Which is throughly incorrect. Of course you can set the state of an output register that’s set to input, that state will just not appear on the pin.
This kind of encoding prevents you from making the programming error of writing to an input pin - which is technically possible, just useless.
However, by default, nonsensical operations should not be possible.
I haven't seen that in ARM microcontrollers, but as peripherals vary across manufacturers I don't think you can assume that it's not done this way.
In other cases, you want to avoid glitching when the pin switches to an output on startup.
If there are multiple chips connected in open-drain, you write the I/O pin to 0 or 1 to either ground or float the pin. You read the I/O pin to see if anyone else has grounded it if you haven't. This was called "wired-OR" in the old days. (It's really AND, but was commonly used in DeMorgan equivalent form for communication buses before tri-state drivers became a thing. I'll get my cane and hobble back to my rocking chair now..)
The context is Rust, where the statement is clearly correct. One can't call a function if the Rust compiler refuses to generate the binary. The author probably should have used a period between the first and second independent clauses and a semicolon between the second and third in order to show the close relations ship between being unable to call the function and it being a compile error.
However, the statement being objected to was under the section about the 3rd level of abstraction: "The Embedded HAL to the Rescue", which is detailing an extra semantic restriction imposed by Rust. The final two sentences of section 2 even allude to the prevented actions being legal but almost certainly being a logic error. To object that the hardware allows nonsensical actions without trapping is to miss the point of section 3 of the article. At the third level of abstraction: Rust prevents something legal but nonsensical.
Now, if this particular board doesn't clear output state when switching a GPIO pin from input to output, then it's a valid criticism that perhaps the developer wants to set the output state correctly before driving the pin and the prevented action is in fact semantically meaningful and perhaps the intent of the developer. However, that wasn't the OP's objection. The OP's objection was simply that Rust was preventing something that the underlying platform allows.
To recap: The article said "Rust doesn't allow you to do X" The OP said "Wrong!!!! The mocrocontoller totally allows you to do X" I pointed out that the article never said the microcontroller doesn't allow X.
One would suspect that other functions in the HAL that control the feature you are looking for directly.