Assuming that the perception enters consciousness as a gestalt with all attributes bound, then when attention is diverted and the apple no longer present, then the decay of memory sets in. The two models neural-net, and "holographic" suggest different parameters for how to approach a measurement.
Under a "holographic" model, the connection might become more vague, but I think it could still be recalled, perhaps with the assistance of a tertiary related stimulus: viz. recalling a memory from an scent.
In a weighted-feedback neural-net model, you'd be looking for the moment when the physical link between the attributes decays, I think. Literally, the moment it becomes subliminal.
I'm thinking of McCulloch/Pitts nets specifically, as described in Marvin Minsky, Finite and Infinite Machines, a classic, but not recent source.
A cortex would be connected through layers to shape detectors and redness detectors, and these would connect to associative areas. The strength of the association would be dependant upon the weights of the neural connections between the two areas, and would begin to decay after the stimulus is discontinued.
... Or maybe I'm rationalizing with big words to save face after being foiled at a snarky answer?? Does any of this make sense? Does it help? I'm sleepy.