I have a header file, whose functionality relies heavily on the success of SFINAE. In present g++ 4.6 it works as expected. Should I assume that, my code will behave seamlessly in the same way for all compilers (C++03 compilers) ?
I find this as an issue, because if something differs it won't result in compiler error and silently would change the code flow.
Yes, you may rely on SFINAE to exist and function properly.
If you have a compiler that fails at it, then it's terminally non-conformant and all bets are off anyway.