The firmware does disable interrupts but only disables the one interrupt needed. For example if a mutex is needed between a timer interrupt handler then the only interrupt that is disabled is the timer interrupt, not global interrupts. Additionally the interrupt priorities are set with the direction interrupt being the highest priority, such that it can interrupt other lower priority interrupt handlers.
Thus asking how long the firmware disables interrupts is a loaded question as it depends on which interrupt is disabled. For the direction pin it is the highest priority and has the lowest interrupt latency and does not get disabled, this is how I can meet the above timing requirements.
With all this said the real world does come into play, and when wiring up the smart steppers most customers run wires between controller and smart steppers. These wires end up adding some capacitance to the step and direction pins. This capacitance slows the rise and falling edges of the step and direction pins based on the pin drive strength and load impedance.
The length for wires and the drive strength of the step and direction pin factors into the timing of the direction and step pins and is often the limiting factor.
I also find that a lot of people think running system at high microsteps is a good thing. Specifically with a lot of stepper drivers higher microsteps rates can create smoother moves, but does not increase accuracy. With the smart steppers the closed loop controller provides the smooth motion, and hence the microstep rate needs to be set to match machine accuracy requirement, usually no more than 16x microstepping and if machine is designed well it would have no microstepping, that is the machine would not count on microstepping for increased accuracy.
Running the step rate faster than machine accuracy required provides no advantages for the smart stepper.
I also find that many motion controllers try to change the step and direction pin at the same time, using one assembly instruction. For example they will set the GPIO port for the direction change and step using one instruction trying to increase speed. This can violate the timing requirements of even hardware based stepper driver chips, worse still is the line capacitance on the step and direction pins will dominate which pin changes first. This can create layer shifting issues as you have stated, even without the smart steppers.
A good motion controller will have some setup and hold time between direction change and step, a great motion controller will have this parameter settable.
Of course controllers like Duet are not designed for using external stepper driver boards. Hence often to use hardware for stuff it is not designed for requires some changes to the hardware or firmware, which is why open source projects like the Duet is great.