Ah, no. The whole point of the IMU factor is to avoid the recomputing: Once the PIM is created from the IMU measurements it is immutable. What's happening here instead is we also compute a Jacobian of the bias change on the result. So, when the bias estimate
changes, we're computing the delta in the IMU factor and update the residual based on that.
Practical advice: do use the best possible estimate of the bias you have, not just zero. So, as your bias estimate evolves over time, give that to the PIM in the constructor.
Best
Frank