How do you balance provider portability with provider-specific features?

I measure portability as migration effort and maintained evidence, not as the existence of an interface. Data formats, stored conversation IDs, cached prefixes, safety controls, quotas, evaluation behavior, and operational tooling can create deeper lock-in than the request syntax. Dual support has ongoing cost, so unused adapters are removed rather than kept as a fictional escape hatch. Critical exits are rehearsed with pinned workloads, capacity checks, and documented differences in quality and guarantees.