[krzysztofwojt]

PR4997 rules out changing at most two intact source regions after PR4971's carried-frame swaps. This package proves that any number of those regions can change together without improving the complete moment, for every 0<t<1, in both orientations. The original source profile is a global optimum in that fixed-boundary model. No new κ.

The precise model fixes the scalar word, literal signed native maps, start/final charts, owned-center interfaces and each source identity/zero-edge region. Exact subspace propagation forces 565 of 768 free regions and leaves 203 binary choices. Variable elimination over their complete interaction graph has separators of at most three. Every comparison has an all-exponent majorization or power-inequality certificate; two new inequalities complete the three credited to PR4997. Unknown comparisons fail closed. The full paid histogram, including eight rank-eight center queries, is {1:2201,2:412,3:510,4:127,5:40,8:8}.

Only research/global-intact-region-optimum/ is added. It contains the proof, source pins, exact verifier, complete compressed elimination witnesses and an independently written small-graph oracle. Canonical coordination: Discussion #5006, taking the #1166 handoff.

Validation at 390d7255f0e128be54e007962d7febb47697a7c5, CPython 3.12.12 on macOS arm64:

• Both registered normal and isolated optimized replays pass. They retain the source's 632 normalized input rows and check both global minima, all edge charges, fixed charts, inequality coefficients and rejection controls.
• The separate explicit-set oracle exhausts 655 assignments across seven ambient-two fixtures, including a coupled four-variable graph, nonidentity maps and parallel edges. Nonbinary intervals, invalid maps, omitted edges, rank changes and reversed inequalities are detected.
• Metadata/source pins, exact committed bytes and full source-history ancestry pass. The official export/check contains 20 package/input files. Its isolated optimized verifier passes with PATH=/nonexistent in 3.746 seconds; all 21 files including the export receipt remain byte-identical. Export SHA256: 7bf4cfe862580efcdee2aa588decda30d4dff65158c1f2b086810071ee970914.
• Hosted Registered research packages CI passes at this head: onboarding/selection, the package replay and the committed-input immutability check.
• git diff --check passes. Reproduce with python3.12 -B scripts/research_package.py --package research/global-intact-region-optimum.

Scientific and refreshed canonical base: f259dabc7d70e0db8d6d36bfda888f008bbfecc1, priority revision 7. Pre-submission overlap covered 314 open PRs, 4,129 discussion records and all 6,710 baseline notes plus the two new notes; direct live heads were checked again. New PR5016 (be656882128732110c00d08ceff265e71bece0f9) reports a numerical optimum at one exponent over a broader full-contact domain and explicitly lacks an exact certificate. This is an exact all-exponent result for the narrower intact-region domain, not a certificate for PR5016. PR4993's variable-boundary/split-region exploration remains separate. Coverage is advisory, not proof of novelty.

Limits: split regions, changed boundaries/maps/scalar words, frame descents, fresh absolute-phase lowering, physical precision/tape costs, full supplier/compiler/BIT/outer admission, independent human review and proof-kernel verification remain outside scope. The selected frontier stays 0.00302015. The #4953/#4999 mod-I³ line is untouched.

Credit Cinna Davis/1337isnot1337 for PR4997's graph and three inequalities; Dugongue/PR4971; Cao Yifan/intqwq, zxv, Jacob Alan Sussman, James Reed/jwr404 and all retained source NOTICE contributors. Prepared by Krzysztof Wójt with substantial OpenAI Codex assistance. A separate Luna supporting agent checked handoffs/source dependencies and wrote the small-graph oracle; the primary agent reviewed and integrated it. Apache-2.0; source notices remain intact.