What about the idea, which I’ve only just recently seen suggested, that RC doesn’t use those carefully-maximised Features, and/or those Tie-points, and/or the sparse point cloud, in Reconstruction. RC uses those only for the purpose of Aligning the photos, then makes no further use of them, but starts afresh (with the hopefully well-aligned photos) for Reconstruction.
If that’s so, then sensitivity/feature detection/feature count need to be only high enough to achieve good photo Alignment, but apart from that don’t contribute to final result quality.