Final-stretch free-timing comparison

Question and frozen evidence

Stretch b4e44ddc0f130ecd45674c78 is the last unresolved stretch. Its specified-field endpoints have passed both processing-case repeat checks. The separate zero-field timing profiles completed eight conditional fits and two controls in 132 native evaluations; all four within-offset comparisons passed. Both nuisance starts reproduced the smaller cost near a 0.093-second alignment, compared with the original +0.15-second bound alignment. Fixing time does not qualify the full fit.

This comparison releases time and all other original active parameters from the verified lower-cost solutions. It uses only the sealed final-stretch profile evidence. No other stretch, model or new processing assumption enters the target population.

Implementation and checks

analysis/tests/ilvis0_last_free_time_worker.py privately adapts the checked two-case free-time executor. Earlier launchers and candidates remain unchanged. The original bounded trust solver, full native stationarity and repeated-center checks, fresh prediction-stability check and repeat-comparison limits remain in force. Sources, selected input hashes, environment, native kernel, finite plan and software qualification are sealed before launch. Shared source directories acquire one exclusive lock each. Every native evaluation is charged before execution, and interrupted charge transactions block continuation.

For each processing case and original starting solution, retain both verified profile offsets and choose the smaller saved cost with the existing deterministic tiebreak. Check the saved starting cost against a fresh native evaluation. Release time within its original ±0.15-second bounds; never use a conditional profile pass as full convergence. Save residuals, full Jacobians, packet boundaries, parameters and failures. Two fresh controls must pass before targets.

Finite allowance

Each target's fitting allowance includes one starting-cost reproduction and at most 80 solver evaluations. At the preceding profile study's approximately 20 seconds per evaluation, exhausting this allowance takes about two hours. Earlier stopping is possible. Completed or interrupted identities cannot be retried, and unused allowances from earlier studies cannot be transferred.

Unchanged assumptions and acceptance

The selected complete interval, original increments, receiver observations, nominal 200 Hz integration, physical bounds, ±1 degree/hour constant offsets, coordinate weighting and h1_v3_tau60 covariance remain unchanged. Both fixed-zero and profiled upstream-field-removal cases remain separate. No science measurement is repaired or interpolated.

Every target needs the original full convergence gates. Primary and repeat must agree separately within each processing case: relative cost difference at most 10⁻⁶ and per-axis prediction RMS difference at most 0.01 metre. Fresh position stability retains the 1 mm axis-RMS and 1 cm epoch-maximum limits. Failed full checks remain failures; no numerical limits are relaxed to obtain a result.

Even success requires a separate saved-evidence review of all six model primaries, controls and required repeats before changing the stretch's disposition. This trial does not establish global optimality, calibrated statistical significance, processing provenance or the cause of the measured background field. Preserve all earlier results and all new failures.