Fixed-endpoint timing diagnostic
The failed primary and repeat for stretch fd7b81e48a4122b149c4925b
have almost identical objectives but reject their local corrections. The
saved audit
identifies time alignment as the largest normalized correction coordinate.
This study checks that direction before another recovery fit.
The scope is fixed before native predictions:
- The earlier matched trial's failed primary and repeat, under profiled removal and the zero-background globe model.
- Its first completed control:
233183f67270980491416678, unsubtracted increments, specified-background globe. The control is selected by the prior task order. - Three endpoints, 34 native tangent calls maximum, zero optimizer starts, no retries, one worker with four numerical threads. The control has ten calls; each failed endpoint has twelve. Failed or interrupted calls remain charged.
- Two identical center evaluations per endpoint, followed by symmetric time-offset perturbations of 0.1 microseconds, 10 microseconds, 1 millisecond and 4 milliseconds. Every other fitted parameter stays fixed.
- For the failed pair only, evaluate the full and half coupled correction from the saved original BVLS audit. No local solve is repeated.
Compare analytic, centered, left and right derivatives in physical position units. Report per-axis errors, signed actual and linearized objective changes, changed packet indices and distances to the nearest packet boundary. Branch indices use the native kernel's left-interval convention and cumulative nominal intervals. The smallest steps may encounter floating-point subtraction limits; larger steps can cross intervals with different measured increments. Distinguish those effects from a wrong derivative rather than declaring every discrepancy a defect.
Keep original increments, nominal clock, weights, physical models, calibration bounds, wider arithmetic and all stopping limits unchanged. Hash original files, selected endpoints, audit corrections, sources and numerical environment before launch. Verify extraction provenance and reproduction of the saved central cost. Atomic results, pre-call journals and exclusive locks support interruption-safe continuation of unstarted identities. Completed work cannot restart.
Expected cost is tens of minutes at recent native-evaluation speeds. This is a numerical diagnostic, not another fit, a revised scientific gate or permission to relabel any stretch. Results determine whether a separate solver change is justified.