The specification of the Coding data type currently (R4 and R5) contains the following statement:

If the system is present, and there is no code, then this is understood to mean that there is no suitable code in the system in which to represent the code.

This seems to be problematic for several reasons, as discussed in this Zulip thread: 

https://chat.fhir.org/#narrow/stream/179202-terminology/topic/Interpretation.20of.20Coding.20with.20system.20but.20not.20code

Some of the problems discussed:

1. Entries with system but no code can quite easily arise by mistake, e.g. if a system is pre-selected in a UI but the user simply forgets to enter the code, or as a result of simple programming mistakes. Assigning a specific, clear intent to such entries (rather than saying they have no clear meaning or deeming them invalid) could thus lead frequent misinterpretations "in the wild".
2. "It's a universal "OTHER" which, for classifications is invalid since a classification is by definition exhaustive. Furthermore, you'd want to know the code system version (because a later version may have added an appropriate code)."
3. Such element may be used to satisfy a "required" constraint on a coded field, but it not clear that it is a valid use, given the possibly more appropriate use of an extension signalling the absence.
4. While such element may be appropriate in specific situations, it does not seem appropriate for Coding quite generally.