Qualify the corrected dense reference

The exact-arithmetic diagnosis shows that the original dense test used a general LU solve on an ill-scaled triangular factor. Rational forward substitution supports the structured calculation and a proper triangular solve. The former 22/24 outcome remains fixed and unqualified.

Change only the dense reference's two whitening solves to triangular solves. Keep the same physical candidates, generated inputs, coefficient bounds, all 26/27 columns, zero linear-control remainder, error allocation, rank cutoff and endpoint agreement tolerance. No production forecast or calibration fit changes.

Freeze a fresh full 24-case qualification with the same finite limits: one numerical thread, two libraries, a three-minute cooperative allowance, 450 native calls / 15,000 packet steps / 300,000 packet-query visits per case. Totals are 10,800 calls, 360,000 steps and 7,200,000 visits. There are no optimizer starts, noise draws or original/protected reads. Unused charges stay charged.

Passing will qualify this native interface for the stated linear problem. Missing nonlinear remainder allowances must still abstain. The result will not qualify nonlinear coverage or release additional observed fitting.