Keeping the shutdown state as a local var in window.coffee causes spec failures because window.shutdown can only be called once in the entire spec suite
If either fails to deserialize from the persisted state
then fall back to the default constructors.
This ensures Atom will still start even if the serialization
format changes and the previous stored state can no longer
be successfully deserialized.
Mouse events that occur in the gutter are forwarded to the
rendered lines with the y-coordinate translated to be the
leftmost pixel of the editor to simulate the event originating
from column 0 of the current line row.
Closes#287
In additional, rename `registerViewClass(es)` to `registerDeserializer(s)`.
This moves us to a situation where any kind of object may want to be
deserialized, not just views.
This closes#233. When an editor is focused, what's *actually* focused
is a hidden input element inside the editor. So clicking on a focused
editor was temporarily stealing focus away from the hidden input, and
then moving it back to the hidden input (which the editor always does
when it is focused). Returning false from the mousedown handler when
the editor is already focused prevents this bouncing of focus and
prevents the fuzzy finder from hiding when you click its editor.
Normally, we don't autoscroll to a cursor when the buffer changes. But
this is a change caused by the cursor itself, so it makes sense to
keep it in view.
This prevents the editor from synchronously redrawing in specs with
cursors in invalid locations. Markers are always updated under the
hood before a change event is emitted from the Buffer or DisplayBuffer.
We still wait to trigger marker observers, but if their position gets
read, it is up to date.
When the font size changes and the editor is detached, it schedules
itself to be redrawn the next time we're reattached rather than
updating the display in a detached state.
Detached display updates worked in the past because we didn't need to
be on the DOM to determine horizontal and vertical positions once we
had calculated dimensions once. So it worked to temporarily attach
the editor when calculating new dimensions, and then continue updates
even when it was detached. That now breaks because we can't ask for
pixel positions if we aren't on the DOM.
When translating a logical screen position (columns/rows) to a pixel
position, the editor now builds a temporary version of the line for the
given row. It then uses the DOM range API to insert an empty range
at the correct text node and offset for the given column and determines
the left position based on its clientRect.
Depending on the speed impact, we may want to optimize this by
recycling the existing line node if it exists on screen rather than
building a new one every time. We will still have to build one if the
line we're moving to isn't on screen yet. We could also increase the
chances of the line being on screen by autoscrolling to the vertical
position first, and *then* calculating the horizontal position. Lots
to explore here.
