Proposal: "void f() { return 7; }" should be illegal

[Don]

In the docs for "return statement" in statement.html it says:
---
ReturnStatement:
     return;
     return Expression ;

Expression is allowed even if the function specifies a void return type. 
The Expression will be evaluated, but nothing will be returned.

----
This makes sense for things we definitely want to allow, such as:
void foo() { return bar(); }
where bar() is another void function. But as currently implemented, it 
also allows bug-prone code. Here's an example from std.utf in Phobos2:

// BUG: should return size_t, not void!
void encode(wchar[2] buf, dchar c){
    if (c <= 0xFFFF)
         return 1;
     return 2;
}

I think that instead, inside a void function, return expression; should 
be the same as: { expression; return; }, ie, like any other expression 
statement, it should be required to have side-effects. I found while 
patching the ICE bug 3344 that this is happening because the expression 
never gets the semantic pass run on it.

In DMD, in statement.c.
Statement *ReturnStatement::semantic(Scope *sc)

=====
	/* Replace:
	 *	return exp;
	 * with:
	 *	exp; return;
	 */
	Statement *s = new ExpStatement(loc, exp);
+	s->semantic(sc);
=====

After making this change, it showed up that bug in Phobos.
I'm not sure if the current behaviour is intended, or is a bug. It 
certainly looks unhelpful, unexpected, and bug-prone to me. As well as 
the one-line compiler patch, I suggest the following change in the spec:

-The Expression will be evaluated, but nothing will be returned.
+The Expression will be evaluated as an ExpressionStatement, but nothing 
will be returned.

Or is there something desirable about the existing behaviour which I 
have missed?