Minimize the unchanged diagonal PSD objective on an independent working space.
Householder QR of sqrt(H)Z avoids a Schur normal-equation factor. Exact flat
columns retain their coordinate when their linear term is zero; otherwise a
zero-curvature descent direction requires the caller's complete-row blocker.
Other numerical rank ambiguity refuses without removing a constraint or adding
objective curvature. The caller retains its feasible native state and alone
checks the original full KKT. Gill and Wong (2014), sections 2-3:
https://www.ccom.ucsd.edu/~peg/papers/genqp.pdf
The finite arithmetic and rank test establish no convergence guarantee.
A normal projection at the affine origin is a separate least-residual dual
candidate, even when a flat direction prevents a working-space minimum.
It is not a second-order-consistent LP basis or a subspace-optimum certificate.
Column-pivoted QR's AP=Q*R convention is documented by LAPACK DGEQP3:
https://www.netlib.org/lapack/double/dgeqp3.f
An unavailable normal projection leaves a valid flat continuation available.
Inputs are borrowed after the refinement owner's complete QP validation;
owned dense coordinates have finite-memory limits, not full-row scalability.
Minimize the unchanged diagonal PSD objective on an independent working space. Householder QR of sqrt(H)Z avoids a Schur normal-equation factor. Exact flat columns retain their coordinate when their linear term is zero; otherwise a zero-curvature descent direction requires the caller's complete-row blocker. Other numerical rank ambiguity refuses without removing a constraint or adding objective curvature. The caller retains its feasible native state and alone checks the original full KKT. Gill and Wong (2014), sections 2-3: https://www.ccom.ucsd.edu/~peg/papers/genqp.pdf The finite arithmetic and rank test establish no convergence guarantee. A normal projection at the affine origin is a separate least-residual dual candidate, even when a flat direction prevents a working-space minimum. It is not a second-order-consistent LP basis or a subspace-optimum certificate. Column-pivoted QR's AP=Q*R convention is documented by LAPACK DGEQP3: https://www.netlib.org/lapack/double/dgeqp3.f An unavailable normal projection leaves a valid flat continuation available. Inputs are borrowed after the refinement owner's complete QP validation; owned dense coordinates have finite-memory limits, not full-row scalability.