[kumagi]

PR #5408's §1 says the ADD lower bound also applies to bare B_f^-1; the inverse must retain Q^(-tensor f) and Q^(tensor f) around it. This isolated review package gives a source-bound counterexample: for one stage, A^-1=(A+12I)/13 has 53 nonzero terms per owner, so the bare inverse uses at most 5,824 ADDs at every width. At f=2925, the literal extension demands 5,826. The actual negative-bank decoder in §6 already has the correct conjugations.

The review independently derives the fresh-stage factorization and contraction argument and checks all 13 saved marker certificates against all 16 negative banks using a separately implemented row-signature join and constraint graph. No further defect was found in §§2–6 under the stated interface premises. The inherited theorem and certificates remain credited to Ali Romman, Dillon, Dugongue, Cinna Davis and their source lineage; substantial OpenAI Codex assistance and automated subagent review are disclosed. Frozen predecessor files and the selected frontier are unchanged.

Validation: registered Python 3.12 replay; optimized replay; exact committed-byte/source-history preflight; official 23-file export and Git-free optimized replay with PATH=/nonexistent. The receipt covers 12,544 code-polynomial entries, 57,344 signed normalizer entries, 512 conjugated rows/columns, all 13 markers, all 16 banks and semantic controls. See research/fresh-stage-independent-audit/REVIEW.md and VERIFICATION.md.

This is a written/executable independent review, not human or formal-kernel acceptance. Endpoint/producer lineage remains a premise; no full producer/compiler/outer replay, new κ or favorable physical invoice is claimed. The 5824-ADD witness is an upper bound, not an optimal algorithm; the source already has its own smaller recoding recipe.

Responds to the explicit independent-review handoff. Canonical claim/result. Base 32890bc5de034ab8fc93ac88ead7279a6a416d04; reviewed source head 256126eeb1607c0bf52996c3d90490f2b786b94a. Refreshed overlap covers 369 open PR bodies, 4,773 Discussion records and all 7,402 base notes; no competing review found within that coverage.