[aliromman]

The emitted G20 caller admits conditional ordinary BIT saving 0.003507410738409, a gain of 0.000001632212840 over PR #5671. One independent H8 helper becomes two independent H4 targets, with no added paid query or owned field. This is a component bound; no whole multiplication kappa is claimed.

Scope, proof, exact checks, source pins, registered replay. Base: canonical main df0137efd35e11b4445734bf612bdd4c7ebc330c, priority revision 7. Contribution head: 715dda75fc4cefb1e492349fa0e331d5f15d0f53.

The exact prefix and suffix cut is new on the PR #5671 word at 31bd16e. Eight scalar ADDs implement the split in each of 18 calls. The complete target change is 36H4-18H8; the paid histogram is fixed. The caller has 11,022 independent inputs, 48 owned fields, 56,223 queries and 443,238 events. Both target normalizer directions and 72 route passes are charged. The balanced 4+4 split is selected over 7+1. A new cut census finds no further split in the emitted common-frame domain with a helper pivot from its boundary group.

Normal and explicit optimized registered replay pass. All 685 local columns pass native and separate adjoint checks at 101 and 10201. A separate global replay checks every independent and dirty column and all chronological query descriptors. Two rational interval engines enclose the raw crossing in (.00350741073845,.00350741073849). Fresh positive fees, compiler thresholds and five ordinary levels give the stated bound. Contaminated cut vectors, missing input/output bases, a stale epoch and an extra old target reject. The committed export passes integrity and Git-free optimized replay.

python3.12 -B scripts/research_package.py --package research/g20-helper-split

Python 3.12 and the standard library suffice. Committed-byte checks pass. The inherited native, ROW/spectator, owned erasure, compiler, backend, precision, tape and analytic recovery premises remain. No huge uniform program or rounded multiplier is executed. Full-history ancestry audit is unavailable in this shallow checkout. Author-side checks are not independent scientific acceptance. No merge or selected-frontier change. Hosted plan job121407969468 ran449 shared tests and failed the two unchanged Current CI navigation checks for the selected review and source_manifest links;21 tests were skipped. The package replay job was skipped. This is the already reported main regression, with repair PR5659 pending. Local scientific checks remain separate.

Credit Dugongue PR5657/5665, ai-nanachi PR5512, Ali Romman PR #5599/5609/5662/5667/5671 and all retained native, helper, controller, ordinary and stock notices. PR #5599 supplies the general split construction; this package supplies its new G20 cut, physical implementation and complete admission.

Open next questions: permit a different physical pivot/output role with complete ownership and routes; find a paid-query removal that uses the new independent input; or bind a stronger compatible Gaussian caller after a fresh overlap check.

Prepared by Ali Romman with substantial OpenAI Codex assistance; Apache-2.0.