At least, that’s my understanding of it.
For any sufficiently powerful system (Turing completeness is definitely powerful enough) there are statements which are true but cannot be proven by that system. Secondarily, no system can demonstrate its own consistency.
However, if you limit your software enough, you can prove that every possible input is handled correctly. That involves removing Turing completeness. E.g. once you accept a language as a configurator, you can no longer prove that. CSS3 is Turing complete. JSON is not.
One requirement is limiting the size of inputs.
Cynics might take this to mean that maths is meaningless. On the contrary, I'd like to suggest that maths is the canonical emergent pattern; any formal system with sufficient complexity to represent itself will automatically exhibit some algebraic structure, and maths emerges naturally from there.
The biggest implication for computer scientists is that, combined with results of Church and Tarski, we should expect that meta-languages are more powerful than object-languages, whether this is by preprocessing and macros like in C/C++ and CL, or by avoiding using meta-features except when necessary like in Python and Ruby.