Improve conditional kappa to 0.00171625567823018 with composition and a frame refit

[mdumitrean]

The composed source-partner/pre-life construction admits conditional κ ≈ 0.00171625567823018 after one additional legal gate-frame refit. This raises the prior composition by about 0.00671578% and exceeds the entire 3046-digit claim of concurrent #2745. Exact rationals are in research/concentrated-prelife-refit/CLAIM.json.

At B1149, the common frame drops from rank six to five, replacing two rank-one children with one rank-two child. The complete local change is H2 − 2H1, with unchanged scalar gates, warp interfaces, dirty-input compensation, inverse cleanup, helper widths and owner allocation. Across 666 calls the complete global change is 666*(H2 − 2H1). All 415 helpers, 72 private FULL buffers, 276,390 owner roles and 66,801 physical slots remain charged.

The original concentrated-prelife-composition package and its pinned source packages are retained unchanged. The new scoped package contains the generator, legal-frame proof, regenerated chained patches, separately implemented word checks, explicit checker derivation, complete cold evidence and portable verification hooks. Its exact fixed-inventory ceiling is C <= 0.0017162556782302, within 1.9800918e-17 of the certified candidate lower endpoint.

Validation:

• Complete fresh public native/global/outer replay: 18 current stages, 18 retained-parent stages and eight strict stages; both native orientations on all 639 arbitrary input columns; 166 physical classes on 512 columns each, with phase and cleanup checks.
• Full supplier/owner/controller rejection controls, 47 supplier and 63 modern outer inequalities, 17 modern outer controls, 19 outer evidence replays and the actual modern ordinary recurrence with 37 sufficient stages. Native/parent receipts are fresh; inherited parent geometry reuse is explicit.
• Upstream registered hook, Git-free read-only export, exact source/claim bindings and meaningful corruption, optimized-Python and output-preservation controls.

Reproduce from the repository root:

python3.12 -B scripts/research_package.py --package research/concentrated-prelife-refit
python -m pip install -r research/concentrated-prelife-refit/requirements-full.txt
python -B research/concentrated-prelife-refit/full_replay.py --work /tmp/FRESH-refit
python -B research/concentrated-prelife-refit/check_admission.py --fresh /tmp/FRESH-refit

Latest-source review: #2745 has exactly the same complete decoded words and call order as the prior composition. Its independently checked alternative fill has the same paid profile; its gain is finer endpoint arithmetic. The full exact claim, source pins and literal allocation comparison are retained. #2736's QQ16 completed-error bound and #2740's restricted dirty-padding obstruction were separately replayed with their conditions recorded. FRESHNESS.json distinguishes complete metadata coverage from source-specific verification.

The supplementary bounded searches exclude 32 public-helper hosting pairs and 238 compatible groups under fixed warp interfaces, and five specific retired-alias profiles even with relaxed packing. Among 914 single-extreme frame closures, B1149 is the only cheaper invoice. These are scoped findings, with reproducible artifacts and explicit open interfaces; they are not global optimality claims.

Credit: Mason Hale (#2696 concentrated assignments), Philippe Van Velzen (#2679 pre-life/refit and #2559 evacuation), Jeremy Kothe (#2673/#2588 order/word work and #2745 concurrent composition, alternative fill and finer grid), and the retained source-partner, native, bit and outer lineage. New work by Marius-Paul Dumitrean with substantial OpenAI Codex assistance, including separate-agent audits. Apache-2.0; source notices retained.

This remains a finite conditional construction and exact exponent calculation. The inherited all-size compiler/common-label, physical hosting/routing, uniform precision/recovery, analytic, bit-supplier and modern outer assumptions remain. The selected repository result is unchanged; human review and proof-assistant verification are not claimed. Open next questions and the revision-5 scope change are in TASK.json.