Rather than changing b.T @ y > tol, this version attempts to use the tolerances from Anderson & Anderson, where tolerances for rho_p and rho_d are different from tolerances rho_A rho_mu and tau / kappa. (Anderson & Anderson don't seem to specify what tolerance they used for rho_g?)

Result:

• test_bug_11617 PASSED
• test_enzo_example_c_with_infeasibility PASSED
• Several other tests (such as test_cholesky and test_solver_select) that specify the solver (cholesky: True or sym_pos: True) now FAILED, possibly because the new tolerances are too high to be numerically stable with those solvers.