[ToikY]

Result and scope

Closes the released PR3943 handoff to match its 960-ADD source-return adapter against PR3940's 1152-ADD adapter before comparing them.

At the same encoded entrance, fixed physical outputs, formal chart semantics, unit arithmetic alphabet and 72 arbitrary dirty helpers, the existing PR3943 word saves exactly 192 unit Gaussian array additions. Both words use 104 halves, 104 restoring doubles and 192 rank-one transports. The difference is two rather than four direct updates for each of 96 nonpivot pairs; both retain the same two 384-entry B traversals.

This is an independent comparative audit, not a newly discovered return construction or scalar optimum. Recursive rank mass is unchanged. PR3940's different fused 208-H1/mass-2576 suffix is not contradicted. No supplier, new kappa or selected-frontier change.

Evidence

Scientific head: f73eef845623a42363ec8d883058fab98fb1335d; canonical base 20ecf2891d80b5acf18d980b2164fb44ea7faa90, priority 6. Exactly 11 added files under research/p104-return-contract-match/. Public root tree 87d61024cb9d2e8c23885fd632eb4ecb90b0cd48 equals the locally tested tree; the public delta is one commit.

• Five unchanged source exports with hardcoded SHA256 checks and recorded original Git blobs.
• Actual pinned predecessor schedule builders, with an independently written Fraction evaluator and inverse equations: all 280 encoded basis inputs, chart obligations, 72 helper returns and complete invoices pass.
• Separate dense symbolic block identities on five rectangular/rational B examples; nine destructive controls, including an unchanged endpoint rejected solely for understating two added operations.
• Normal/optimized checks, Git-free read-only export, registered Python 3.12 replay and publication-byte/source-history preflight pass. Source-builder reuse is explicit; neither predecessor arithmetic evaluator is imported. No independent human/kernel review is claimed.

Reproduce, standard library Python 3.12:

python3.12 -B scripts/research_package.py --package research/p104-return-contract-match
python3.12 -B scripts/check_research_commit.py --package research/p104-return-contract-match --check-source-history

Local package replay takes about one second on CPython 3.12.14; runtime is host-dependent. Hosted CI is separate and initially pending.

Selection, overlap and remaining obligations

Handoff-first refinement of p104-side-correction, with an advance claim on PR3943. Main refreshed at entry and before submission; 654 open PR bodies including drafts checked at entry, then all 657 refreshed across seven pages and the repository overlap helper rerun. Live #1166/#1163 reads are bounded, not a complete Discussion/closed-diff survey. New PR3954 explicitly excludes this comparison and updates the downstream two-axis task; its claims are context, not freshly replayed evidence.

The complete original channel, actual binary hyperplane/native/global-phase realization, signed-floor precision, physical owners/holding/routing/control/tapes, opposite orientation, selected cold replay and full supplier admission remain unrun/unproved. The adapter comparison uses the predecessors' formal encoded entrance. Stop this comparison and release the claim; further scalar optimization does not repair the failed recursive allowance. The next consequential task needs a changed all-role encoding/consumer and a favorable complete invoice beyond PR3954's rejected scaffold.

Credit

Chris R / Clear1nsight with substantial OpenAI assistance owns PR3882's representation and both PR3940/PR3943 adapters, including the 960-ADD construction. Credit parallel PR3954 and the retained Rohan Arun/PR1972, Chafik Boukhalfa/PR1991, Adam Gall/PR1940, terminal-floor/occupancy, Josie Elliston/E8, Sagar Thakrar/PR1012, Douglas Colkitt, Dominik Scholz and caller/native/compiler lineage. Full notices and exact source pins remain in the package.

New comparative audit and checks by ToikY with substantial OpenAI Codex assistance, including separate AI algebra review/test implementation. Apache-2.0. Maintainer edits disabled; no merge or frontier change.