Hi Hamish (or Harvey?),
as you know from above, the exact use of TEST points is also unclear to me.
CPs simply tell RC which points on different images should be in the same spot in 3D.
The weight setting defines how many images in the calculation process it simulates.
E.g. RC has 7 images of one particular spot but there is a problem.
So you add a CP with weight setting 10, then the CP counts like 10 images (as opposed to the 7 from RCs automatic detection) and corrects the geometry accordingly.
GCPs are the same as CPs only with, as you said, given coordinates - be they from GPS or theodolite survey or other means.
And yes, they will pull the spot to the “real” coordinates within a certain limit.
However, you need to check this - RC gives you the error measurements.
Sometimes it can happen that even though there are GCPs with coordinates, the model will stubbornly remain in a faulty geometry.
Then you need to delete ALL components and re-align…
The steps you mention result froman area that is covered by two groups of images (in the same component) that are not or only insufficiently linked by tie points (the automatic ones detected by RC). This can be remedied by adding sufficient CPs on images of both groups.