Thanks for the reply Frank.
For any readers interested in doing similar, some notes:
I was able to implement a multi-session SLAM system by keeping track of the pose graph constraints of the individual sessions outside of the GTSAM solver.
When a loop closure was established between two different sessions (a primary and a secondary session), the trick was to use that loop closure relative transformation to transform the initial guess for the secondary session nodes into the fixed frame of the primary mission.
Then after that, subsequent inter-session loop closures could be inserted into the iSAM solver - as in the single session case. In particular, the usual iSAM2 LM solver was stable and reliable.
I didn't feel the need for explicit support for multi-session SLAM was needed.
bw,
- Maurice