Like all logic chips today, with a standalone programmer. You make program on your desk computer then send it to chip. Electronic-mechanical converter would setup your device.
Without computer you could do it old-way with a panel of switchable bits and clock signal. You enter required bits with user-switchable levers, then switch clock lever up and down, then enter another machine-word. Or make your program on punch cards.
Apparently it runs fast enough to be programmed interactively. Keyboard input via mechanical linkages is a solved problem: consider the humble typewriter. The Selectric typewriter even had macroscopic digital mechanical "circuitry"; each keypress moved a lever with lugs that encoded a binary code specifying coordinates for where on the typeball the character was located, and special linkages called whiffletrees converted the digital code into analog rotations for the typeball mechanism in the manner of an R-2R ladder or similar in the electronic realm.
The idea is that the CPU and peripherals would communicate via mechanical linkage (say, levers moved, or not moved, according to whether the signal is "high" or "low") rather than electronic voltage levels on a wire. For some peripherals, electromechanical adapters could be built that use relays to convert electronic signals to mechanical ones and back, allowing you to theoretically even plug in standard PC keyboards and displays into such a mechanical computer.
I'd have thought a mask ROM would be incredibly easy to build. And the whole system would be nonvolatile anyway, so storage of mutable data would be straightforward.
Ultimately the code would compile down to binary; getting it into the machine is just like any other interface.
