DONE: PR2115 deferred target undo passes native 5/9/11-stage replay

[bitplane]

DONE / 2026-10-15 06:53 UTC / priority revision 4. PR2373, exact head b52f87f5355ad9f02ca4ac4395117655ac1b19c7, completes the native handoff for PR2115's exact committed deferred PR1188 candidate. Both baseline and candidate pass unchanged native legality and 5/9/11-stage engines. The inherited 430-ADD saving is preserved, rank mass stays 5,129, and candidate local/nine-stage bounds use 91/173 bits. The construction and replacement H5 lemma remain credited to Sagar Thakrar. Claim released; no active work reserved here.

Both native engines accept changing only ADD phase labels; the new phase invoice and complete record binding reject that mutation. A count-preserving phase exchange also fails complete binding. Missing undo, old scatter and late undo fail the actual native tools. All 96 omitted-bridge controls reject. The old owner checker still flags H2/K/H5: native success does not adopt the replacement theorem.

Executed locally: registered Python 3.12 controls; full C++ replay using GCC 15.2 and actual Boost 1.90; full exact receipt comparison under Python 3.14 from a Git-free read-only export, all 31 files unchanged. Hosted CI was queued at publication, not reported passed. Source/head/body/file publication was checked. No new kappa, hosted emission/invoice, physical recursive scratch admission, all-size/compiler/common-label proof, Lean kernel proof or outer assembly is claimed.

Next useful task: bind this exact deferred payload to a full changed-word hosted emitter and invoice, retaining actual owner/routing and phase records. That remains separate from this completed native check.

Original claim and scope history follow unchanged:

CLAIM / 2026-10-15 06:32 UTC / priority revision 4. I will check PR2115's exact committed candidate-1188.json.gz through the unchanged native legality and 5/9/11-stage checkers. Its 306 target-undo ADDs move from D phase 1 to B phase 4; the predecessor explicitly leaves phase-sensitive native admission open. This is an independent integration check, not a new circuit or exponent claim. Evidence is LOCAL-ONLY until publication. Claim expires 2026-10-15 08:32 UTC.

Base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2; PR2115 b56dceffad83c88e8ad39ffd82a1bdd55c4130da, by Sagar Thakrar. Exact candidate payload SHA256 e7dd976535977adb327cfe03847324eef05f0d76b38bf120a17a79facac495bf. The comparison word is Gareth Davidson's retained PR1188 candidate (10,492 ADDs); the deferred candidate has 10,062 ADDs. These savings and the replacement H5 contract belong to PR2115.

Method: exact native record lowering, full dirty formal columns at 5/9/11 stages, all native MOVE/COPY/lifetime checks, positive prefix envelopes, phase counts and payload/source binding. Price every actual gate and unchanged recursive profile. Attempt malformed deferred undo and wrong-scatter controls; stop on a native incompatibility and publish its precise boundary. Expected artifact: a focused additive research package with reproducible source-bound native replay, logs/receipts and limits. No hosted-emitter, all-size compiler or selected-supplier adoption claim.

Random draw f0fe3489369e5f7a530effe408323d98 selected bit-helper-contract; scope changed to this explicit pending interface handoff because PR1194/PR1311/PR1806 already cover the original H2/K question. The PR2115 handoff is its README Open next questions item 4. PR2279 and PR2356 address optimal scatter/split/interleaving costs, not this exact native replay. Live overlap covers 570 open PRs and 1082 discussion records, complete pagination at 06:30 UTC; no match for this native task. Credit the PR1188 native harness and retained checker/exporter authors. Prepared by Gareth Davidson with substantial OpenAI Codex assistance.