[mihir-s-05]

PR #4462's cross-cell transfer was explicitly left without a different-contributor check in #4473 and the handoff thread. This package corroborates its two conditional lemmas at 89548dc186c9c90d459085b6d1be4834977705b9, with a written all-p argument and separately implemented rational matrix checks.

The audit makes the combined rounding ledger explicit: across-cell errors must be added to the within-cell ratio error. The existing constants still fit: 2delta + 2(p-1)/(64p^4) < 1/(32p^3). It derives the correctly ordered nonnormal residual H^-1 - P_L = Q^L H^-1 and the section-agreement radius. Exact six-by-six counterexamples show that dropping either the row or column hypothesis can break the claimed spectral inverse bound.

Only research/delta-cross-cell-independent-audit/ is added. The original claim/verifier snapshots retain exact hashes and attribution and are never executed by this audit.

Evidence

On Windows, CPython 3.12.0, standard library only:

• 511 exact scalar parameter cases, 24 mixed-cell matrices and 384 sign-vector responses.
• Exact nonnormal polynomial identities at lengths 1-4 and an interior/boundary locality fixture.
• Six meaningful failure controls: row-only, column-only, excess rounding, missing unit diagonal, reversed residual order, and near-face agreement.
• Registered metadata and replay, optimized isolated replay, committed-byte preflight (12 package/input files), export integrity, and optimized standalone export replay all pass.

python3.12 -B scripts/research_package.py --package research/delta-cross-cell-independent-audit
python3.12 -O -I -B research/delta-cross-cell-independent-audit/verify.py
python3.12 -B scripts/check_research_commit.py --package research/delta-cross-cell-independent-audit

Scope and coordination

This is a conditional matrix proof with finite supporting checks. It does not establish the ideal Gaussian gaps, production page generator, Schur/tensor/core propagation, native/stock/outer obligations, or the full multiplication theorem. No new kappa or selected-frontier change. #4481's Schur follow-up is separate.

Base and refreshed main: 2a13eb1827784206202db90e5dc8327cf8a250db, priority revision 7, selected conditional kappa = 23/10000. Final overlap snapshots cover 352 open PRs and 3,378 discussion records; local note coverage is sparse and is not a novelty certificate. Canonical investigation: #4498.

Linux and formal-kernel verification were not run. The selected baseline command stopped at an omitted package in this sparse checkout; no selected cold replay is claimed. The optional full-history preflight was not run in this shallow clone. The audited unmerged head is retained as package-local evidence and is not a runtime dependency.

Credit: Ryanson Jonathan / pinterbanget (#4462), James Reed / jwr404 (#4442/#4451/#4452), the #4449 rounding-contract and retained LOG-A/FINE-PRIME/rad20 contributors. Prepared by mihir-s-05 with substantial OpenAI Codex assistance; Apache-2.0. Separate implementation after reading the claim and code; no blind replication, independent human review or formal proof-kernel review claimed.