Preserve shared rate-sample errors in ESA intervals
The ESA adapter converts rate samples into increments by integrating a straight line between adjacent samples. Two neighboring intervals therefore share a measurement at their common endpoint. An independent-increment noise law does not describe that conversion.
Add a separate, explicitly named rate-point noise law. Declare the covariance of white errors per rate sample, stationary changing-bias covariance and bias correlation time for each sensor family. Do not infer point variance from an IMU name, signal variation or an unspecified bandwidth. These declarations are not the continuous white-rate PSD used by the incremental-data component.
Retain twelve hidden coordinates: six changing biases and six preceding white point errors. Initialize the first sample once. At each interval, use the trapezoidal mean of the two endpoint errors. The new endpoint error drives both that increment and the next hidden state. Preserve its covariance exactly, including unequal intervals, all axes and zero/singular components. This is the logged point-conversion law, not the integral of an unobserved continuous bias path between samples.
Compare complete generated increment covariances with independent dense point-error covariances transformed by a trapezoidal integration matrix. Explicitly test neighboring white-error covariance and distinguish the result from the continuous-OU integral formula. Use at most sixteen deterministic cases and twelve intervals per case, one numerical thread and a cooperative two-minute allowance. No random draws, native calls, fits, original reads or protected reads are allowed. Freeze a fresh source/environment qualification.
This component alone does not qualify the native trajectory adapter for its larger error state. That integration, persistent receiver errors, calibration at bounds and nonlinear prediction coverage remain subsequent gates. The incremental-data law and all earlier frozen evidence remain unchanged.