[Clear1nsight]

At PR #3945's actual lender333 boundary, a lazy signed-affine warp can feed the three immediate consumers and compose into the next rank-one child's right monomial. This package proves the same four arbitrary caller-array outputs while avoiding 1024 real writes against a materialized comparator that already shares source loads and phase negations. Reads and negations do not improve against that stronger comparator. Recursive child count and owner width are unchanged.

The source-bound construction retains affine translation 458 and global i³, and applies PR #4002's phase compensation to the actual entrance/exit/successor canonical frames. It is a finite f=1 implementation comparison: routing setup, address/control arithmetic and the native general-f tariff remain outside the claimed runtime vector. No complete total-cost, latency, supplier-admission or κ claim is made.

Verification covers all 2048 complex basis columns of the four-array map across three implementations; 512-column independent source warp and successor-child checks; 512 columns for each actual compensated frame; two dense dirty inverse-return witnesses; eight semantic rejection controls; and a constructive eight-word private scratch schedule with explicit entry/return clears. Standard-library registered normal and optimized Python both pass. The source lowering compiler is retained only for optional generation; verification uses the credited PR #3987 sparse evaluator and a separate new dense circuit implementation.

The package is isolated under research/staged544-paid-native-fusion/. Historical base, original card, exact source bytes and replay commands are pinned. Open next questions identify the general-f, paid-setup and whole-supplier replay obligations.

Credit: ToikY/PR #4002 for phase tags; PR #3987 for signed recipes and independent sparse evaluation; Dugongue/PR #3945 for the native words and source compiler lineage, with the complete retained notices for dataisfire, CinnaDavis, Tekalign, icekylinx, DreamingOfClouds, Sagar Thakrar, roger-sato, ArhamTahir and earlier authors. New work by Chris R / Clear1nsight with substantial OpenAI Codex assistance. Same-session AI adversarial review prompted the stronger comparator and narrower write-only result; no independent human or proof-kernel review is claimed.

Frozen publication: exact head 4406317ff99ddef1d3c69ce42eef0c6bbddf64c4, one canonical-main 20ec parent,25 added files. Committed-byte/source-history, official export/integrity and read-only Git-free normal/optimized replay pass, with 26 export files unchanged. Root independently matched all 325324 public bytes. Hosted CI is a separate status. The investigation takes PR4002’s paid-boundary handoff; selected frontier unchanged.