Hi Götz,
My goal is to reuse the exact same calibration and lens distortion parameters. Therefore no computation should take place and resolution should not have any impact whatsoever.
Indeed, considering the current computation time, my guess is that RC still estimates new parameters and does the feature detection and matching - and would be fast enough if that was the actual task. 
From the documentation, the prior being “locked” means “camera position and calibration will not be changed”, and is “fixed in further calculations”, so my understanding is RC should consider these cameras as already calibrated. But it is obviously not the case here, why !?
I am not sure I follow your last point. The source cameras are already calibrated, with parameters loaded thanks to XMP files from previous calibration. But the input images are still distorted at this point. The goal of this whole process is to undistort them according to the input lens distortion parameters.