For example, if your language lacks a CopyTagedMatrix then you just write one, it gets the correct size, allocates the memory if necessary, copy the elements and return the new matrix. Now you can use it without remembering the implementation details, and it’s much clearer than the 10 lines that necessary to get the bounds and the fors to iterate over the matrix.

An extreme example is the GMP bignum library. You use each function as magic, and never read the implementation details. (Unless it’s in a very tight loop in a very hot part of the most important feature of your program, but I doubt that it’s good to use GMP in that place.)

Gates had a complete sense in his head of how his program worked at the level of machine language. But not at a lower level of CPU microcode, or transistors. So even he, even then, was not claiming to have "total" understanding in the sense I think you mean.

Likewise if the Linux kernel (or your distributed web app) is organized into a reasonable set of "chunks", then it is possible to have usefully complete sense of how it works. The implication being that if it's not organized -- or organizable -- like that, you have a problem.