Corrected saved-context trial for the nearby timing pair

The earlier trial completed both fresh controls in six native calls. Both target identities stopped before native evaluation: their newly attached GPS context contained receiver dictionaries instead of the numerical latitude/longitude/height rows used by extraction. The trial is complete and its failed starts remain charged. Never restart it.

The correction builds the context from the sealed original endpoint arrays, saved relative receiver epochs and preselected receiver records. It reproduces the extraction convention exactly: latitude and longitude in radians, height in metres, and relative time equal to receiver UTC time minus the frozen first IMU interval start. Validate this equality before reading originals. No measurement, epoch, physical model, bound, native kernel or stopping tolerance changes.

lll.ilvis0_saved_context provides the checked conversion. The v2 worker privately reuses the frozen executor and overrides context construction only. Old source files, prepared artifacts, failures and allowances remain untouched. Each v2 identity has a new suffix and a separately frozen source/input/environment set.

Finite limits

Two fresh smooth controls at three calls each and the original 370cec66878c4c1cf84b4ceb profiled-removal specified-field-globe primary/repeat at 26 calls each: four identities, 58 native calls, zero optimizer starts. Maximum 402,048 clock queries, six bounded affine solves, four total numerical threads, no retries. No other stretch or model is included.

Qualification is one 33-test batch with a 120-second ceiling: the existing 21 candidate controls, nine saved-context checks and three v2 worker checks. At most 1,024 software callbacks, 81 compiled-motion calls and six compiled clock comparisons. Tests may read sealed metadata and endpoint arrays, never originals. The real-source regression verifies all 554 receiver epochs for each target.

Require both controls before target evaluation, unchanged source reproduction, full two-sided gates, source-to-final movement limits and primary/repeat agreement (relative cost 10^-6, per-axis prediction RMS 0.01 m). Record calls before execution; an incomplete transaction blocks resume. Completed or charged identities may not restart. Preserve finite limits even if a task fails before using its allowance.

Publication performs no native calls. A successful result still needs the complete saved companion/model review before changing the current 51 globe / 0 flat / 2 unresolved findings. The other stretch remains outside this trial.