Preselected recording calibration-transfer grid

Version ilvis0-joint-z-profile-followup-v1, study replication. Selection is fixed by docs/ilvis0-prediction-audit-20261011/replication-selection.csv and its receipt: only the earliest predeclared interval, d9d91c8edba5721931856fc9, task 4876a722de50db44fd7a, file ILVIS0_applanix_57289_610_LVIS_38614965.013. Match the original 53-stretch selection, reference and source hash before freezing. Never substitute another recording if it fails. The complete interval is 439 seconds at median speed 793.61 km/h. Selection used no model fits.

Grid: Z gain {−2%, 0, +2%} × constant Z offset {−1, 0, +1 degree/hour} × {specified-field globe, zero-field globe, stationary disc}: 27 identities. Use only bias1_unsubtracted, fixed-zero removal. Original integration, header timing, remaining bounds, original solver and initialization stages match the first grid. Free normalized coordinates start at zero. Reestimate remaining parameters at every node; this tests calibration robustness, not prediction with earlier calibration fixed.

One start/identity, at most 200 native sweeps/start and 5,400 total, no retries or reserve. Count derivatives and both final predictions. The nominal-clock prediction is fixed-parameter, not a clock refit. Execute after the removal grid with one worker and two threads. Save all failures, derivatives, boundaries and residuals separately from original results. Interpret shape first, field only after complete globe-favoring comparisons. Grid nodes are not confidence limits; another date is not proof of independent hardware. Preserve sealed hashes, measurements and environment; charge before starting, retain interruptions as failures, and never restart after completion. No second recording, additional case, prediction replication or corpus rerun is included.