[sbarratt]

Claude-generated PR.

Fixes #252.

Problem

With solver="MOREAU", CVXPY moves constraints on variables (x >= 0, nonneg=True, PSD=True, and SOC/exp/power cones on variables) out of the rows of A into data["dir_cones"]. cvxpylayers built Moreau's Cones from the cone dimensions alone, so the layer silently ignored these constraints. On cvxpy 1.9.3 this also caused 26 failures in the existing test_moreau.py / test_moreau_dual_variables.py.

Changes

• Pass direct cones to Moreau. get_solver_ctx passes dir_cones to MOREAU_ctx, which sets them on moreau.Cones (as CVXPY's own interface does).
• Scale PSD direct cones. Moreau expects sqrt(2)-scaled off-diagonals and CVXPY stores them unscaled, so the solver works in x_solver = scale * x. MOREAU_ctx precomputes a scale for each CSR nonzero. The Torch and JAX paths scale P, A and q and unscale the primal, and so does the constant-P/A setup() path. The JAX warm-start cache stores the scaled primal, since that is Moreau's coordinate system.
  • This runs in the solver context, not by rescaling the parametrization matrices at build time. param_prob is cached by CVXPY and shared between layers built from the same problem, so rescaling it in place would compound.
• Direct-cone duals. Moreau returns them separately as z_x. Both backends now append z_x to the dual vector, and _build_constr_id_to_slice maps those constraints after the conic rows. So constraint.dual_variables[0] for e.g. x >= 0 still works, including gradients.

Only MOREAU produces dir_cones, so other solvers are unaffected.

Tests

New tests/test_moreau_direct_cones.py covers each direct-cone kind: nonneg constraint, nonneg attribute, SOC, exp, power, PSD (quadratic) and PSD with a parameter in A. It checks:

• Values against CLARABEL, for Torch and JAX, unbatched and batched, plus JAX warm start
• Gradients against finite differences (CLARABEL at tight tolerances), for Torch and JAX. Exp/power gradients match to ~3e-3, so the gradient tolerance is 5e-3. Moreau gives the same gradients whether the cone is direct or in the rows of A.
• Duals of nonneg and PSD constraints that became direct cones

Locally (cvxpy 1.9.3, moreau 0.4.1, CPU): pytest tests/ gives 288 passed and 3 failed. The 3 failures (test_torch.py::test_sdp and 2 DIFFCP dual gradchecks in test_dual_variables.py) also fail on master. CUDA paths were not tested.

Found while working on this (not fixed here)

Building a second MOREAU layer from the same cp.Problem gives wrong results, even on master and without direct cones (reproduced with MOREAU only; CUCLARABEL/MPAX share the code path but were not tested). get_solver_ctx re-permutes CVXPY's cached param_prob.reduced_{P,A}.reduced_mat in place on every build. I'll leave that for a separate issue/PR.

🤖 Generated with Claude Code

[PTNobel]

I think #248 is the better approach here

[sbarratt]

Closing in favor of #248, which also forwards CVXPY's direct cones to Moreau (and fixes the shared-problem parametrization bug noted above). The direct-cone tests from this PR pass on #248's branch.