Matched timing before another long screen
The integer implementation preserves all checked native results. Its speed has not been established. Measure that separately before a costly comparison.
Use the same two 128-packet zero-field sources as the completed runtime diagnostic, one under each processing assumption. Evaluate the old and new implementations once per source, in one single-thread worker. Alternate order between sources. Exclude profiler instrumentation and compilation from the propagation timer; preserve the old physical boxes, caps and 0.32-second query. Compare every prediction, coefficient matrix, binding and remainder afterward.
Freeze four bound identities: four reference/source/derivative setups, 512 source and derivative appends, 512 body-bound reservations and 4,096 logical bound steps. Retain 45 seconds per bound with a separate 240-second overall ceiling. No native trajectory oracle, observed fit, LP, retry or protected data. Record both timing outcomes, even if a bound fails or the new method is slower.
This is a small engineering benchmark, not a general performance guarantee. Only after useful matched results should a new six-source full-history screen test original timing limits. Do not repeat the 48-region comparison simply with more time. Diagnose horizontal widths from sealed artifacts separately.