meas_modelfit has some issues with Eigen 3.3.4. I am making this a separate ticket from DM-14305 in hopes that these can be fixed in a way that is backwards compatible.

The first issue is that this does not compile:

ndarray::asEigenMatrix(*ix) = component._mu

    + std::sqrt(_df/rng.chisq(_df)) * (component._sigmaLLT.matrixL() * _workspace);

but the solution to that is trivial: swap component._sigmaLLT.matrixL() and _workspace, though Jim Bosch suggested a solution that reduces duplication with nearby code.

A more serious concern is that the following code in TruncatedGaussian.cc TruncatedGaussian::maximize raises an assertion error that a matrix is empty when running tests/test_truncatedGaussian.py:

Eigen::FullPivLU<Matrix> solver(G.topLeftCorner(n - k, n - k));

To fix that problem Jim Bosch suggests returning a vector of all zeros if n == k.