds. The only thread that runs this long is the handler (waitFor (secs 8.0)) but it has already been killed. So I'm scratching my head a bit.
Short answer: use -threaded.
The runtime is waiting for a worker thread to complete before it can exit; even though your Haskell thread has been killed, there is still an OS thread executing Sleep() which was started by threadDelay, and this OS thread has to complete before the RTS can exit. We should really terminate the thread more eagerly, but since this only affects the non-threaded RTS fixing it isn't a high priority.
Sweet, thanks. A note in the docs somewhere about this implementation detail would be nice; something that says you may want to consider -threaded because the implementation of threadDelay uses an OS thread. I suppose the best place would be in the docs for Control.Concurrent.threadDelay, although perhaps too a mention in section 4.10.7 of the ghc user's guide, because in there it says: "Note that you do not need -threaded in order to use concurrency; ..." It could say "Note that you do not need -threaded in order to use concurrency (unless you use threadDelay - see Control.Concurrent for details); ..." Alistair