The complete list is here: http://wiki.hl7.org/index.php?title=Publicly_Available_FHIR_Servers_for_testing
There are no formal recommendations. Stability will vary as the specification evolves and updates are applied to different servers, the amount of bandwidth their maintainers have at a given point in time as well as server limitations (load, temporary bugs, etc.) The most reliable server this week may be less reliable next week. Implementers are encouraged to test against a variety of servers so they don't become over-dependent on the capabilities (and quirks) of particular implementations.