What the fixed-point checks explain
All 28 evaluations completed in 5.13 minutes, with zero optimizer starts. The six failed-pair endpoints and two passing controls reproduced their saved residuals and derivatives exactly. The original fit and repeat labels remain.
Timing and calibration move together
The two IMU21 zero-field globe endpoints differ by about 19 ms in timing. Under both processing assumptions, their horizontal costs are approximately 5.56418 and 5.57307. Swapping only their time offsets increases cost to about 18.73 and changes predicted horizontal positions by 4.43 m RMS. Moving the complete parameter vector halfway instead gives cost 5.56950, with horizontal changes of only about 4 mm from either endpoint.
Timing therefore cannot be investigated independently of its compensating calibration coordinates. The full midpoint is not an optimized profile point; these checks alone do not establish separate minima or a timing-derivative error. The saved affine corrections make micrometre-scale horizontal changes and do not bridge the repeat disagreement.
The flat fit has an unobserved vertical direction
Changing only gravity correction between the two unsubtracted flat endpoints leaves the horizontal predictions and cost unchanged, while changing predicted height by 668.6 m RMS. In this flat formulation, that direction is invisible to the horizontal objective. This is a concrete consequence of removing the height constraint, not measured vertical motion of the aircraft.
The complete midpoint changes horizontal positions by about 0.66 mm RMS and height by 326.6 m RMS. The full affine step from the repeat endpoint increases actual cost despite a tiny predicted improvement; the original 1.21 mm north-axis repeat difference remains above its 1 mm limit. A midpoint or favorable probe cannot replace an independently converged repeat.
What follows
The completed timing-path diagnostic compares independent finite differences with the timing Jacobian and samples the complete path between endpoints. Its 29 calls find no gross derivative fault. It uses zero optimizer starts and changes no clocks or physical bounds.
The separate horizontal-only result remains one complete globe preference and one unresolved comparison. Its unscored height errors can be enormous; it establishes a horizontal score advantage, not adequate full-flight paths or gyro-only evidence. The original full-position findings remain 53 globe / zero flat / zero unresolved.