In Julia, type specialization (the secret sauce that makes it faster than Python but still dynamic) occurs at function calls. Nothing at compilation time can be inferred about the type of ddψ since it depends on a non-constant global. The for-loop can’t regain knowledge of the type, since it lives in the same function as when disaster struck.
Things would be different if the for-loop was in its own function (see this part of the performance tips).