Very much agree on these points. In my experience the pragmatic path for larger projects has been a hybrid of keeping visual programming graphs self-contained -- roughly within the "asset" boundary -- and then using a hierarchical format with textual representation to assemble those assets and drive their interface parameters. This is roughly how animation studios use USD (openusd.org), for example. Arguably this is just a strategy of combining a couple domain specific languages, with edges where the optimal tradeoffs of each domain flip. It's a very interesting question whether these limitations of visual programming are essential complements of their benefits, or if there is some good way to provide better abstraction, re-use, etc. Certainly Houdini, Nuke, Katana, etc. all provide limited forms of those things (ex: Houdini OTL's; scripts that version-bump nodes and try to auto-upgrade the parameters), and they do see lots of use in industry.
so I started my studies in animation actually, I used a lot of those tools, you can think of statebox as applying category theory to restructure aftereffects or https://en.wikipedia.org/wiki/Shake_(software)
when done right, we claim, you can target many different things ("semantics")
what we claim is something like, the compositional aspect of many such node based systems can be described as a certain type of mathematical object (monoidal category) ~ we can build an editor for that and then map that "dsl" to particular targets (image processing, state machines, etc)
