The reset logic in visual runs doesn't take the line offset into account. The for loop iterates over the line's char indices, but the line-local index i is then used to index into the original_classes for the whole paragraph.

Consider the string "aa טֶ", which consists of five chars with the following subranges:

• 0..1: a
• 1..2: a
• 2..3: space
• 3..5: ט
• 5..7: \u{5b6}

What happens in detail: When resolving the line 3..7, the code previously iterated over both chars, and for the last char i=2 resolved to the original class of the space, setting reset_from to Some(2). Then, all bytes from 2 to line_str.len() (which was 4), were reset to paragraph level. As a result, the first half of ט was reset to 0.

The result of all this is that the reordered visual runs now slice the ט in half because its level changes halfway through. In addition to a fix for the problem, I added the above example as a test case.