A fix for this is on master now, with commit 8a5dad2045434fcc06f2878de55f7dcdefa01a1b. There were a number of issues in the implementation of FnTransport that were preventing exceptions from bubbling up as they should have been.
No APIs were harmed in the course of implementing a fix. Also, no keepalive mechanism was required (which makes me slightly suspicious that the issue has actually been fixed, but we'll see). (Not TCP keepalive; that's irrelevant to the issue, and wouldn't help even if it were reliably standard/present.)
New tests added, and manual verification against nREPL servers running in different processes looks good too. I'll put out an 0.2.0-beta10 shortly so that people can start banging on this to see if it resolves issues in downstream projects/tools.
Note that I also found that it sometimes requires two sends (messages here, not writes to a socket) to provoke an error when writing to a transport that had been connected to a server that was closed or otherwise went away. Maybe it's a buffering issue? I don't have a good theory. Check out e.g. nrepl-test/transports-fail-on-disconnects if you want to play around with it.
Task
Closed
Major
If it helps, I did observe in some testing that after the server side of the socket goes down, writing to and flushing the client socket's output stream twice would bubble up a client exception. That seems a little keepalive-ish, but I'm not sure whether that sort of thing would do other transports any good.