This question is inspired by this answer and related to unsolved post "Association's Attributes change between 10.3.1 & 10.4".
Consider the case of pre-evaluated Association:
Clear[y]; assoc = <|"x" -> y|>; y = 0; assoc <|"x" -> y|>
However with the above definition for y we get for a new Association:
<|"x" -> y|> <|"x" -> 0|>
The first question: what prevents assoc from being further evaluated to <|"x" -> 0|> and why the new Association doesn't hold its arguments?
The second question is related to the structure of Association. Let us compare evaluated and unevaluated Association:
AtomQ@<|"x" -> y|> AtomQ@Unevaluated@<|"x" -> y|> True False
Depth[<|"x" -> y|>] Depth[Unevaluated@<|"x" -> y|>] 2 3
Despite the above TreeForm in the both cases shows that we have atomic objects:
TreeForm[<|"x" -> y|>] TreeForm[Unevaluated@<|"x" -> y|>] 
So, is Association atomic or not? How to understand this controversy? Is there a special internal definition for Association causing this behavior?
AtomQpart of question I do not see a difference to sayGraph. Unless with raw objects like stings or numbers, the head must be evaluated so that the atomic quality is recognized, mustn't it? For example, a Graph-object with a false body will not be an atom... $\endgroup$TreeForm[Plus[1, Unevaluated@Association["x" -> y]]]willl not show an atomic structure. This is very depending on the evaluation process... $\endgroup$Associationis both a constructor and a container? For example,Hold[<|1 -> 1|>] === With[{x = <|1 -> 1|>}, Hold[x]]evaluates toFalse, indicating that there is some construction step going on that makes a constructed association different from the raw input. $\endgroup$