We had very inconsistent use of the braille end-of-line indicator.
Current feedback on the list is that it makes sense to include it at
the end of all editable lines as well as at the end of a line we know
is code/preformatted text. There was also an indication that making
things configurable would be nice.

This commit is the first step of the first item: It removes the
web and webkitgtk scripts' override of _generateEol and places the
agreed-upon rules (end of all editable or code/pre-formatted lines)
into the default's _generateEol.

The second step of the first item is to go through all of formatting.py
and remove the role-specific inclusion (or exclusion) of the indicator.
That will be done in a subsequent commit.

Making it fully configurable is on the TODO list.

Most relevant when building from release tarball.
The meson script already allows to not be inside a git tree by falling
back to setting revision='' when git returns no revision.
Let's go one step further and set revision='' also when git is not found;
if this were a git clone, the system would likely have git installed.

Some implementations return an offset of -1 if the caret is not in the
text object. Currently we are treating that as an error condition,
leaving it up to all the individual callers to correct (e.g. via
`max(0, offset)`). It would be easier and simpler to do the correction
in AXText.

Commit c4a7b391 handled the case where Orca is immediately quit after
a new window is activated and the KeyBindings.addKeyGrabs() gets called.
But it wasn't handling this scenario for the modifier grabs which is
(currently) taking place in the default script. For now handle that
with one additional check.

Spiel is getting it's first tagged release, so let's bump the
version so it can continue to be user-tested.

Those testing with Spiel should run `meson update subprojects`
before `setup`, then re-build. See the README for more details.

Commit 18d012de modified getFirstCaretPosition to take children
into account which makes sense in some cases (e.g. containers with
text children), and not in others non-editable combo boxes. We need
to audit all callers of this function and see where it makes sense.
Until then restore this function's original behavior.

Fixes issue #465.

Some users may want to build providers from source themselves, so
offer basic guidance on how to do this. Include a note indicating
users may need help from their distribution for build configuration.

closes #463

For a single-speaker scenario, the Spiel API intends for a single
`Spiel.Speaker` to be used. Ensure we're not accidentally halting
speech if an ancillary server is shutdown.

closes #463

Spiel has a stable tag, so let's define it in the subproject to
reduce breakages for testers.

We should only restore keygrabs if the key pressed during bypass mode
is not a modifier key.

Fixes issue #470.

Treat negative width or negative height as invalid so we don't try to
see if it fits in the window's bounding box. That bug should be fixed
in the player or Qt. But we don't want to regress the user experience
by expecting valid coordinates from objects in flat review.

See issue #469.

locusOfFocusChanged was unconditionally refreshing grabs if the new
locusOfFocus was not in a document. We should only do this work if
we were in a document before. Otherwise, things like arrowing in a
browser's menu will refresh grabs every time the selected item
changes.

If the action presenter menu is dismissed and the underlying app does not
fire a window activate and/or focus event, we won't update our location
and active script. This can cause things like caret navigation and
structural navigation to not work. Fix this by doing the following:

* Connect to the "hide" signal of the action-presenter menu
* When the menu is hidden, clean up by clearing the focus manager's
  state, setting the active script, the active window, and focus.

This should cause us to present the restored location as well as
refresh keygrabs.

If old focus is None, refreshing keygrabs is not unnecessary.

When any script is activated, default.activate() adds key grabs without
checking what should be enabled/disabled. As a result, if a web-script
based app was launched, all the navigation commands were enabled by
default and had to be toggled off before native navigation would work.

Prior to b9a11edd, every single time
focus changed, we refreshed keygrabs. That was extremely non-performant
and unnecessary. However, now that we no longer do all that work, we
need to ensure the initial grabs are done with knowledge of what should
be enabled.

* Override default.activate in the web script
* Suspend (or not) navigation commands based on the locus of focus
* Call default.activate so the normal activation work, including adding
  key grabs, can proceed.

That only works in pyatspi.

See issue #300.

* slideAndTaskPanes had no callers so remove it
* drawingView was only called by slideAndTaskPanes so remove it
* isDrawingView was called by the now-removed drawingView and
  slideTitleAndPosition. The latter's logic is outdated. While we
  could fix it, the only thing that was calling it is onNameChanged,
  and we don't seem to be seeing that event when the current slide
  changes. In addition to that, testing slide navigation with Orca
  running seems to regularly cause Impress to crash.

Ultimately we should create dedicated scripts for the different
features of LO, separating Writer, Calc, Impress, Math (if it's
accessible), etc. into their own scripts. In the meantime, there's
no need to keep this unused, outdated logic around.

Writer returns "", 0, 0 for the current character when the caret offset
is at the end of the text. Handling that scenario was removed as part of
code clean up. Restore that check. Also merge the workaround for gnome-
shell, which returns an offset of -1 when the caret offset is at the end
of text, into the default script's echoPreviousWord.

