This is by design for a free-threaded server. COM takes your word for it and allows stubs to use arbitrary RPC threads. You cannot make any assumptions about the thread identity, the RPC thread is picked from a pool and is recycled. Unfortunately it often picks the same one when the calls are sequenced so it will look like it works fine at first. But trouble starts as soon as more than one concurrent server call is made. There is no option to make it selective, a free-threaded server promises to not care. Nor could that work well in practice, it would either scale horribly or induce deadlock.
You therefore cannot use a mutex to implement locking, it has thread affinity. A semaphore is a good choice.