[theon199]

Finite complex-supplier circuit improvement; no new kappa. Carrying whole-role signs through the frozen PR5599 word reduces its complete ordinary scalar allowance from 220232 to 219258, with all 112 final restorations paid. The local word has 25253 events instead of 25475. Its 18 full caller invocations save 17532 real-component primitives. This is an exact tariff comparison, not a measured runtime improvement.

The complete MOVE itinerary, 748 independent public inputs, 540 helper targets, owned lifetimes, ADD operand order and coefficient magnitudes remain identical. Negative nonunit scales retain their positive magnitude. The paid and target histograms are unchanged, so selected conditional kappa remains 0.00310045.

Reproduce with Python 3.12, standard library:

python3 -B scripts/research_package.py --package research/p104-sign-gauge

Expected: normal/optimized PASS_P104_SIGN_GAUGE, PASS_P104_SIGN_INTEGRITY, and PASS_REGISTERED_RESEARCH_PACKAGE. The package supplies the full transformed word, deterministic constructor, proof and exact receipt. Checks cover every source cut, equal per-role prefix envelopes, all 571536 endpoint coefficients, chart/owned-state accounting, six semantic controls, actual source/receipt/word corruptions, lossless gzip repacking and read-only Git-free replay.

Base: df0137efd35e11b4445734bf612bdd4c7ebc330c. Source: the accepted PR5599 word at 09dd0965711ca7eeb169057a420abab1c9052137; its word and tariff are bound by Git blob and SHA256 at the base. No unmerged objects or network are needed for scientific replay. Only research/p104-sign-gauge/ is added.

OS seed 34 selected the complex-supplier track. Three candidates and bounded overlap checks are in CLAIM.md; there was no pinned sprint issue. Ross Richardson's PR919 and Scratch-net's PR1153 already supply same-frame sign-normalization precedents. This extends the application to the actual selected P104 word, across frame moves and owned assignments. Concurrent PR5713/5718 change paid/target profiles on newer words; their gains are not used or added here.

Credit Ali Romman for the source word, tariff and transpose-check method; retain Pierre-Loic Doulcet, Joel Pulikkan, DreamingOfClouds, thu-nmrc, Gareth Davidson, David Arias, Cinna Davis, Jacob Alan Sussman, zxv, Cao Yifan and the cumulative native/BIT/compiler lineage. Prepared for Charan / theon199 with substantial OpenAI Codex assistance; no subagents or independent human review. Apache-2.0.

Native request bytes are unchanged, but native recipes, full caller/admission and rounded/tape implementation are not freshly executed. The five inherited framework assumptions remain. The catalogue card is explicitly marked changed scope; this scalar result does not meet its changed-recursive-profile gate or prove a multiplication theorem. Hosted CI is separate from local validation; the source-base CI documentation failures reported in PR5718 are not repaired here.