While it is likely on many hardware architectures that the context switch will be constant time, or close to it, I disagree that this should be an important design goal for an RTOS developer. First, because RTOS makers don't control the hardware settings on which they are run. Data caching is one simple way in which the prior context might be brought into the registers much faster than worst-case.

Additionally, RMA is necessarily pessimistic. Application developers who care about ensuring all deadlines are met will never stop caring about worst-case times. Still, making parts of the code run slower than needed is never a good idea. Any spare cycles that pessimistic RMA counts as unavailable can still be used in the field–by low priority tasks that don't have any critical deadlines to meet.

If you want completely deterministic application behavior, you'd do well to ditch the RTOS altogether. Low jitter functionality should be placed into hardware directly (e.g., an FPGA), executed in interrupt handlers, or run as part of a "cyclic executive" backgroun loop with deterministic timing. An RTOS can support provability/predictability with respect to deadlines with RMA. But an RTOS will never give you determinism.

Cheers,
Mike