Saved linear corrections cross timing boundaries

Both failed corrections cross all 119 integration boundaries; neither successful control crosses one. All four replays match the saved predicted cost changes and step norms before their timing movement is interpreted.

What this explains

The numerical check proposes a correction using a local linear approximation. At the failed points, the timing correction moves into an adjacent integration interval for every query. The observed cost rises rather than falling as that linear approximation predicts. Both successful controls keep their timing fixed.

This identifies where to test the approximation. It does not demonstrate that the recorded derivatives are wrong, or establish that either failed point is an optimum. A two-sided timing check is needed before changing the solver.

Scope and safeguards

Twenty-six known-answer software checks pass. Four saved 27-column linear problems are replayed once, with their original bounds and 100-iteration limit. All four solves were charged before execution. No nonlinear fit, native kernel, noise draw, original aircraft record or protected confirmation group was used. The failed fit labels and the 53-stretch findings remain unchanged. Calibration and independent validation remain closed and blocked respectively.

Four-record CSV · Fixed review scope · Timing proximity check · Next analyses.