> All language keywords are “magic” to some degree.
Agreed. Every keyword adds complexity to the language; the fewer your language needs, the better. If your language has first-class functions and polymorphism (and any serious language does these days), there's no need for special-case control flow keywords; better to have a design like e.g. Smalltalk, where if/while/... are just normal functions.
> The control flow that try/catch creates has never been surprising or invisible in my experience.
One of the biggest production bugs I saw happened because of removing an unused variable (the function looked correct, but actually the unused variable right at the top of the function was throwing an exception; the fallback code path was correct, but the function itself was implemented wrong). It's the same problem as https://glyph.twistedmatrix.com/2014/02/unyielding.html - you can't tell which function calls might throw by looking at them, and most of them don't, but some of them do. And on top of that you have the goto-like "action at a distance": starting from a given catch there's no way to find the corresponding throw (or vice versa). You can't even tell when a catch is completely unused.
Easily refuted observing that you could strip languages of many of their keywords (for example replacing for and while with if and goto), and you'd end up with less readable code. Keywords are often added to make languages simpler, at least because they declare the intention of the developer.
Otherwise, Brainfuck would be the least complex language to write code in.
You're conflating the complexity of the language itself with the complexity of code written in that language. Brainfuck is a famously simple language; brainfuck codebases are quite complex. Sometimes building complex functionality into a language might be worthwhile, if it allows you to simplify code written in that language. But since a developer working in the language has to be able to understand both the language and the codebase they're working on, you need to be careful about that tradeoff, only adding to the language those features that are general and powerful enough to simplify a lot of codebases. Otherwise you blow the whole complexity budget on language functionality, leaving the developer with no spare mental capacity to understand the specific codebase - even if their codebase makes no or minimal use of some of those complex language features.
Better, where possible, to implement general-purpose reusable functionality in libraries written in ordinary code in the language. That's a win-win approach: reusable library code can simplify codebases written in the language, but since it's plain old code that follows the normal rules of the language, a library doesn't add language complexity that the developer is forced to keep in mind.
Fundamentally, the key to making codebases in your language understandable is to be compositional: make sure that the developer can easily understand the combination of a and b if they understand a and b separately. Library functions do that, because you can understand how code that calls a library function behaves without needing to understand what the library function does. But language features like throw/catch don't do that: you need to understand what throw does if you are to understand code that calls a function that throws, even if the code you're trying to use doesn't throw itself.
Sounds like you’d enjoy an R5RS scheme. Minimal syntax goes a long way. Unused variables throwing exceptions definitely seems like an over accumulation of technical debt.
Agreed. Every keyword adds complexity to the language; the fewer your language needs, the better. If your language has first-class functions and polymorphism (and any serious language does these days), there's no need for special-case control flow keywords; better to have a design like e.g. Smalltalk, where if/while/... are just normal functions.
> The control flow that try/catch creates has never been surprising or invisible in my experience.
One of the biggest production bugs I saw happened because of removing an unused variable (the function looked correct, but actually the unused variable right at the top of the function was throwing an exception; the fallback code path was correct, but the function itself was implemented wrong). It's the same problem as https://glyph.twistedmatrix.com/2014/02/unyielding.html - you can't tell which function calls might throw by looking at them, and most of them don't, but some of them do. And on top of that you have the goto-like "action at a distance": starting from a given catch there's no way to find the corresponding throw (or vice versa). You can't even tell when a catch is completely unused.