Tested and running on the test server. Sorry about the strange regex delimiters. Several versions of Perl don't agree on how to escape the embedded delimiters and I blame mod_perl in this case (my copy of Perl agrees with me).

   # Insert spaces to prevent the nodelets from getting too wide.
   # We leave the loopholes of using a bunch of "&nonentity;"s or
   # "<!--> -->" to intentionally make the nodelets wide (intended for
   # /msg'ing to yourself) as the problem is more accidents than abuse
+.
   # "&123" and "&lt" work in some browsers, but we might put spaces i
+n
   # the middle of them (if you don't like it, then remember the ";").
   my $len= 0;
   $text =~ s[(\s+)|([^\s<&]+)|(<[^<>]*>)|(&#?\w{1,10};)|(.)]`
      if(  $1  ) {
         $len= 0;
         $1;
      } elsif(  length( $2 )  ) {
         # $2 is the only case that can be "0" (ie. false)
         my $res= $2;
         my $tot= $len + length($res);
         if(  18 < $tot  ) {
            my $max = 18 - $len;
            my $min = $max - 9;
            $min = 0   if  $min < 0;
            $res =~ s[
               (  \S{$min,$max}
                  (?: (?<!\W) (?![\w\[{(;,/])
                    | (?<![\w\$@%&*]) (?!\W)
                  )
               |  \S{$max}
               )(?=\S)
            ][$1 ]x;
            $res =~ s[
               (  \S{9,18}
                  (?: (?<!\W) (?![\w\[{(;,/])
                    | (?<![\w\$@%&*]) (?!\W)
                  )
               |  \S{18}
               )(?=(\S+))
            ]{
               length( $1 . $2 ) > 18 ? "$1 " : $1
            }gex;
            $res =~ /(\S*)$/;
            $len= length( $1 );
         } else {
            $len= $tot;
         }
         $res;
      } elsif(  $3  ) {
         $3;
      } else {
         my $res= $4 || $5;
         my $add= $5 ? 1 : int( length($4)/3 );
         $len += $add;
         if(  18 < $len  ) {
            $len= $add;
            " $res";
         } else {
            $res;
         }
      }
   `egis;

   return $text;
[download]

It tries to not put spaces in front of any of [{(;,/¹ and not after $@%&* because they can be Perl sigils.

(Updated)

¹ The first 5 because of Perl syntax, the last two because of IE silliness -- the "," is included for two reasons. IE won't wrap on " ," nor on " /".

Update2: I realized that [ will be encoded as &#91; and will be matched separately so I can remove the \[s from the regexes and probably revert to my best-practice method of using [ ] delimiters for regexes despite the mod_perl(?) bug. This also means that spaces won't be inserted in front of other characters that get encoded, namely any of <>], which is probably not worth trying to work around.