I do not think in terms of "sorting" encoded-words before presentation,
and I doubt this is the best way to describe the process.
IMHO the encoded-words should be decoded left-to-right, that is, in the
order in which they appear in the message. However, this should not
constrain the ordering of presentation of RTL text. If multiple encoded-
words are needed to encode a single segment of RTL text, the characters
within that segment should ideally be encoded in logical order, so that the
first encoded-word contains the characters that are read first, the
second encoded-word contains the characters following that, and so on.
Thus it may be necessary for a receiving MUA (at least, one that generates
the display output in strict LTR order) to decode multiple encoded-words
before it can display any of them - up to a single line of output.
(However even an MUA that cannot display RTL text may need to buffer
multiple encoded-words in order to generate a visually-pleasing
presentation of the header field, so I don't see this as an additional
burden for supporting RTL text)
It should in any case be possible to decode all encoded-words, in
the order in they appear, into a single string of Unicode symbols and
then present that string. If the charset in any of those encoded
words is implicitly visually ordered then it will be necessary
for the corresponding sequence of Unicode symbols to be marked as
visually ordered also.