Last maximum absolute normalized physical multiplier times nonlinear constraint value; null denotes unsupported normalization.
Conservative direction, physical certificate, optional arc and generic fixed metadata visits; excludes selector visits and camera geometry.
Last baseline QP's21 scalar fields, including terminal checks; arithmetic qualification does not imply requested exact-gap or nonlinear convergence.
Infinity-norm forward-error bound to the exactly normalized stored dual aggregate.
Largest explicitly computed Float64 affine-row summation allowance, not a nonlinear gate tolerance.
Gap completion passed actual-row checks, or stalled completion passed separately bounded reconstruction checks; the latter does not attain the requested sub-ulp gap.
Maximum absolute multiplier times original affine slack.
True only when gapUpperBound meets the requested target; numerical-limit qualification alone cannot set this.
Shared tilt; strictly negative supplies descent and active-row inwardness.
Complete simplex pricing gap, in the model's objective units.
Normalized-dual gap reconstruction discrepancy bound, not an achieved requested gap.
Outward upper bound for the exact normalized stored-dual gap; differs from the computed gap.
Actual bounded pairwise transfers.
Maximum positive violation across every authored affine inequality.
Error bound between computed and exactly normalized-dual minimum price.
Original pure QP objective, half squared step norm plus gamma.
Actual residual excesses are within separately recorded normalized-dual reconstruction bounds.
Actual returned-row residuals passed their dot-evaluation checks; never rewritten by reconstruction bounds.
A stalled direction passed numerical reconstruction checks; actual returned-row failures remain visible and this is not nonlinear convergence.
Absolute computed simplex mass residual, checked separately from objective gap units.
Separate Float64 reconstruction allowance for the reported metric stationarity residual.
Independently reconstructed maximum |d+sum(lambda*a)| in metric coordinates.
Actual convex iteration stop, independent of the numerical certificate.
Conservative scalar visits, including certificate reconstruction.
Latest optional arc's20 fields and nullable five-field tightening record; it may predate the terminal QP/normal check.
Exact nonzero coordinate union used in the working transpose.
Whether one failed full trial requested selected-constraint correction support.
Latest one-use correction refusal, separate from the qualified mixed direction.
Whether that single minimum-metric correction met its support conditions.
Prepared current-model scalar target; null without an independent correction factor.
Smallest/largest original working singular value; null without a factor.
Largest selected physical equality residual before the single refinement.
Largest selected original physical equality residual.
Largest original physical affine inequality violation.
Tilted coefficient actually tried: nominal or original-row-supported.
Nominal coefficient min(1, physical SQP norm squared), before support work.
Whether nominal mixing required the bounded original-row segment branch.
Whether that branch supplied a raw-row-qualified new physical direction.
Whether the complete untilted affine QP received a numerical certificate.
Working-space support or refusal reason, independent of nonlinear status.
Whether a nonzero selected physical equality residual was corrected once.
Positive-weight constraint rows selected from the baseline direction.
Largest original physical quadratic stationarity component.
Scale of the original physical objective gradient.
Conservative visits charged before scans, factorization and correction.
Last physical normal passed the unchanged finite/residual checks; baseline direction certification is independently necessary.
Last checked objective-row multiplier, or unavailable before a direction supplies it. This scalar is not a multiplier history or independent dual certificate.
Last maximum original-coordinate Lagrangian gradient component; null denotes unsupported objective-multiplier normalization, sentinels denote computed nonfinite values.
Last original objective-gradient scale, independent of geometric error units.
Actual finite positive generic normal tolerance; support floor is its square root and the inner gap target its square, not separate knobs.
Whole-fit11-field path/outcome marginals and selector visits; no path-by-outcome table or selected-path labels. lastFailure can predate later progress.
Cost decreased but failed the unchanged Armijo condition.
Original tilted full trials and subsequent baseline halving evaluations.
Full comparisons using a same-base correction arc.
Actual half comparisons on qualified accelerated paths with fixed correction.
First rejection cause was null or nonfinite actual cost/qualification.
Latest actual rejection cause; null before any rejected trial.
Full comparisons using the nominal physical-norm mixture.
First rejection cause was cost greater than or equal to current cost.
Full comparisons using the original-row numerical-support mixture.
Complete actual qualification passed but a lower-cost or earlier equal candidate was selected.
Generic trial-vector and fixed attribution visits, separately charged.
Complete fixed feasible-fit diagnostics, separate from accepted geometry. All nonfinite quantities have explicit JSON sentinels. These are computed algorithm reports, not independently reconstructed certificates or history.