Work through a KV-cache allocation

Allocate a twelve-block cache as requests arrive, grow, and finish. Then account for a shared prefix without counting it twice.