Upcoming CVXPY change (#3525) breaks ConvexProgram.to_conic #20

Hello Tobia, we've been cleaning up some CVXPY code and unfortunately it has a small effect on GSCOPT.

Here's the suggested fix message which was generated by Claude (but it summarizes things quite well!):

An upcoming CVXPY change, cvxpy#3525, moves cone formatting out of the individual solver interfaces and into its own reduction, ConeFormat, appended to the solving chain immediately before the solver. Previously each conic interface applied the layout itself inside apply().

That changes what gcsopt/programs.py:222 gives you:

chain = cp_convex._construct_chain(solver_opts=solver_opts)
chain.reductions = chain.reductions[:-1]   # strips the solver only

ConeFormat is now reductions[-2], so [:-1] keeps it and the program comes back already in the solver's cone layout — Zero rows negated, and multi-cone SOC/exp rows interleaved cone-major. _constrain_in_cone (programs.py:76) decodes CVXPY's stuffed layout instead, so the two no longer agree.

The test suite here catches it. Against a CVXPY build with #3525, tests/test_programs.py::TestConvexProgram::test_to_conic fails with the conic round-trip returning inf where the convex program gives -0.95477…; the other 19 tests pass.

Fix

Filter the solver-facing steps rather than slicing off the last one — the same thing CVXPY does internally in reductions/dcp2cone/canonicalizers/perspective_canon.py, which needed it for exactly this reason:

+from cvxpy.reductions.solvers.solver import Solver
+try:  # cvxpy applies the solver's cone layout as its own chain step
+    from cvxpy.reductions import ConeFormat
+    _SOLVER_FACING = (Solver, ConeFormat)
+except ImportError:  # cvxpy <= 1.9
+    _SOLVER_FACING = (Solver,)

         chain = cp_convex._construct_chain(solver_opts=solver_opts)
-        chain.reductions = chain.reductions[:-1]
+        chain.reductions = [r for r in chain.reductions
+                            if not isinstance(r, _SOLVER_FACING)]

Thank you! and please let me know if you have any questions about this :).