The first tls_read may and up reading more than the response line, but we were using another buffer in copy_body so we loose up to the first 1K of every document.

It's easier to allow the stopping of not-existent requests rather than track which tabs have still ongoing connections and what not and handle this bit of logic in the UI.

close_with_err does a close_conn which in turns drops req from the queue. Thus it crashes on error, avoid that.

Crashing is some situation is not a good idea. If the other process is dead, it has probably already crashed, so crashing again would overwrite the core file, and in every case it doesn't tell us much. It's more interesting to see why the first process crashed in the first place.