It doesn't matter which property you are looking at. If it's undecideable, you can sidestep the issue by forbidding programs for which the checker is not sure.
That doesn't mean you have to lose Turing Completeness. You just have to chose properties that don't imply that the program halts. Static typing for instance doesn't mean the language isn't Turing Complete. Simply typed lambda calculus is the exception here.
I sometimes workaround this by proving for "outputs the correct result, or loops forever". Here, you don't have to prove for halting, since your program is not expected to halt in some cases anyway. But then in some particular cases, you can prove that your algorithm is correct. You do not have to lose Turing-completeness to do this. But it gets trickier when you want to built a type system that can give all the bugs in your program, in general. You can definitely build something that finds all the bugs "for all intents and purposes" but not something that can be mathematically proven that will give you all of them. To be concrete, you can build a program, given arbitrary Python code, gives most of/all the bugs in this program. But you cannot build a program, given arbitrary Python code, that is proven that it will correctly know whether this program will output correct result for all possible inputs. You can do this for some restricted Python programs (as you've been arguing for last two comments) but in general you cannot do this, which brings us to this fundamental division.