Suppose an observer $A$ is falling into a supermassive black hole — so massive one that gravitational acceleration at its event horizon is much smaller than that on the Earth's surface. Now, suppose $A$ has slowed down by some rocket engine to a negligible speed by time he's at arm's length from the event horizon (and could touch it if it were tangible), after which the thrust is shut off, so that $A$ is now in free fall. Since there're no spaghettifying tidal forces, he easily stretches his arm (which crosses the horizon without this being noticed) and then retracts it (and low acceleration lets this happen before the whole $A$ crosses the horizon). From the point of view of this observer nothing special has happened.
But to a faraway observer $B$ this arm stretching and retraction should look like entering and exiting a very real event horizon, from which nothing is supposed to be able to return. So the description in the paragraph above must be missing some crucial detail that would make this impossible.
What detail is missed here? What should prevent observer $A$ from retracting his horizon-crossed hand?