I've run into strange behavior with a model that runs fine in one case, but if I change the order of the variable definitions then initialization reaches a different solution and the simulation fails. Oddly, the number of algebraic variables also changes.

Unfortunately, I cannot share the model. I know it isn't very helpful without a test model, but perhaps someone can explain why this might happen?

Hmm, interesting. I will look at the presentation (once it is posted) to see how you perform this magic. Thank you.

I've seen in the change log that numeric initialization has been removed because it is deprecated (r24058). Can someone please explain why this very useful capability has been removed? Is there an undocumented functionality that replaces this?

Numeric initialization is critical for a whole suite of problems where symbolic initialization is not possible (as far as I know), particularly when there are multiple solutions to the initialization problem or when the initial values of state variables is only approximately known. For example, consider a 2D spring-mass-spring system under gravity where the spring supports are closer together than the total length of the springs (causing it to form a triangle or truss). There are two solutions: 1) the mass hangs down in the direction of gravity, and 2) the mass is held up by the springs in a higher total energy configuration. How can I tell the initializer to go to one solution or the other? With numeric initialization all that was needed was to give it some bias to put it in the local region of the solution and let ipopt (or other optimizer) find the exact solution, which is the best anyone could really hope for in complex problems.

Any insight as to why this was done and any possible workarounds/replacement capabilities would be very much appreciated.