Maybe a bit of a nitpick, but I'm not sure how I feel about this test. It's really testing catch_unwind rather than handle_vm_panic. Basically the only thing this ensures about our code is that handle_vm_panic doesn't panic itself.

I agree for the automated case and it might not be a text book style test. However, it is very handy for manual inspection of the console output (which revealed the Any debug problem) as well as ensuring the function can be used in isolation with the weird input type. Do you have an idea how to improve the test?

The guard pattern maybe? A bit verbose but no parameter needed?

static VM_WRITER: RefCell<Option<Box<dyn Writer>>> = RefCell::new(None);

fn with_writer<W: Write + 'static, F: Fn(&dyn Writer)>(w: W, func: F) {
    let prev = VM_WRITER.get();
    VM_WRITER.set(Some(Box::new(w)));
    access_writer(func);
    VM_WRITER.set(prev);
}

fn access_writer<F: Fn(&dyn Writer)>(func: F) {
    VM_WRITER.with_borrow(func);
}

[...]

access_writer(|w| writeln!(w, "hello world!"));

Well, that code itself won't work immediately because .with_borrow isn't enough and it needs a fallback case to io::stdout() but close enough to illustrate what I mean.

And the overhead is alright since it's a rare case to occur anyway (hopefully never)

No strong opinion on if/how to improve this. But if there are no easy wins, I'd merge this as is and leave it up to you to improve it later.

Alright, then let's merge as is for now.