There are languages that model code like this. Using one for a while will likely teach you why. If you think "spaghetti" is an apt description for text code, wait until you see flow based code that's evolved in place for a while.
Perhaps I'm imagining an alternative way to manage source trees, or modules, rather than the code in individual source files. A hierarchy like a traditional file system but with extra kinds of relationships. Perhaps it'd have to be curated like reference documentation, perhaps it could be automatically generated from source directives. I dunno, just idle thoughts...
These features allow the user to iterate rapidly through the development process to create a prototype, perhaps even gaining an average 100x development-time increase over using a compiled text-based language. The spaghetti is evidence of that-- the user moved so quickly that they were able to hold the relevant abstractions in their short term memory to produce the final write-only diagram. And at least in my experience, the final write-only diagrams tend to work to produce the desired output, even if they were written a decade ago.
In fact if there were a button users could push to "abstractify" some pre-spaghetti sub-diagram I believe that would negatively impact the language. The point is that when you've got mental model of all inputs and outputs within a single page visual cheat sheet, you can instantly connect any input to any output to take the next step in the development process. For certain domains like audio synthesis the fact that the user can leverage the diagram-based system to get any output at all is worth the seemingly prohibitive price of write-only programming.
It’s my opinion that one of the things that turns visual code into spaghetti is that refactoring is difficult, because the code grows in 2d, as opposed to 1d (I don’t count line length). It’s sort of like if you could have infinitely many columns of concurrently executing code in text a based function, where any line in a given column can be connected to any line in a different column.
I think what could be useful is a hybrid model, where you have something like text based modules whose communication is orchestrated by a visual diagram. For what it’s worth, Simulink can be used like this to some extent(e.g., hooking S-functions together) and I find it orders of magnitude more pleasant than labview.