Getting results like that could mean, that something is inappropriate. It could be the camera parameters, computed positions or rotations or precisions. What is the resolution of the used images?
Aligning using the locked pose is quicker as the app is using these information as determined, so it doesn’t need to compute their positions. But if the parameters are not precise, you won’t get the proper sparse point cloud and then model (your sparse point cloud is quite sparse using the locked positions. For ordinary cases it should be more dense).
Yes, each new alignment in the project takes something from previous alignments. So, each new one could be more precise. Or, if it was wrong at the beginning, it could be worse…
Would it be possible to share your data with us? If so, I will send you the invitation for the data upload.