Thanks for the feedback!
I take it that you've already considered other register OISCs like subleq, of course, but you might not have considered Fractran. It's a OISC register machine with mul as its only operation, worth a look if you're rooting for register/tape machines.
Looking the other way, toward BLC, you might enjoy Iota/Jot as a ultra-minimal reduction system. Implementing such a system is even faster than doing something like bfs.
My opinion is that all these sequential machines are going to fall out of favor for highly parallelizable systems in the near future. For a future system of a size comparable to brainfuck. I'd look at interaction nets, with only 6 reduction rules, interaction nets are a likely candidate, or are at least worth a look. It's just a stack of interaction between nodes that can be reduced in any order to achieve computation.
Moving beyond, it's likely that cellular automata systems will prevail over everything else that I've mentioned above. Something like a 4-state(a-la wireworld, qu-ant), is easily communicable with pictograms, highly parallelizable, and in some cases reversible.
On communicating ideas in the far future, the movie Into Eternity, talks about how to communicate a danger to the people who would stumble on a nuclear deposit. There is a segment about language that might be interesting to you.
Anyways, I hope that you keep on researching this.