Be sure we don't have an empty queue.

When the user deletes text, we want to present the text at the new
location. Unfortunately, when the user deletes a message in Thunderbird,
Thunderbird fires text deletion (and insertion) events that appear to be
it re-using the current list/tree item. This causes us to speak one or
a few characters unexpectedly.

We do not want to handle these deletions as user-triggered text changes.
Fortunately, we can distinguish these deletions from typical deletions
because Gecko appends the "system" suffix in this case. Use that in
isDeleteCommandTextDeletionEvent to filter out these events.

Every time the selected mail message changes in Thunderbird, we are
spammed by text changed events we don't want to present to the user.
These all have the "system" suffix, so we can ignore them in the
Event Manager rather than queueing them up for processing.

Yet more event spam from interacting with Thunderbird's message list.

Changes in Thunderbird are impacting presentation of deleted items.
Before, the deleted item would actually be deleted. Then we'd get
a selection change event and present it. Now the events seem to be
as follows:

1. The user presses delete
2. Thunderbird changes the name of the deleted item to reflect the
   newly-selected item (and fires an event)
3. Thunderbird fires the selection-change events

Given the above, Orca gets a selection event as expected, but does
not present it because the selected accessible object has not changed.
Thus we need to present the name-change instead.

Thunderbird was returning early for name-change events that were not
from the document in relation to a deletion. So stop returning early.
In addition, the web script was always returning True (event was
handled). That's a bug. Return False if the event wasn't handled.

There were two sources of this regression:

1. We had cached an invalid value in "treatAtTextObject". We should
   always treat editable objects as text objects regardless of their
   contents.
2. When given an offset that corresponds to the end of a string, some
   text implementations return the correct char/line/word/etc.; others
   do not. So even if we treat the entry as a text object, we could
   fail to display anything for braille when the offset was at the end.
   Fix this by reducing the offset by one when it's the same as the
   character count. All implementations seem fine with this.

Orca almost always uses a single source of truth for preferences, to
ensure broken states are hard to achieve.

To keep the UI fully functional, we intercept Orca's attempts to read
from saved settings, until the user willfully saves the settings. The
user must have either kept or change the speech system, so the override
property is cleared and Orca returns to its normal operation.

closes #472

Meson has a tool run with `meson devenv` that collects all the
installed location and ensures Orca is being run with the build-time
dependencies.

This is as simple to use as `bash` and `exit` and makes testing for
volunteers much easier.

Failable initialization of Spiel was being done outside of the
usual `_init()` function. Catch these exception with the rest, to
avoid leaving Orca in a silently broken state.

Sometimes when `speak()` is called we defer to another function which
will create an utterance, so only create on when necessary.

When selecting a fallback voice, the server assumed it would always be
passed a non-empty default.

Use the same function for retrieving the target language to avoid
starting up with an unacceptable fallback.

As a side effect of 621d4efa, Orca got
super chatty in Thunderbird whenever one was typing the subject of an
email. This was caused by our no longer ignoring most name-change events
from Thunderbird due to changes in Thunderbird (reusing objects and
changing their name rather than creating new objects).

In order to solve this particular regression in the user experience,
ignore name changes that are fired on a frame if focus is in an editable
field and the complete text of that editable field is contained within
the new name of the frame.

When GetAttributeRun, for example, is not implemented, or
fails otherwise for a text, we know of no attributes for
it, and do so for the complete content of it.

Thus, it makes sense to claim said range – the complete
text – has no attributes.

That makes users like splitSubstringByLanguage still use
the string, even though the attributes are not known.

Signed-off-by: Markus Göllnitz <camelcasenick@bewares.it>

See issue #477.

Prior to 621d4efa, we were treating
(incorrectly) all name change events as handled, and functionally
ignoring those events. Now that we no longer do so, Orca is getting
chatty. One instance of such chattiness is when fast-forwarding on
YouTube. YouTube briefly changes the name of the frame each time
shift+right is pressed and changes it back when released. We should
not be presenting those changes.

Rather than trying to handle specific instances of name changes on
browser frames, it should be safe to ignore them: When a new page is
loaded, we present the name of that page which typically matches the
frame name. Plus back when we were ignoring all name changes, no one
reported Orca failing to present the changed name of the current window.

We were only doing so when the event was for the active window. However,
we sometimes correct the active window in response to an input event, so
it is possible to get a window:activate event for a window that we set
as our active window. The script associated with that window should be
used for processing object events.

This reverts commit 8325d8af.

We were correcting for this error condition, but did so by setting
the offset to 0. As a result, Orca was saying the first character
when it was expected to say something else. Change that by always
treating -1 as an error when the character at offset is sought.