I'd like to add that this behavior breaks the Unicode bidirectional algorithm (UBA), and hence invalidates Emacs' claim of full conformance, or indeed of weak conformance, for that matter (so-called 'implicit bidirectionality' -- see section 4.2 of the UBA specifications).
The reason is that section 3.4 'Reordering Resolved Levels' of the algorithm states (I replaced the bullet points in the original by numbers):
> * The characters are shaped into glyphs [...]
> * The accumulated widths of those glyphs (in logical order) are used to determine line
The Emacs line-wrapping algorithm does not use the logical order of the glyphs to determine line breaks, as evidence by the example given in my original post, which I shall link to again: http://imgur.com/Bckn7zP