Meanwhile, in this form, I personally find the test fully convincing that Windows doesn't properly unwind the rest of non-volatile regs and we need to just restore %rbp and be good (while zeroing out other regs is maybe not necessary in practice). Also, I tested separately with MSVC and /EHa (to do it quick and avoid any complications which can be affecting a more complicated test) that access violation in WM_CREATE does not get caught and app just aborts this way (so executing SEH handler through user mode callback doesn't work on Windows).
So maybe if Alexandre thinks adding a complicated test for such an unwind is not really necessary we can just go with the change restoring rbp and be good?