Factor a working transpose by column-pivoted Householder QR, without AA'.
Only structural identities reduce rows: a nonzero free ax_j=0 fixes x_j=0,
irrespective of a's magnitude/sign, and an all-zero zero endpoint is tautological.
Every other numerical rank ambiguity refuses rather than discarding a row.
Original coefficients/bounds are borrowed read-only; all factors are owned.
Gill and Wong (2014), sections 2-3, require an independent working basis:
https://www.ccom.ucsd.edu/~peg/papers/genqp.pdf
Binary64 rank readings and affine reconstruction remain separate from the
refinement caller's unchanged full original-row KKT and geometry admission.
Factor a working transpose by column-pivoted Householder QR, without AA'. Only structural identities reduce rows: a nonzero free ax_j=0 fixes x_j=0, irrespective of a's magnitude/sign, and an all-zero zero endpoint is tautological. Every other numerical rank ambiguity refuses rather than discarding a row. Original coefficients/bounds are borrowed read-only; all factors are owned. Gill and Wong (2014), sections 2-3, require an independent working basis: https://www.ccom.ucsd.edu/~peg/papers/genqp.pdf Binary64 rank readings and affine reconstruction remain separate from the refinement caller's unchanged full original-row KKT and geometry admission.