[Clear1nsight]

Result

This source-bound fusion combines PR #3981's actual copy-free late FULL delivery with PR #3943's compressed query cleanup, using two actual correction consumers. It preserves all 280 input-column identities, endpoint rows and formal frames, and all 72 arbitrary dirty helper returns.

Matched complete module	Before unit ADDs	After unit ADDs	Halves / doubles each	Native mass
PR #3882 encoded correction	5472	4704	2616	2520
PR #3940 fused source-preserving correction	5856	5088	2712	2520

Each word saves exactly 768 unit ADDs, with no extra scaling, native transport, private storage, or unpaid donor restoration. The source-preserving variant returns every X=x at FULL and every Y=x+y in its indexed hyperplane; it avoids the separate post-encoding adapter's extra192 native mass. These are matched constructive invoices, not scalar optima.

Actual chronology and evidence

The original final 416 query-cleanup events define D. Exact extraction verifies D=T[I B], where T has 32 coefficients ±2. After the actual late deliveries and donor normalization are complete, Q=h−Dx. The PR #3882 correction creates q=x_J+B x_N in persistent FULL residuals. In the PR #3940 source-preserving tail, q exists at the exact internal cut after 1360 tail operations excluding the already-paid104 lifts—after every scatter donor is restored and before the inverse pivot stream removes q. At either legal cut,64 read-only unit ADDs implement Q+=Tq, replacing832 old cleanup ADDs.

• A new exact Fraction evaluator runs the entire joined words on280 independent inputs and verifies actual cut rows, residual donors, FULL arithmetic availability, endpoint frames, and every dirty return.
• Each variant's complete chronological native transport subsequence is literally identical to its own matched reference. Prior native phase evidence and final physical/global/all-size qualifications remain inherited; no fresh native recipe audit is claimed.
• Eight actual semantic faults reject, including incorrect cleanup timing, an unrestored quartered donor, wrong source-consumption ordering, and source-preserving cleanup after q is destroyed.
• Seven actual optimized CLI corruption fixtures each pass a positive baseline before rejecting Boolean/float receipt values, duplicate keys, false savings, changed/missing source, or changed frozen predecessor bytes.
• Separate AI peer inspection and bounded reproduction passed; it reuses this evaluator and frozen builders, so it is not represented as a second independent implementation or human/formal review.

The unchanged conditional PR #3954 two-axis gate still fails by 32448. Scalar fusion changes no native quantity. New global caller hypotheses are separate tasks; there is no supplier admission, all-size theorem, selected-frontier change, or new kappa claim.

Scope and provenance

Published fork head fa933662771f241e6db5b70cdad571df8d8d14c0, tree 6e25bd5ed638f936f11db953ffc42e0e81b9cbbb. Exact committed bytes, source-history availability and official export integrity checks passed. All three registered commands passed from the read-only, Git-free official export, with unchanged bytes and permissions. Export SHA256: 1ff3430d59e34b65d0b5ddade1743de30aa73244a39f9699deca68b4ccc1fa91.

Twelve files only under research/p104-fused-late-cleanup/, based on canonical main 20ecf2891d80b5acf18d980b2164fb44ea7faa90, priority 6. Four unchanged predecessor builder exports retain original Git blobs, SHA256 hashes and public heads: PR #3981 cdb7dcf3…, PR #3943 2cba526a…, PR #3882 scientific 9d56eeed…, and PR #3940 51ba88c0…. The verifier authenticates original Git blobs as well as exact bytes; main-only exports require no unavailable unmerged history.

The full 702-open-PR body snapshot at 08:30:14UTC and a fresh latest 100 update check found no equivalent fusion. The helper lacks a Discussion cache; the root coordinator separately checked live boards and posted the scoped claim: #3954 (comment) . Main remained unchanged at the publication checkpoint. This is bounded overlap evidence, not a novelty certificate.

Credit Chris R / Clear1nsight's PR #3981 late delivery, PR #3943 cleanup/return adapter, PR #3882 correction and PR #3940 fused source-preserving tail. The new result is their verified coupled chronology and full matched bill. Credit Rohan Arun PR #1972, Chafik Boukhalfa PR #1991, Adam Gall PR #1940, icekylinx PR #531 and PR #1862 compiler/interpreter lineage, and Josie Elliston, Sagar Thakrar, Douglas Colkitt and Dominik Scholz's E8/P104 work. PR #3954/PR #3965 and the distinct PR #3961/PR #3962/PR #3966 constraints remain credited. PR #3956's existing adapter comparison is not claimed anew.

Prepared by Chris R / Clear1nsight with substantial OpenAI Codex assistance. No merge, closure, selected-frontier, or shared workflow change.

Reproduce: python3.12 -B scripts/research_package.py --package research/p104-fused-late-cleanup. Hosted CI passes at the exact published head above: planning and all three registered replay commands succeeded, with committed inputs unchanged.