From RFC 2047:

(3) As a replacement for a 'word' entity within a 'phrase', for example,
    one that precedes an address in a From, To, or Cc header.  The ABNF
    definition for 'phrase' from RFC 822 thus becomes:

    phrase = 1*( encoded-word / word )

    In this case the set of characters that may be used in a "Q"-encoded
    'encoded-word' is restricted to: <upper and lower case ASCII
    letters, decimal digits, "!", "*", "+", "-", "/", "=", and "_"
    (underscore, ASCII 95.)>.  An 'encoded-word' that appears within a
    'phrase' MUST be separated from any adjacent 'word', 'text' or
    'special' by 'linear-white-space'.

The reason for this is that things like '@' are syntactically significant in headers and so must be encoded.