FYI, this looks like a bug in wsgiref.handlers.BaseHandler.finish_response(), which should probably be using a try/finally to ensure .close() gets called.  (Which would catch a failed write() to the client.)

I'm kind of amazed this has gone undetected this long, but I guess that either:

1. other servers are probably catching errors from .run() and closing,
2. they're not using BaseHandler in their implementation, or
3. most apps don't have anything that essential in close() and/or the clients don't disconnect that often.  ;-)