Humble revamp of @trust, @safe and @unsafe

[#ponce]

I think that  in the current design of safety, @trusted function and normal functions are quite similar. An @unsafe proposal has been rejected because of complexity.

But here is a case that is left.
Sometimes in D1, I found that a function I tought trustworthy is in fact completely buggy. I mark it "deprecated" to let the compiler found all calls to this function.

Here is a humble proposition for another safety policy which account for that need.

1/ All "normal" functions are @trusted by default. They can call any function. Remove the @trusted annotation.

2/ @safe functions call call @safe or normal functions.

2/ introduce a @crap annotation  (or maybe @__@) for objects and methods. The @crap annotation hilight bad code, bad design and provides a handy metrics when reviewing code. Similar to the previously proposed @unsafe annotation.

3/ remove the -safe switch and introduce -unsafe. In safe mode, any function calling a @crap function becomes @crap. In unsafe mode, @crap is not viral.

4/ @safe and @crap are of course mutually exclusive

In short:

- @safe can call:
   - @safe functions
   - normal functions

- normal function can call:
   - @safe functions
   - normal functions
   - @crap functions (but become @crap if in safe mode)

- @crap functions can call anything.


I don't know if it's feasible. It supposes to trust the programmer more than the current design, but the current design may lead to @trust abuse imho...

The main point is that functions are trusted by default.
This proposal makes @crap viral, to incent the programmer to eradicate it. 
It lowers _a lot_ the guarantee you have when compiling -safe code, because it implies the programmer to mark manually unsafe functions. This is a very weak point.
If one want to prove more safety, one may aswell put @crap: on top of the module you are making safe and turn it to @safe progressively.

It accounts for two different needs : making guarantees about program safety, and fixing a dangerous function without breaking anything.
It's also non-intrusive if the programmer is only interested by the second goal.

What do you think ?