I think that #485 wasn't quite the right fix for #474 - applying locked axes to an inertia tensor zeros out one or more of its rows, making the matrix singular - taking the inverse then results in a matrix that's all NaNs, making the delta_ang_vel also all NaNs, which causes the angular momentum to be unaffected by anything that timestep (including applied torques!).
The problem with the status quo ante seems to be that a similar thing happened - rather than trying to invert the (singular) locked inverse inertia, the calculated Jacobian was (near-)singular in a lot of cases when axes were locked, and depending on round-off errors you'd either get NaNs or just an outright cancellation of the angular velocity
Solution
Calculating the delta-ω from gyroscopic effects with an unconstrained inertia matrix and then applying the locked axes to that angular velocity increment should be more reliable, and should avoid undesirable NaN values showing up
Objective
I think that #485 wasn't quite the right fix for #474 - applying locked axes to an inertia tensor zeros out one or more of its rows, making the matrix singular - taking the inverse then results in a matrix that's all NaNs, making the delta_ang_vel also all NaNs, which causes the angular momentum to be unaffected by anything that timestep (including applied torques!).
The problem with the status quo ante seems to be that a similar thing happened - rather than trying to invert the (singular) locked inverse inertia, the calculated Jacobian was (near-)singular in a lot of cases when axes were locked, and depending on round-off errors you'd either get NaNs or just an outright cancellation of the angular velocity
Solution
Calculating the delta-ω from gyroscopic effects with an unconstrained inertia matrix and then applying the locked axes to that angular velocity increment should be more reliable, and should avoid undesirable NaN values showing up