Unfortunately, no. I wouldn't like everyone in my roster to be able to shake my client window. Instead I would like to implement a sort of white list for this. And it's hard to represent in XEP-0115 hash that somone is allowed to ask an attention query and others aren't.
In that case you would just ignore the attempted shake for people who you dont want to be able to do it (and probably return an error of some sort), but you can still publish that you allow it in your caps if you allow it for some people in your roster, I think your "whitelist" implementation is going to be pretty much in a minority and people are likely to just implement it how it works in MSN messenger and the like with a global on and off, normal users (Aunt Tilly) arnt going to be bothered with individually whitelisting specific users to allow it... Things dont need to be made overcomplicated.
IQ is perfectly compatible with unsupported clients. They definitely reply with an error (if the client isn't an outdated Psi, which was known to swallow unsupported IQ stanzas).
I would defianately vote to use IQ rather than message too. Also regards to the memory leak issue mentioned im not sure what thats all about, we shouldnt be hampering the design of optimum protocol just because of the possiblity that someone might have bugs in their software...
Richard
