An analytical model joining stochastic geometry and queuing theory is devised to study the performance of adaptive LoRa networks with dynamic Spreading Factor (SF) allocation. LoRa devices are perceived as interacting two-dimensional Discrete Time Markov Chains (DTMC)s. Each chain jointly tracks the number of packets in the buffer and the node's protocol state while accounting for Duty Cycle (DC) restrictions and quantifying the imperfect orthogonality of SFs. The network performance is characterised in terms of coverage, delay and Pareto frontiers under different orthogonality assumptions and for various adaptation settings highlighting insights useful for the design of application-aware decentralised or semi-decentralised SF adaptation schemes.