Testing a faster calculation

All nine small checks pass: three source-transformation checks and six compiled cases. They used one released numerical core. The running noisy-data study is unchanged.

The proposed optimization keeps the wider-precision position calculation and uses ordinary precision for the slopes that guide the optimizer. Predicted positions and residuals must stay bitwise identical in the first checks, and all 26/27 derivative columns must meet the fixed comparison tolerance.

This is an isolated software prototype. No speed improvement or long-record accuracy is established yet. The next gate is a separately frozen long-record benchmark, followed by training and prediction qualification. It does not authorize an observed fit or alter a scientific result.

Verified test evidence · Test scope and limits · Reproducibility · Roadmap.

The completed long-record benchmark preserves positions and residuals exactly but fails the derivative tolerance on all six cases. Its speed improvement does not qualify a replacement for fitting.