Nit: I'm not sure replacing

pub struct BumpBox<'a, T> {
    ptr: NonNull<T>,
    _allocator: core::marker::PhantomData<&'a ()>,
}

with

pub struct BumpBox<'a, T> {
    value: &'a mut T,
}

... is necessarily making the intent clearer.

However removing the destructor

impl<T> Drop for BumpBox<'_, T> {
    fn drop(&mut self) {
        // Safety: The pointer is valid and we own the data
        unsafe {
            self.ptr.as_ptr().drop_in_place();
        }
    }
}

... is absolutely NOK. Why did you do that?

Otherwise, I totally like having miri as part of the CI.

The new code does not assume that the memory is initialized, before it was written to (I think the old one was okay too, but not sure)

It is OK (in that it will never, ever lead to a program crash or anything like that - it is just a UB if you read from the uninitialized memory which was assumed initialized, but we don't do that, and even if we did, it would be "undefined behavior" but not a crash).

Miri found this in the original code let ptr = t_buf.as_ptr() as *mut T; where a *const T is cast to *mut T

Yeah. Maybe just let ptr = t_buf.as_mut_ptr() and be done with it?

Nit: I'm not sure replacing

pub struct BumpBox<'a, T> {
    ptr: NonNull<T>,
    _allocator: core::marker::PhantomData<&'a ()>,
}

with

pub struct BumpBox<'a, T> {
    value: &'a mut T,
}

... is necessarily making the intent clearer.

However removing the destructor

impl<T> Drop for BumpBox<'_, T> {
    fn drop(&mut self) {
        // Safety: The pointer is valid and we own the data
        unsafe {
            self.ptr.as_ptr().drop_in_place();
        }
    }
}

... is absolutely NOK. Why did you do that?

Otherwise, I totally like having miri as part of the CI.

I think I misunderstood a part of the code, yeah &mut T isn't sensible that should be a pointer. I thought of it as a reference to the array and not as it having ownership, which it should have... I will undo those changes.