Long-history screen of compiled conservative bounds
Run only after every compiled native control passes. This is a new finite study; the earlier screen, including its two time-limit failures, is preserved.
Use the same six generated 1,600-packet sources, sector 4, 7.2-second query, physical domains and six-support training certificates as the earlier screen. Evaluate the separately qualified compiled implementation once per source. Compare reference centers, all coefficient matrices, common-error bindings and training conditioning exactly before retaining a per-axis minimum allowance.
The purpose is to check that the arithmetic change completes the previously slow calculations without weakening their bounds. Report every axis and each identity's elapsed time. A compatible bound is not itself a useful complete prediction range; no full-sector target can pass from this screen alone.
Finite allowance
- Six identities, six setup reservations and 36 exact support verifications.
- Six bounds and 19,200 two-pass packet steps.
- Zero native trajectory evaluations, LP proposals or optimizer starts.
- One worker with four numerical threads.
- 120 seconds per bound, checked cooperatively; 900 seconds overall.
- No retries, expanded allowances, aircraft-original reads or protected ESA data.
Freeze sources, generated inputs, environment and compiler. Compile from frozen sources and retain arithmetic source/binary metadata. Count failed attempts and their support reservations in publication. The norm-constrained body-source interval is not installed by this screen. Its integration and a complete prediction comparison remain separate gates.