Saved linear-correction timing review
Question
Does the correction used by the saved numerical stability check cross an IMU integration boundary? Compare the two failed starts with the same two settings from the preceding, successful noise family. Boundary proximity alone is not a diagnosis; the successful controls are needed too.
Fixed scope
Use the closed fresh calibration only. Select seeds 281474977010692 and 281474977010693, rate-sample errors, specified-field globe, field removal allowed, starts 0 and 1: four existing records. Preserve their failure labels.
Replay each saved bounded linear least-squares correction once, with all 27 columns, its original bounds, tolerance 10⁻¹⁰ and 100-iteration ceiling. Verify its predicted cost change and step norm against the saved stability probe before interpreting the timing step. Use the source-bound, hash-verified clock routine to report which receiver queries cross boundaries.
This is saved-matrix algebra, not a new nonlinear fit or a calibration campaign. It permits four charged linear solves, one thread and a ten-minute cooperative ceiling, with no retries, native calls, regenerated measurements or new noise draws. Interrupted solves remain charged. Known-answer software tests must pass before reading the four study matrices.
Next gate
Publish both controls and both failures, even if a linear replay does not match. The result can identify a useful derivative test; it cannot qualify a failed fit or prove a physical cause. Any native probe, revised solver or fresh campaign still requires its separate reviewed finite plan.