In Safari, when the cursor is positioned on either side of a widget, IME input may behave unexpectedly.
Even when the widget itself does not change, an IME composition can trigger reconstruction of cm-widgetBuffer nodes. This causes the DOM structure around the composition range to change during input, which may lead to abnormal IME behavior.
Although the widget content remains unchanged, this unnecessary reconstruction modifies the DOM during composition and can interrupt or corrupt IME input.
What your demo is doing is always collapsing the second character of the document. And the screenshot suggests you’re seeing the issue when starting a composition after the first character of the document. Which will then immediately be hidden. That this disrupts the composition seems rather obvious. Am I missing something?
Demo has some issues. The point I wanted to demonstrate is that, during view.composing, a DOM update can still be triggered even when the only operation performed is decorationSet.map(update.changes), without any decoration recomputation or recreation.
Looking into this, I found two misbehaviors that may be related to what you’re seeing. But since you didn’t provide a clear reproduction, I am not sure whether the patches below help.
