By "shaping across space" I mean that the shapes of the words might be very different depending on whether there's a space or whether they're on second lines. That basically doesn't happen in Latin script (though one could possibly imagine a very fancy cursive or a stunt). So I'll give an example in Nastaliq. One of the example images[1] on Wikipedia is "خط نستعلیق" (meaning "Nastaliq script") where the "خط" is tucked under the other word when they're on the same line. There's a space between them, so it's a valid line break. If they were on two separate lines, the total width would be a lot wider because of that shaping behavior.

This is not conceptually a difficult problem, you can just shape all possible substrings between candidate line breaks to get their widths, at which point Knuth-Plass will give the optimal line breaks (relative to your objective function). But given that shaping is expensive, you really want to avoid that in the 99.99% or so cases where the width of the word isn't altered by other words beyond space boundaries. That's what the "safe to break" logic is about - letting you know when you can make that assumption, as opposed to needing to reshape to get the precise metrics.

[1]: https://en.wikipedia.org/wiki/Nastaliq#/media/File:Khatt-e_N...