Try fix Korean IME on iOS

[kerams]

What does the pull request do?

Tries to fix #16956 on iOS.

What is the current behavior?

What is the updated/expected behavior with this PR?

See #16956

How was the solution implemented (if it's not obvious)?

With ㅅ in a text box, when the user types ㅓ, iOS invokes IUIKeyInput.DeleteBackward to remove ㅅ, and then IUIKeyInput.InsertText with 서. Further typing ㄴ should result in another delete followed by the insertion of 선.

However, the second IUIKeyInput.DeleteBackward does not currently happen and ㄴ is inserted instead, producing 서ㄴ, probably due to how the previous delete was handled internally (resulting in the wrong return value from IUITextInput.TextInRange). With my change it is possible to type 3- and 4-character Korean syllabic blocks, but on only the first syllable. Additionally, manual deletion does not remove the entire syllable but only its last part (선 becomes 서 when deleting), which is the norm for Korean.

Sadly, further syllables continue to be broken, and manual deletes without calling keydown only select the final character. I'd appreciate any tips, especially describing why and how the double key up/down backspace invocation is required.

Checklist

• Added unit tests (if possible)?
• Added XML documentation to any related classes?
• Consider submitting a PR to https://github.com/AvaloniaUI/avalonia-docs with user documentation

Breaking changes

Obsoletions / Deprecations

[kerams]

Does Avalonia do any internal postprocessing of text entered in a TextBox? When I set the implementation of IUIKeyInput.DeleteBackward to be empty, I'd expect to see ㅅ서선 when typing ㅅ, ㅓ, ㄴ, because IUIKeyInput.InsertText appends text, yet the text box correctly contains just 선 despite disregarding the issued deletes.

I don't know how to explain this without the said postprocessing, and not sure I can fix the problem without understanding how this behavior is possible.

[kerams]

Nothing? It's quite embarrassing when a user reports to me that such a basic but crucial component as text input is not working properly...

I really want to get this fixed, but at this point I've spent hours of my time debugging the problem both on Android and iOS, so I could use some hints.

[cla-avalonia]

• All contributors have signed the CLA.

[Gillibald]

"basic"

[kerams]

Is there someone else on the team besides this person that can help?

[Gillibald]

I have already given you plenty of information on the corresponding issue. IMEs communicate with the text view via: https://github.com/AvaloniaUI/Avalonia/blob/master/src/Avalonia.Base/Input/TextInput/TextInputMethodClient.cs

Some IMEs require a special text store. We try to fulfill that requirement but it might not work as expected. First, try to understand how the iOS inputContext is working. Multiple protocols are involved. We might not implement some needed parts correctly. All you can do is to debug the order of events.

https://developer.apple.com/documentation/uikit/uitextinputdelegate?language=objc
https://developer.apple.com/documentation/uikit/uitextinput?language=objc

[kerams]

So it turns out the issue is very much akin to what's happening on Android. When iOS calls DeleteBackwards and InsertText to replace the last Korean syllabic block, SurroundingTextChanged ends up being raised, and the handler (which executes in the same context) feeds text change events back to iOS, supposedly messing with its keyboard/composition state. The solution is very adhoc, implemented based on AvaloniaInputConnection for Android - when we react to events from iOS that mutate text, we set a flag and then do nothing in SurroundingTextChanged. If you have a suggestion for a better solution, I want to hear it.

a.mp4

[avaloniaui-bot]

You can test this PR using the following package version. 11.3.999-cibuild0054522-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]