It looks like there is some misunderstanding regarding to what limit_conn actually limits. It limits concurrent connections *processed* by nginx (not keep-alive ones), and only after header has been received (and thus configuration for request has been determined).

Since nginx is event-based, with one worker process you shouldn't expect requests to hit limit_conn unless they block at some stage (i.e. responses bigger than socket buffers if sendfile off, replies bigger than sendfile_max_chunk if sendfile on, proxy_pass ...). With many workers limit_conn may be hit without blocking, but this generally requires _very_ high concurrency for small requests.