Fixes#521
We were not initializing the dead key state. The dead key state is read
on all translation calls so it must be initialized to 0. This was
working before because coincidentally whatever the memory was always
worked (it probably didn't, but it never didn't work for someone to
notice).
In debug modes, Zig sets uninitialized memory to 0xAAAAAAAA. Amusingly,
this is a valid dead key state value that UCKeyTranslate ignores and
thus produces the correct value. Therefore, we were only seeing bugs in
release modes where the uninitialized memory was a bad value
(consistently 0x1 on my machine) causing the wrong key to come out.
We were getting lucky before, this has always been a bug.
The surface runs on the same thread as the app so if we use the app
mailbox then we risk filling the queue before it can drain. The surface
should use the app directly.
This commit just changes all the calls to use the app directly. We may
also want to coalesce certain changes to avoid too much CPU but I defer
that to a future change.
macOS translates inputs such as shift+tab into the control character tab
(ascii 0x09). Linux/GTK does not translate character inputs except to
printable characters. We don't want control character translations
because these are all handled manually by our key encoder (i.e.
translating ctrl+c to 0x03).
To determine the logical key that was pressed, we previously just
trusted that the translated text would have the right value. But if
modifiers are pressed, the text may not translate.
For example on macOS, Ctrl+C does not produce any text. As a result, we
would fall back to the physical key. On layouts like Dvorak, the
physical key for "C" is "I". This means "Ctrl+C" sequences weren't
working.
Instead, if there is no text or the text doesn't map to a key, we
translate again using no modifiers to try to get the raw text of the
input and then base the key on that.
Fontconfig in particular appears unsafe to initialize multiple times.
Font discovery is a singleton object in an application and only ever
accessed from the main thread so we can work around this by only
initializing and caching the font discovery mechanism exactly once on
the app singleton.
