Incorporate QP dual residual into daqp_refine_active

[Copilot]

daqp_refine_active only corrected the LDP primal residual (r_p = M·u − d), ignoring the dual residual (r_d = u + M^T·λ). For QPs with a non-trivial Hessian, errors in r_d can degrade solution quality after refinement.

Changes

src/auxiliary.c — daqp_refine_active

When Rinv or RinvD is available (non-identity Hessian), runs a full KKT-informed correction instead of the primal-only fix:

Step	Operation
1	r_p = M·u − d (primal residual, as before)
2	r_d = u + M^Tλ* (dual residual, new)
3	eps = Rinv·r_d (solve R·eps = r_d via packed upper-triangular Rinv)
4	rhs_lam = r_p + M·eps (enriched RHS)
5	delta_lam = (LDL^T)⁻¹ · rhs_lam
6	λ* += delta_lam, u -= r_d + M^T·delta_lam

Step 6 simultaneously zeroes both residuals. No new allocations — work->xold (unused during daqp_ldp) holds r_d, and the first n elements of the existing n+1-element work->zldl temporarily hold eps before the LDL D⁻¹ step overwrites them.

Pure LDP case (Rinv == NULL && RinvD == NULL) is unchanged — falls through to the original primal-only correction.

interfaces/daqp-eigen/tests/08_qp_refinement.cpp

New test for the QP-informed path: sets pivot_tol = DAQP_INF to force daqp_refine_active on every solve and checks the result against a reference solve at default settings.

---

📍 Connect Copilot coding agent with Jira, Azure Boards or Linear to delegate work to Copilot in one click without leaving your project management tool.