The compiler should not "just" know it. It would know it because we told it how.

Consider a function that takes a float and returns a complex number. I then change the function to take a complex number ("Complex Float" in Haskell if I read the docs right), and returns a float.

I could then tell the compiler, by implementing an implicit cast operator, how to cast float to complex. The implicit part is then that the compiler tries it without me telling it to use the cast explicitly.

Then any code that worked with the old function should work perfectly fine using the modified function without modifications, since per definition the reals are contained in the complex numbers.

This is how I do it in several languages I've used.