Fix Gamma Mixture example broken by Optim v2.3.2

[bvdmitri]

Summary

Scheduled CI has been failing since 2026-09-21. The only failing notebook is Problem Specific/Gamma Mixture, which errors with Optimisation procedure for point mass estimation did not converge.

Root cause

• Optim v2.3.2 (Compute the line search's alpha = 0 slope from the gradient again JuliaNLSolvers/Optim.jl#1286, released 2026-09-14) changed how the line search computes its initial slope.
• RxInfer's default point-mass optimizer for bounded parameters is Fminbox(GradientDescent()). With the new Optim it now reaches the iteration limit on the Gamma shape parameters as, even though it has effectively converged.
• Bisected locally on Julia 1.10: pinning Optim to 2.3.1 fixes it, while pinning LineSearches to 7.8.1 does not.

Fix

• Pass a custom optimizer to PointMassFormConstraint that uses Fminbox(LBFGS()).
• Add Optim to the notebook's Project.toml.

Verification

• make examples FILTER="Gamma Mixture" (Julia 1.13): 1 successful, 0 failed.
• Results match the currently published page to about 1e-6 (inferred mixing 0.3728 / 0.6272).

Follow-up worth considering: a more robust default in RxInfer's default_point_mass_form_constraint_optimizer.

🤖 Generated with Claude Code