Research: share JP8 cleanup coefficient work, saving 19,688 primitives (no new kappa)

research/complex-jp8-cleanup-fusion/CLAIM.md

@@ -0,0 +1,74 @@
+# Local sprint claim
+
+- Track: complex supplier; assignment seed **31**, drawn once with Python
+  `secrets.randbelow(100)` (operating-system randomness).
+- Exact base: `66849e1475c930fd7732b1a0067ffee7e1e95f60`.
+- Selected result at this base: #626/#622/#623 with #543,
+  `kappa = 51110230636731/31250000000000000`. Pending PR titles do not
+  supersede that record.
+- Object: the selected JP8 canonical word's final scalar cleanup. Odd seed
+  selects JP8 rather than CK; no search over alternative suppliers.
+- Hypothesis: commuting elementary additions can be fused or cancelled in
+  the fixed final cleanup without changing its arbitrary-dirt linear map,
+  recursive profile, helpers, source-access interface, or frame boundaries.
+- Method: exact rational commutation identities, deterministic bounded
+  rewriting, full-column replay in both orientations, primitive-cost and
+  prefix-envelope checks, and controls that break dependency or restoration.
+- Deliverable: one standalone finite circuit result (or an explicitly scoped
+  negative result), pinned inputs, executable checker, and exact evidence.
+  No new kappa is promised.
+- Expected finish: 2026-10-13 14:05 Asia/Seoul (05:05 UTC). Implementation
+  stops by 13:55 for verification and handoff.
+
+Three narrow candidates were considered:
+
+1. Fixed-boundary B-phase frame optimization: declined because #698 already
+   exhausts the selected CK case; repeating the same method on a different
+   seed is insufficient novelty.
+2. Raw coordinate-total factory resynthesis: declined because #681 already
+   supplies a verified finite local rewrite of that mechanism.
+3. Selected JP8 final cleanup compaction: chosen. This changes scalar-word
+   dependencies in the restored cleanup, unlike #696's unchanged-macro
+   primitive lowering and #681's raw factory synthesis. Existing gate-move
+   and cold-end work is predecessor work, not claimed as new here.
+
+The central review ledger, selected construction/review, open PR bodies
+(including #651, #679, #681, #683, #687, #696, #698), and open issues were
+inspected. GitHub's pinned-issues query returned no entries; the sole open
+issue was #650, concerning Discussions. No reservation issue was opened.
+The claim is therefore recorded locally. Scope is restricted to this file
+family; any delegated review must stay within it.
+
+Contributor: @ellipsi2, with substantial OpenAI Codex assistance in research,
+implementation, mathematical checks, and writing. Apache-2.0.
+
+## Experiment update
+
+The complete initial scan found no same-edge addition before its first
+noncommuting obstruction. The one evidence-driven pivot stays within the same
+JP8 final cleanup: share the existing coefficient-work stream across contiguous
+non-unit, equal-magnitude, common-source additions. The scalar word itself stays
+unchanged. The new fanout block clears work at its boundary rather than after
+each original scalar macro. The favorable budget was 999 forward blocks and
+19,688 fewer real-component primitives; the exact finite checker confirms this.
+The backward scan has zero eligible blocks. No kappa or global admission claim
+is made. The completed search and verification are described in README.md.
+
+## Completed result and handoff
+
+The finite circuit contribution is complete: 999 forward blocks save 19,688
+real-component primitives, with unchanged exact dirty-register endpoints and
+prefix envelopes in both orientations. See the [result and command](README.md),
+[exact certificate](EXPECTED.json), and [validation record](VALIDATION.md).
+
+The final full-queue overlap refresh inspected 536 open PRs through #717.
+#696 remains the closest related contribution, with a distinct individual-macro
+lowering. New #708 and #710 concern F2 bit circuits, and #717 records format
+limits; none supplies this JP8 shared-coefficient block. No pending patch is
+imported. A refreshed fetch still resolves main to the exact original base.
+No pinned coordination issue exists, so this completed claim stays local.
+
+Submitted: [PR #722](https://github.com/CrocSwap/integer-mult-bounds/pull/722),
+as a research note / finite circuit improvement. The immediate pre-submission
+refresh reached #721; #719's CK scalar-cleanup obstruction is compatible with
+this distinct JP8 primitive-lowering result and is credited in the PR.