My point was to point out the limitations. I might be wrong, but people seem to often ignore these limitations. This oversells the benefits and leads to frustrated people like the one whose comment I originally responded to.
99.5% is not nothing, but not something you can live with in most systems so you are probably going to need something more. Also, you are going to be a lot happier if the vast majority of those hundreds of lines in your cross-system call stack are completely agnostic to the types of your problem domain and you want to be careful to minimize the coupling between components living on different sides of the jumps you mention. You can do that, of course, with static types, but it takes more care and intentional design and you really need to acknowledge the issues before you can get it reasonably right.
If you look at a typical call stack - in any language - going from the frontend all the way to the database, there will be hundreds of lines. Two, maybe three of those will be remote jumps.
Every one of those calls is an opportunity for a contract mismatch. Even if static type checking misses a half-percent of your calls, it's still working to enforce 99.5% of your contracts. The value of that is hard to oversell.