Matched complete-domain runtime check
Launch only after all 24 exact-report native qualification cases pass. Use the six original sector-zero child domains from the direct-source full-history screen, with all three candidates and both processing cases. Keep their complete 1,600-packet inputs and 7.2-second query, source/reference, conditioning, clock and calibration ranges unchanged.
Run the unchanged and midpoint-only implementations once each in every case. Alternate their order by the fixed parity of model index plus processing flag: three pairs in each order. Compare complete raw and converted reports exactly. Measure each bound without a profiler. Keep preparation and compiler time separate from the measured bound call.
Finite limits and acceptance
- Six identities and twelve bounds; 172,800 maximum logical bound steps.
- Six reference/domain setups and child proofs, 36 parent support checks.
- Twelve source, derivative and point setups; at most 19,200 visits in each separately reserved source stream. No new LP proposal or optimizer.
- Four workers with one numerical thread each; four total threads.
- Keep the original 120-second cooperative ceiling for each bound. Use a separate 600-second overall ceiling. No retries, signals or extra starts.
- Zero native trajectory oracles, original-aircraft reads or protected ESA data.
Every pair must finish and match exactly. Before seeing timings, require a median candidate/predecessor time ratio at most 0.95, with no individual ratio above 1.10. This is a finite engineering screen for a useful speed improvement, not a precision benchmark or guarantee that every other domain will finish. Preserve all timings, failures and reservations regardless of the verdict.
Only a passing result permits preparation of a separate finite missing-future certificate comparison. Neither this test nor local equivalence checks relax the prediction-width target, qualify recording-scale uncertainty, authorize aircraft fits or change any completed study's outcome.