Given that several of the thumbnailing re-architecture goals are intertwined, it seems like implementing a working solution on VM will help us progress towards a real solution for production. Not having to make each piece production-ready will help us move faster and check if all the complicated moving parts work well together. I expect that once a VM-based full solution is built, it will be the basis for one or more RfCs to solve these issues the same way in production.
The likely core service of this architecture would be Thumbor.