Constraining the step removes the predicted large timing jump

All 18 linear proposals stay inside their timing intervals. Both sides and all three spacings propose the same internal boundary for the failed cases. The successful controls retain their existing physical timing limit.

These offsets are parameters of generated-data controls, not measured device calibrations. The 18 matrices describe four related saved points, not independent experiments.

What changed in the proposed step

The earlier full correction could cross a change of slope while compensating with other calibration parameters. This component solves the full 27-column linear problem separately inside each feasible timing interval. All physical nuisance bounds and unidentified directions remain present.

All 18 solves succeeded in 21–25 iterations, below the fixed 100-iteration cap. For the failed cases, the largest predicted relative cost improvement is about 2.1 × 10⁻¹¹, well below the unchanged 10⁻⁸ cost-stability limit. Both sides and all spacings were retained; none was selected because it looked favorable.

What this establishes

The component passes 26 software checks, including independent face enumeration, known smooth and kink minima, physical bounds and a rank-deficient problem with all 27 columns. The saved-matrix review charged exactly 18 solves and used no native calculations, noise generation or nonlinear fitting. Its objectives were independently recomputed from the saved matrices.

These remain linear predictions. The proposed parameter vectors have not yet been evaluated in the physical calculation. They do not qualify the failed fits, change the original 53-stretch findings or provide a calibration threshold.

The next gate is a finite evaluation of every saved proposal using the unchanged physical model. A complete numerical method needs further qualification before another optimization or calibration campaign.

Verified report · All 18 proposals · Frozen plan · Previous physical diagnostic · Next analyses.