Own sparse LDL-transpose factors of [diag(P), A'; A, -diag(H)]. H is zero
for equality rows and s/z for inequalities. Symbolic elimination and ordering
are reused within this solve; only diagonal values change between iterations.
Davis's left-looking elimination-tree LDL algorithm computes actual fill.
Signed static/dynamic pivot stabilization affects factors only; every solve
refines against the original, unregularized KKT matrix.
Twofold product sums retain cancellation in elimination, substitution and
original-system residuals; rounded residuals cannot certify their own error.
The regularization/refinement constants retain Clarabel 0.11.1 defaults.
Finite refinement is not a convergence proof: the interior-point owner must
still establish the original primal/dual residual and objective-gap tests.
Own sparse LDL-transpose factors of [diag(P), A'; A, -diag(H)]. H is zero for equality rows and s/z for inequalities. Symbolic elimination and ordering are reused within this solve; only diagonal values change between iterations. Davis's left-looking elimination-tree LDL algorithm computes actual fill. Signed static/dynamic pivot stabilization affects factors only; every solve refines against the original, unregularized KKT matrix. Twofold product sums retain cancellation in elimination, substitution and original-system residuals; rounded residuals cannot certify their own error.
The regularization/refinement constants retain Clarabel 0.11.1 defaults. Finite refinement is not a convergence proof: the interior-point owner must still establish the original primal/dual residual and objective-gap tests.