Polish a Solved native diagonal convex QP in its original coordinates.
Exact coefficient copies retain intersected interval metadata; every original
row remains in final feasibility, stationarity and complementarity checks.
Independently feasible and complementary native values first retain their
primal while signed-dual coordinate corrections seek original stationarity.
Otherwise exact interval intersections supply compatible active endpoint equations
for corrections near the original native primal and signed dual. Only
identical coefficients share an equation; every original row and endpoint
remains in independent final admission.
Only a finite candidate satisfying every original KKT condition at the
caller's unchanged tolerance replaces the preceding native primal and dual.
Failed or wrongly selected active sets retain the original candidate, native
status and actual polishing diagnostics. An AlmostSolved or infeasible native
result is never promoted. No extra native solve, slack expansion, objective
amplification or permanent ridge is introduced. One selected path owns the
caller's existing correction budget; no failed path adds a second budget.
Stellato et al. (2020), section 4, describe active-set polishing and correction
against the unregularized KKT:
https://cse.lab.imtlucca.it/~bemporad/publications/papers/osqp-paper.pdf
Float64, finite numerical factors and the existing work budget do not imply
uniform accuracy or convergence. Original KKT admission remains independent
of downstream geometric, anatomical, Float32 and rendered acceptance.
Polish a Solved native diagonal convex QP in its original coordinates. Exact coefficient copies retain intersected interval metadata; every original row remains in final feasibility, stationarity and complementarity checks. Independently feasible and complementary native values first retain their primal while signed-dual coordinate corrections seek original stationarity. Otherwise exact interval intersections supply compatible active endpoint equations for corrections near the original native primal and signed dual. Only identical coefficients share an equation; every original row and endpoint remains in independent final admission.
Only a finite candidate satisfying every original KKT condition at the caller's unchanged tolerance replaces the preceding native primal and dual. Failed or wrongly selected active sets retain the original candidate, native status and actual polishing diagnostics. An AlmostSolved or infeasible native result is never promoted. No extra native solve, slack expansion, objective amplification or permanent ridge is introduced. One selected path owns the caller's existing correction budget; no failed path adds a second budget.
Stellato et al. (2020), section 4, describe active-set polishing and correction against the unregularized KKT: https://cse.lab.imtlucca.it/~bemporad/publications/papers/osqp-paper.pdf Float64, finite numerical factors and the existing work budget do not imply uniform accuracy or convergence. Original KKT admission remains independent of downstream geometric, anatomical, Float32 and rendered acceptance.