Any insights as to how to get effective determinism without pure functions?
Pure functions win here because we only have to reason about the function arguments. Reasoning is easier when you have to less of it! Without pure functions, the state space explodes, because anything anywhere could have a side-effect — madness, I say! — so how would you figure out equivalent states? Special case inspection I suppose?
There is a continuum between "anything goes" and "pure functions". From a certain light, all the varieties of encapsulation can be seen as explorations of this theme. (An idea: perhaps there are languages where side-effects can be circumscribed in ways that don't rely on the high bar of pure functions?)
The arguments against pure functions appear to be somewhat contingent: current adoption levels, practice, convenience, and taste. Things that are contingent have a way of changing over time as new methods emerge and new values take hold.
I don't hold shallow human preferences in high regard. As an example, many humans and organizations in practice trade security for convenience, at significant cost — especially when that cost is external to them.
At the same time, onerous tasks encourage work-arounds. This is the crux of it I think: unless people perceive the value of pure functions (to do what they want to do), some will seek shortcuts. So we need some combination of (1) making them easier to work with; (2) motivating people to aspire to higher principles; namely, high quality software that does not open to door to exploitation, misuse, botnets, etc; and/or (3) internalize the external costs.
As an example of (3), if people and companies faced predictable and commensurate costs when their flawed software inflicts harm on others, they would be more inclined to care about correctness.
At least with many GUIs, the time it would take to know, document and/or test all combinations of UI behaviors quickly exceeds the time available to humans. It's at least tractable to start with vastly more states than could ever be measured, and then gate or special-case certain ones to try to make things minimally safe/sane/reliable for the user.
Pure functions/immutability help a lot because tests become representative and cheap. I’d only push back on “tests of all equivalent scenarios” being sufficient, the space explodes and many real failures live at I/O/concurrency/distributed boundaries. My intended claim is that FP/ADTs/types reduce the state space and improve the ROI of tests, not replace them.