Will this solution work if the UR caller doesn't pass a phEvent to urEnqueueCommandBufferExp, and then does a command-buffer release? Note, this is what your CTS test does.

Good point. It seems for CUDA and HIP we'd need to create the event regardless if SYCL RT asked for it or not and then cache it in the command-buffer. I can't really see any other option 🤔

Yeah I can't think of another solution either. Can you check if we have wording in the sycl_ext_oneapi_graph spec about this behavior on executable graph destruction? If we're doing extra work ontop of backends to give users a nicer experience then we should at least advertise that. I'm guessing we have SYCL e2e tests already that check this, since I think you said you saw some fail when the SYCL-RT code was updated but not the UR code.

Can you check if we have wording in the sycl_ext_oneapi_graph spec about this behavior on executable graph destruction?

I don't think we're mentioning anything about this in the spec.

If we're doing extra work ontop of backends to give users a nicer experience then we should at least advertise that.

Ok, I'll add something to the spec about that.