Boundary-aware recovery trial

Complete: 17/24 fits passed using 2,399 native sweeps. All six completed controls, nine of fifteen unfinished primaries and two of three repeats passed. No test stretch became fully resolved, so the method did not qualify for rollout. See the results and saved-result diagnosis. The finite plan below records the study's original design; do not restart completed work.

This trial tests a numerical repair on the same three unfinished stretches and completed control used by the earlier recovery comparisons. The observations, physical models, calibration bounds and error assumptions stay unchanged.

The new solver takes a bounded linear step, checks its actual nonlinear cost and shortens rejected moves. It removes only roundoff-sized slack at parameter limits. All successful fits still require the original fresh stationarity and stability checks. The known-answer comparison passed 40/40 controls, but it cannot establish success on the aircraft recordings.

The finite plan defines 24 starts / 14,448 native sweeps: 15 unfinished fits, six completed controls and three predetermined perturbed repeats. Each start has at most 600 fitting calls and two diagnostic predictions. No retries or replacement starts are permitted. At the ceiling, runtime is estimated at 10–14 hours, with one worker and two total numerical threads.

All planned fits must pass; completed controls and perturbed repeats must also agree in objective and predicted positions. Original results remain 12 globe preferences, zero completed flat preferences and 41 unresolved stretches. Numerical trial outputs abstain from calibrated scientific decisions.

Reproduce

Prepare software evidence and the separate frozen trial from the repository root:

env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 ~/venv/bin/python docs/ilvis0-boundary-recovery-trial-20261012/seal_controls.py
env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 ~/venv/bin/python analysis/tests/ilvis0_boundary_worker.py --prepare-only

Run the prepared study with a durable worker:

env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 ~/venv/bin/python analysis/tests/ilvis0_boundary_worker.py --detach

The launcher records the frozen finite limits automatically. It preserves exclusive locks, source/input/environment checks, precharged append-only journals and atomic hashed outputs. Interrupted identities remain spent and unresolved. Results publish automatically when the trial completes; --package-only publishes without fitting. Do not restart completed work or reuse its unused allowance.