Frozen scope of the recovery qualification trial

Use the three failure examples and completed control selected in the existing recovery preparation. Do not select intervals or repeats using model costs or desired shape preferences. First reproduce each original objective at its saved point; reconstruct its original initialization, physical bounds and observation covariance. Reuse already converged reference fits in unresolved stretches.

• Numerical method: ilvis0-warm-continuation-v1, bounded dense trust-region least squares with Jacobian column scaling, no cold initialization stage.
• Fitting starts: 15 unfinished fits, six completed control fits, three repeats.
• Each start: at most 600 tangent sweeps, including fresh final/stability checks, followed by at most two value predictions. Total: 24 starts / 14,448 sweeps.
• Final stationarity: projected relative gradient at most 1e-4.
• Stability: relative cost change at most 1e-8, relative residual change at most 1e-6; three stable accepted steps or a fresh bounded-correction verification.
• Internal tolerances: ftol=None, xtol=1e-11, gtol=1e-14.
• Repeats: one unfinished fit per failure example, targeting a different model in the fixed order rotating globe, zero-field globe, disc; deterministic fallback to the first unfinished fit if that example lacks the target model. Use the first original processing case available, without looking at costs. Shift each normalized saved coordinate by 0.01 toward zero, with zero shifted to −0.01; keep all coordinates within the original box.
• Physical assumptions: original nominal-200-Hz clock, h1_v3_tau60 GPS covariance, ±1°/hour constant gyro offsets, original remaining calibration bounds, no field removal versus profiled shared removal. Fit all three physical candidates fairly.
• Completed-control agreement: relative objective difference at most 1e-6 and per-axis prediction disagreement at most 0.01 m RMS in the common reference metric.
• Repeat agreement: same objective and prediction tolerances; every repeat and its primary must converge. Calibration parameter equality is not required.
• Qualification: all planned starts pass numerical checks and all control/repeat agreement checks pass. Otherwise report which criteria failed; no automatic rollout.
• One worker, exactly two BLAS threads; no concurrent numerical jobs. No retries, no replacement starts, no allowance reset after interruption.
• Preserve original failures and conclusions. New comparisons always abstain from calibrated scientific decisions. This does not prove global optima, statistical significance, the source of a rotation field, or a sidereal-versus-solar rate.

Build controls and seal sources before preparation. Preparation hashes four compressed originals but does not decode/integrate them or fit observations. Full original and extraction-provenance checks occur after exact finite authorization, before fitting. An expansion to all 41 needs its own qualified method, selection freeze and allowance.