Hy Cyril
If you know camera positions you can import them as reference.
You say you also can have prior rotations with really high accuracy... that sounds good, but I think this is harder to achieve even for a robot, so my suggestion would be to set camera locations accuracy to a very tight value (for exampe 0.00001m) and let rotations accuracy set to say 1degree
that should be enough of a constraint and would provide enough freedom to estimate rotations and interior orientation during alignment and refinement wile your initial camera posioions would be leaft almost untouched.
But if you are 100% positive that your encoder-estimated locations and rotations have negligible error, you can import them and apply "quick layout" script. After that you can build the tie points and if you are right and your orientations are good, Metashape will create a sparse cloud without modifying your reference priors.
Regards,
José MartÃnez
Geobit & Accupixel CTO