struct TextEdit

One reversible edit: replace a range of the buffer with new text.

Offsets are UTF-8 byte offsets, not String.Index and not characters. Indices are invalidated by the very mutation being recorded, so they cannot survive in a history — but the choice between bytes and characters is a performance one, and it is the difference between a usable editor and an unusable one on a large file.

A character offset costs String.distance(from: startIndex, to:) to produce and the same to consume: a full grapheme-break walk of everything before the edit. Recording one keystroke needs three of them (the edit position, the anchor, the focus) and restoring the caret needs a fourth, so typing at the end of a 10 MB buffer cost ~64 ms per keypress. A UTF-8 offset is O(1) in both directions — String.Index stores the encoded byte offset, so producing one is a subtraction.

Bytes are no less grapheme-correct here, because every offset stored is produced from a real String.Index, which is always on a character boundary. Converting back snaps down (see index(atUTF8Offset:)) so a buffer edited into a different shape can never leave a caret inside a "\r\n" or a ZWJ sequence.