They definitely implement it; they're one of the implementations used to show implementation experience of the Web Annotations Model per the implementation report linked to from the W3C Recommendation.

As I mentioned in another comment, these tests can be trivially passed. You just paste in some JSON data. To pass "all" the tests you actually only need to implement a small handful of the fields. The rest of the tests are qualified by something like "if present" -- so you pass those by just not having anything present in your test case.

When I use the hypothes.is extension, I see queries in my browser's developer console that do not look anything at all like they are querying annotation collections as defined in the spec.

The data returned from these queries is close to the spec but includes many fields that aren't in it. These extra fields seem to be conveying important information for their application.

Edit: I ran one of their annotations through the W3C tests. Five failed tests. Including one of the most important fields:

       {
          "name": "1:2 Implements **Annotation _id_ key** which has a **single value** that is a **string of format uri** - [model 3.1](https://www.w3.org/TR/annotation-model/#annotations)",
          "status": "FAIL",
          "message": "assert_true: ERROR: Annotation is missing id key or its value is not a single string of format uri.; expected true got false"
        },

This is why I said I'm not confident they are actually using it.

As far as I'm aware, there's nothing in the spec that prohibits extensions to the set of fields, provided they use the required fields in their required way.

Of course, you run the risk that you end up with de-facto standards and everyone having to implement one implementation's extensions (oh, hi WebKit-prefixed properties!).

> there's nothing in the spec that prohibits extensions to the set of fields

Correct. But since it has all this "JSON-LD" stuff baked in, you need to also carefully define and publish the extra fields.