Exactly, the reconstruction is subjective and can lead to damage. It is better to make copies of the parts and reconstruct those

you convinced me with this concise argument

so the fragment-recombining future looks like robot archaeology: fragment collection done by careful robots documenting and keeping statistics on each grain of sand or potential fragment, those datasets will constantly be repuzzled with gradient descent and better and better loss functions, and humans go to the museum or dataset download page of reconstructed artifacts, or at least watch the funniest ones among an unlikely large cloud in likelihood space...