At some point in the distant past (7f36f3a) annotation deletes were
handled by adding the "deleted" property to annotations and anonymising
them. Now we actually delete them.

This commit removes code that was intended to handle such annotations.

Remove "show_unanchored_annotations" feature

We've discussed this feature and decided that it can't be turned on in
its current state. This commit removes it.

Note that the feature flag in `h/features.py` remains until these
changes are deployed out to production.

Merge pull request #2739 from hypothesis/trello-177-change-the-default-behavior-of-the-embed-to-highlights-on-by-default

Enable highlights by default everywhere

Fix the missing threads in the stream

This reverts commit 5a5b4e345c050a71e619fb5732426b2306f29e64.

Add a new, undocumented separate_replies=True option to the search API.

If separate_replies=True option is _not_ given to the search API, then
it reverts to its previous behaviour: _do_ include replies in the "rows"
list returned. This is the same behaviour that the search API had befor:
it returns both top-level annotations and replies in the one "rows"
list, but without any guarantee that if some annotations/replies from a
given thread are in the list then all annotations/replies from that
thread will be in it.

If separate_replies=True _is_ given then the API follows the new
behaviour: "rows" contains top-level annotations only, and a separate
"replies" list containing all replies to the annotations in rows is also
inserted into the result.

Do the concatenating of these loadAnnotations() instead of in multiple
places that call it.

Fixes #1916.

Get rid of "Message not found" messages on the /stream page.

Problem:

The stream requests the most recent n annotations from the
search API, but what it actually gets is the most recent n annotations
_and replies_. In some cases a reply may be in this set while the parent
annotation that it is a reply to is not.  When that happens, the frontend
just displays "Message not found" in place of this thread (parent
annotation and replies).

Solution:

Change the semantics of the search API: only return the top-level
annotations (not reply annotations) in "rows". Replies are always
filtered out from the Elasticsearch query results. "total" (and "limit" and
"offset") only apply to these top-level annotations.

Additionally, always do a second Elasticsearch query to get _all_ reply
annotations to the matching top-level annotations, and return these in a
new "replies" key in the API response.

This means that whenever the frontend calls the search API, it always
gets back all of the replies to all of the annotations that it gets, so
it always has the complete threads and feeds them all into the threading
code, which can safely assume that it has the complete threads and no
longer produces any "Message not found" messages.

Enable highlights by default in the Chrome extension, bookmarklet and embed.

They were already enabled by default in the Chrome extension and
bookmarklet so this only changes the embed.

Highlights are turned on by default in annotator/host.coffee if no
showHighlights option is found.

The extension and bookmarklet no longer need to set showHighlights: true as
this is now the default. The embed (which was not setting showHighlights: true
and so was getting the default setting of false) will now get the new default
setting of true.

For people embedding h into their own pages, they can turn off highlights by
default by including:

    <script>
      window.hypothesisConfig=function(){return{showHighlights:false}};
    </script>

Faster database tests

Remove old feature flag data

Don't log 404s to Sentry

Allow turning on debug query logging with an env var

Remove the 'groups' feature flag

Merge pull request #2718 from hypothesis/2713-change-permissions-when-moving-annotations-btwn-groups

Change perms when moving annotations btwn groups

Keep the permissions of new annotations in sync with the group, when
changing the focused group.

This fixes an issue where a new (not yet saved to server) shared
annotation in group Foo would become a private annotation if you changed
the focused group to Bar.

The issue is that model/@annotation.group gets changed to Foo, but
model/@annotation.permissions still contains a permissions object with
group Bar's ID in it, and as a result the permissions service thinks
that it is a private annotation.

The fix is to update both .group and .permissions when the focused group
changes, and to update the drafts service _after_ doing this not before
(otherwise a copy of the annotation containing the previous group's ID
gets saved to the drafts service, and then later gets read from the
drafts service and overwrites the correct data).

Fixes #2713.

This is deployed and toggled on everywhere, so we can simplify our code
by removing this feature flag.

There's one small wrinkle here, which is that there are still two
distinct designs for the "top bar" -- one for the sidebar and one for
the stream. I've simplified this in this commit, and removed any
reference to groups as a toggle for the two flavours of sidebar, but at
some point we will probably want to unify the two experiences.

Replace angular-websocket with a minimal WebSocket wrapper & fix push notifications after reconnection

When the socket is closed, the 'socket' instance is null.

streamer.connect() exposed the WebSocket instance as
a private _socket property on the response for use
in the tests.

This removes the need for that by capturing the most
recently constructed FakeSocket instance in the constructor.

The additional dependencies in websocket.js (EventEmitter,
util, retry) resulted in an increase in bundle size from
541K to 555K vs. master.

Reduce the size by using 'tiny-emitter' instead of EventEmitter
and 'inherits' directly instead of the large 'util' package.

This reduces the bundle size of app.min.js to 546K

Python 3 compatibility

Don't update timestamps of unsaved annotations

Fix a bug where the updateTimestamp() function in AnnotationController would
try to update the @timestamp property of new annotations that have no
model.updated timestamp because they haven't been saved to the server yet.
This result in nonsense values being generated and updateTimestamp() being
called multiple times per second for each unsaved annotation.

Save permissions to drafts service

Delete unsaved drafts on sign out

When reconnecting to the push notification service after a disconnection,
re-send the configuration messages that specify the client ID
(this is a per-session ID) and the current annotation filter.

Without both of these, the client will not receive future notifications.

We used to have our own WebSocket wrapper, which was replaced in
cf410d0bd2127973739208ef94b2df59cc0b720e with angular-websocket,
since that provided the functionality that we needed on top
of WebSockets, namely:

 * Integration with Angular $scope by wrapping onmessage callbacks
   with $scope.$apply()

 * Automatic reconnection in the event of a disconnection

 * Queuing of messages sent prior to a connection being established.

Unfortunately, angular-websocket also brings in a somewhat unnecessary
dependency on the "ws" module. That got out of date and resulted in no longer
being able to build H with the current version of Node.
See https://github.com/AngularClass/angular-websocket/issues/47

This gave us either the choice of fixing angular-websocket or
replacing the dependency.

This commit removes the angular-websocket dependency in favor of
our own WebSocket wrapper, but unlike before, we outsource the retry logic
to the "retry" package. A benefit of this is that we get better logging
of what is going on during automated reconnects.

Easier to mock in tests.

Simplify the group share links

Fixes #2701.

Clean up debug helpers

Fixes #2716.

Remove unused claim mail merge module

This got renamed in core-js without a semver-compliant version bump. Of
course.

Appropriately encode the tab URL for a query string

Small design tweaks to the group share page

- Put a <br> between the document link text and the document hostname
  (if there is one)
- Double the amount of vertical space between each <li> in the list of
  documents
- Increase the number of characters after which document link text or
  hostname is truncated