Status report: The GPF problem that had been bugging me (and the code) for three weeks has been solved and what I before suspected to be a GC bug turned out to be something else. All aspects are properly triggered and, more importantly, exceptions are properly relayed on the way back for all relevant scenarios in managed-to-managed (inproc, outproc, outproc/interop for SecureMethod).
Testing is underway for unmanaged-to-managed. The latter is actually the most significant difference to whatever can be done with the managed context. These attributes are actually working for unmanaged (read: COM) clients as can be seen in Excel's VBA stopping on an aspect-fired exception in this screenshot.
Still, stability is still an issue under some circumstances and there seem to be a couple of memory leaks deep down in the guts.
(By the way, Ingo asks over there, why I have a stateful variable in the code he's quoting. Answer, because I don't set the completion flag, state is held. SCs aren't uncontrollably stateless. Ah, and, if that variable was public, you could even put an aspect on it and I'll intercept that call as well.)