Regrouped orientation-cover usefulness comparison
The new native construction passes all 24 qualification identities. Compare it with the completed orientation cover using the same byte-identical generated 400-packet inputs, original physical bounds, all three candidates, both processing cases and eight closed yaw sectors. Preserve all earlier results.
Keep the same six training times, 1.80-second prediction, 1 mm declared training allowance, 10 micrometre arithmetic allowance and 1 micrometre center check. Use the qualified regrouped native bounds in place of independent packet remainders. Retain every sector not excluded by a checked necessary constraint. An LP failure cannot discard a sector. Evaluate future truth only after the sector's training region, center and future bounds are fixed.
Report the full union for each of the six cases, including worse or unchanged ranges. Show ratios to both the original broad-box result and the previous orientation cover. Preserve the original component target: every axis in every case must be at least ten times narrower than the original broad-box baseline. Correctness requires all known parameters to survive their sector's necessary constraints and every future truth to remain in its complete union. Even meeting this short-control target is not recording-scale qualification.
Save all native values and 27-column Jacobians, training observations and inequalities, coordinate scales, chosen domains/centers, future derivatives, support certificates and bound terms. These enable subsequent diagnosis without repeating native evaluations. Hash every side artifact through its result.
Freeze 48 identities / 192 native evaluations / 76,800 native packet visits, at most 384 bounds / 153,600 reserved bound packet steps, and at most 288 one-thread LP proposals / 576 exact support certificates. Charge before each operation. One worker reserves four numerical threads. New qualification timings are 1.75–2.07 seconds per broad-case bound plus its native probes; allow a separate 900-second cooperative ceiling, approximately eight minutes expected. This is a new study's allowance, not an expansion of any completed study. No retries, extensions, process signals, observed fits or original-file reads.
Verify the complete qualification receipts, unchanged native/clock sources and all predecessor input hashes before launch. The source/input/environment freeze must include the old comparison and the new execution/serialization code. Partial or failed studies remain terminal and published. A failed correctness gate blocks dependent work; a correct but broad result blocks qualified observational predictions. Actual error-law adequacy, global numerical treatment and independent confirmation remain separate gates. Protected ESA groups remain sealed.