Research: complete the scalar layer with one clean carry

[petalas]

Summary

Implement the complete PR3094 prelude/A scalar layer with one clean carry, conditional on exclusive temporary access to two operand banks. At W=84, both providers and orientations save 82 clean carry bits and 1,938,818 unweighted logical gates per complete layer.

ADD    signed powers -> wired MAJ/UMA -> restore source and carry
SCALE  guarded shifts + dirty-bank modular negation -> restore auxiliary and carry

The changed bill has 7,677,700 CNOTs and 3,616,848 Toffolis, versus 6,806,482 and 6,426,884 in PR3094. CNOTs increase; this is a logical circuit/access-contract result. Actual native bank ownership, movement, other operations and the complete consumer bill remain open. No supplier admission, new kappa or selected-frontier change.

Only research/rank7-one-carry-scalars/ is added. It includes the uniform proof, literal implementation, exact source pins, complete invoices and reproduction instructions. The full-width saving is 25108W−170254 for 23<=W<=256. PR3081 already supplies the local wired one-carry dyadic ADD; this result composes all actual signed-power ADDs and negative SCALEs while charging the dirty-bank negation.

Evidence

• Before: PR3094's complete 9,992-macro scalar implementation needs W−1 clean carries. After: all 9,992 macros in both literal providers and both orientations rebind to the same native dispatch and guard ledgers, with one clean carry and the entire changed gate bill charged.
• Registered Python 3.12 cold native/capacity replay passes. The union of both stream alphabets gives 124 exhaustive circuits, 314,560 complete truth columns and inverse checks, plus 1,550 fixtures under an independent Boolean-list evaluator. Twelve optimized behavioral groups reject wrong signs, omitted cleanup, dirty carry, stronger hidden alignment, malformed SCALE recipes and altered source bytes.
• Publication-byte preflight passes all 45 package/dependency files at 66ee36401536625714c9937e36bbe7ef115da05f. Canonical main remained 997e6a9 at the submission refresh.
• A Git-free export with exactly the 30 declared runtime inputs reproduces the committed receipt byte for byte.
• Ruff correctness checks and formatting, warnings-as-errors source compilation and internal read-only review pass. The inherited native engine is executed, not independently reimplemented. Hosted CI and independent human review are separate.

TMPDIR="$RUN/tmp" uv run --python 3.12 --no-project python -B \
  scripts/research_package.py --package research/rank7-one-carry-scalars

Use a session-owned external RUN directory with tmp/. The README gives direct replay and exact receipt comparison. Canonical scope and finding.

Credit Cuccaro, Draper, Kutin and Moulton; PR3081/3088/3094/3090/3080; dcharb78, Jeremy Kothe, Marius-Paul Dumitrean, Cinna Davis, DLamdata, eumemic, icekylinx, Tekalign and the retained native/geometry/owner lineage. Source notices remain unchanged. The package records substantial automated development and internal-review assistance.

Merge Danger

Door: two-way.

Blast Radius: isolated.

One opt-in research package. Shared code, frozen inputs, priorities and the selected result remain unchanged. The SCALE auxiliary bank and exclusive mutable-source access are explicit unproved native implementation obligations.