[kumagi]

Change and mathematical scope

PR3044's explicit handoff requested an external cold replay of its T74 R1+R2 release. This isolated audit package completes that replay and adds two separately implemented standard-library checks: exact rational/GF(2) reconstruction of R2 on both actual words, and fresh source/native/invoice binding plus accounting.

All 13 cold stages pass. The local delta remains H2 - 2H1 < 0 for positive exponent; the complete rank mass is 4,816,512, physical stock is 66,813, and required reserve is 1,510 < 2,048. The published rational endpoint is reproduced, not improved. Selected frontier and predecessor files are unchanged.

The inherited native/outer code was independently executed, not independently reimplemented. Native semantics, chronological owner legality and analytic inequalities still rely on that pipeline; all-size/compiler, physical implementation, precision/tape, caller and universal block-energy contracts remain. This does not admit a supplier or certify the full multiplication theorem.

Task and overlap

• Priority revision 6, independent-moment, refined to the explicit PR3044 handoff. Canonical claim.
• Refreshed canonical base: 997e6a9ca7b4f95d50e366f61412fd640d72a88a; audited PR3044 head: 1f922b25abff2d7fd87fa5b29619160862a06b59.
• PR3044 already supplied author checks/release; PR3050 checks different six-skip words, and PR3114 concerns restore-only optimality on the PR3042/3054 words. This adds exact historical-head cold execution and separate local/binding checks, not a claim to the current leading candidate.
• Full refreshed overlap covered 268 open PRs and 1,902 Discussion records; the earlier stale snapshot and its public correction are documented in TASK.json. Final main fetch and focused PR/Discussion checks found no matching additional cold replay. Unpublished work and every historical diff are outside coverage.

Validation

Python 3.12.13, pinned scientific dependencies and g++ 15.2.0; commands/environment are in the package README and ENVIRONMENT.json.

• Original complete recipe, network disabled after exact-object fetch: 13/13 stages, 745.54 seconds.
• Independent local checker: both published and freshly regenerated words; 12 semantic rejection controls; normal registered and isolated optimized Python pass.
• Separate fresh-run audit: 642 source files, 646 admitted-package files, 24 native files, 22 invoice links, 52,184 recounted idle children, actual hosted-template binding; 11 rejection controls pass.
• python3.12 -B scripts/research_package.py --package research/pr3044-independent-cold-replay passes. This bounded hook checks the local result and saved-evidence integrity, not a fresh native replay.
• Publication-byte check passes for all 46 package/input files; git diff --check passes. Small receipts, stage logs and exact final invoice are retained. No selected-frontier replay or independent full-theorem proof was run. Hosted CI is separate from these local results.

Credit

Nick Petalas (R2, original independent checks and release), Jeremy Kothe (R1), Marius-Paul Dumitrean (joint rank-7 words/allocation), existing independent auditors and the retained native/geometry/owner lineage. Full attribution is in README. Hiroki Kumazaki with substantial OpenAI Codex assistance for analysis, implementation, execution, writing and separate AI-agent review; no independent human or proof-kernel review claimed. Apache-2.0.

Open next questions

The package's next question is an independent interpretation of the R2 native phase/residual-adapter slice, starting from the two source-bound native caches. This would reduce reliance on the inherited engine; artifact equality alone does not prove native semantics. Recheck current family handoffs before starting.