Admittedly, I am a green-horned newbie when it comes to clang, but even so I succeeded debugging the problem with some hints from Doug Gregor (on IRC) in two hours. The fix arrived on short order after the conception of the solution idea.
That I could get this working in a few hours is an astonishing feat (that I am pretty proud of) and a tell-tale aspect of clang's awesome design. Naturally, I still have to survive post-commit review, especially w.r.t. performance regressions; OTOH I am rather confident that I got the semantics right. Some loose ends in testing remain, which I hope to wrap up this weekend, so that I can see my application linking with clang (LLVM) on monday. That will burst up the doors towards static analysis...
PS: Hopefully I won't need months to make this working ;-)