The test is mechanical, which is the point. Scale the render to match the plate, drop the opacity to about forty per cent, and look at the edge. A shoulder that has grown, a cap that has gained a ridge, a base that has widened: all obvious in overlay and all invisible when the frame is judged on its own at the end of a long day.
Forty per cent is not magic. It is the opacity at which the eye reliably catches a two per cent shift without either image dominating.
A frame that fails goes back to the plate and is regenerated with tighter conditioning. It does not go to a retoucher, because the drift is in the generation and retouching leaves the cause in place.
Can this be automated?
The comparison can, with an edge-difference threshold, and it is worth doing on large sets. The judgement about whether a given difference matters commercially still needs a person, which is why we keep it as a gate rather than a filter.
What if the product is meant to be seen from a new angle?
Then it needs its own plate for that angle, cut and verified the same way. Testing a three-quarter render against a front plate produces a false failure every time.