This was exactly the situation for us at PBworks when I wrote my version of OP's system — huge existing business logic and object model in php but with new requirements and performance demands. Move the business logic into a stateful backend REST daemon instead of the frontend web servers, without throwing away all of the existing codebase. In one move you get the ability to load balance more naturally and persist data in-thread instead of e.g. huge memcache traffic for each user request.
The idea is pretty cool actually, but I am wondering: Is the advantage is really that huge over just running a normal PHP stack on the backend server in combination with local memcache?
It is if you need to consider so many objects during a request that the memcached request and deserialization becomes too much time per API operation (requests like visible object ids in an unbounded folder with acls/permissions applied per item)
