Owner repacking on PR622: conditional κ = 0.001635425470780892

[jeremydiviney]

Change and scope

Repack PR622's literal owners to reduce fragmentation, raising its conditional κ from 0.00163542231500688 to 0.001635425470780892. This is a small real packing improvement (about +0.000193%), not a rounding refinement. Exact κ is 408856367695223/250000000000000000; the conservative value is 0.001635425467912872.

The spare-rank profile changes from {1:127,3:19,7:1,21:1,48:2213} to {1:128,3:17,4:1,7:1,22:1,48:2213}, with the same total rank. There are 5,027 changed placements and 3,146 additional packed-offset changes. All 492 helper identities and widths per call, the jp8 ×65 / ck ×7 dispatch, construction recipes, verifier code and source contracts are unchanged.

Dependency: this is stacked on #622 at f69477f4f6d3c7eb1f740b166881bffed4ffa9f6. Its inherited commits appear in the comparison against main while the dependency remains unmerged. Our numerical change, exporter and evidence are isolated in commit 036753e. Our change is confined to research/complex-571-cold14-k492/ and its new sibling repacking-audit directory. The repository's selected certificate is not changed. This follows our earlier #618 contribution, now applied to the newer supplier family.

Concurrent result: #623 at e39a615ca601ce5d62abeb24f3fcdb8644ca1005 now claims the higher κ 0.001635490664920455 with ck ×8 / jp8 ×64. This submission keeps #622's dispatch and supplies a reproducible improvement over that pinned base; it makes no current-record claim.

Validation and reproduction

• Fresh strict default-SPEC package replay: all 64 stages pass, with no reused baseline or provider outputs. 23 owner mutations and 114 assembly mutations reject; phase and generated-identity controls pass. All 51 receipt artifacts and 64 raw stage-log hashes were checked.
• Both exact moment engines, the conservative calculation, strict outer constraints, owner/assembly mutations, phase checks and generated-identity controls are exercised by the unchanged verifier.
• The standard-library tools/replay_repacking.py reproduces PLAN bytes and decompressed OWNERS content from pinned PR622 inputs. Five malformed-input/optimized-mode controls reject before output. All 52 other inherited package files are byte-identical.
• Independent fresh reproduction of unchanged PR622: all 64 stages pass; all 51 receipt artifacts and 64 raw stage-log hashes checked.
• The shared repository make verify suite passes, including 84 isolated test modules and gen4 verification, on our earlier PR617 repacking worktree. Makefile/scripts/tests and all other shared infrastructure files are identical. This broad suite was not rerun on the PR622 worktree; SHARED-REGRESSION-SCOPE.json records that scope separately from the new package replay.

See OWNER-REPACKING.md for exact source pins and commands. Recorded moves, as-run search code, intermediate maps and contributor evidence are included. Simultaneous one-/two-helper swaps found the improvement where simple consolidation and single-helper swaps did not. The search score and restricted minimizer are not admission dependencies; no globally optimal packing is claimed.

The cited PR543 bit supplier is repriced, not rebuilt in the package replay. All-size compiler, common-label/mutable-source realization, whole-call hosting, precision/tape/exact recovery, analytic transfer and bit H1–H5 assumptions remain. This is finite contributor evidence awaiting independent review, not an unconditional theorem or full Lean certification.

Construction credit belongs to jeremykothe2 (PR622 and PR617), DreamingOfClouds (PR604), and all predecessors retained in NOTICE. This repacking, nomination exporter and contributor validation were prepared by Jeremy Diviney with OpenAI Codex assistance.