I've started to have a look at this, founding a couple of interesting tests at test/language/unclassified/test_attribute02.e and test/language/unclassified/test_attribute03.e

Currently they "obviously" fails. So I tried using ISE online compiler at https://www.eiffel.org/codeboard with this small code

class    APPLICATION
create     make
feature {NONE} -- Initialization
    make
        do
            print ("Hello Eiffel World!%N"+e.out)
        end
    e: INTEGER
        attribute
        Result:= 42
        ensure Result>0
        end
end

ISE compiler answered with this warning:

Warning code: VWAB

Warning: attribute body is never executed because the type of the attribute is self-initializing (i.e. expanded or detachable).

What to do: remove any instructions in the attribute body.

When running the above example you get:

======================
Hello Eiffel World!
0
======================

Showing that the attribute body is actually not executed.
So at least this means that at least for expanded attributes we may "safely" consider this a non-bug.

The issue would be much different whenever e is a reference object, such as a LINKED_LIST, like this:

class    APPLICATION
create     make
feature {NONE} -- Initialization
    make
        do
            -- Add your code here
            print ("Hello Eiffel World!%N"+e.out)
        end
    e: ARRAY[INTEGER]
        attribute
        Result:= <<1,2,42>>
        end
end

that gets:



    application.e

Hello Eiffel World!
ARRAY [INTEGER_32] [0x7F405BDFE658]
  area: SPECIAL [INTEGER_32] [0x7F405BDFE628]
    -- begin special object --
        0: INTEGER_32 = 1
        1: INTEGER_32 = 2
        2: INTEGER_32 = 42
    -- end special object --
  object_comparison: BOOLEAN = False
  upper: INTEGER_32 = 3
  lower: INTEGER_32 = 1