According to RFC  2047 5(1)

> However, an 'encoded-word' that appears in a
  header field defined as '*text' MUST be separated from any adjacent
  'encoded-word' or 'text' by 'linear-white-space'.

So, it seems like splitting on whitespace is the right thing to do (see MUST).

While your solution works for your case where the charset and cte are utf-8 and q respectively (not a general case for random chatsets and cte), it seems like a hack to get around the fact the header is non-conformant to RFC.

IMO manipulating the original header (value.replace in your patch) isn't something we should do, but @r.david.murray would be the right person to answer how we handle non-conformant messages.