Replay deferred target undo through native 5/9/11-stage checkers

[bitplane]

PR2115 moves 306 target-undo gates from D to B but leaves its native implementation unchecked. This replays its exact committed PR1188 candidate through the unchanged exporter, legality checker and freshly compiled 5/9/11-stage engines. Both the 10,492-ADD baseline and 10,062-ADD candidate pass; paid rank mass stays 5,129, and the candidate's exact local/nine-stage bounds use 91/173 bits. The circuit, 430-ADD saving and replacement H5 lemma are Sagar Thakrar's results.

The replay also establishes a coverage boundary: both native engines accept relabeling the 306 undo ADD phases. They do not use those labels. A separate complete record binding and phase invoice reject it; a count-preserving phase exchange is also rejected by the complete record binding. Missing undo, wrong scatter and late undo fail actual native checks; all 96 inherited omitted-bridge controls reject. The old owner checker continues to report H2/K/H5 false, so native success is not adoption of the replacement interface.

Validation:

• Registered CPython 3.12 replay: exact payload reconstruction, all 661 dirty columns forward/reverse, local invoice, owner flags, four semantic controls and three source/typed-receipt controls.
• Full native C++ replay under Python 3.12 with optimization enabled, GCC 15.2 and actual Boost 1.90 headers.
• Full native replay and exact receipt comparison from a Git-free read-only export under Python 3.14; all 31 exported files unchanged.
• Separate scoped native workflow added. Local execution is recorded; hosted CI is not claimed passed.

Predecessor: #2115 at b56dceffad83c88e8ad39ffd82a1bdd55c4130da. Its candidate/proof/README and expected records are pinned. The PR1188 literal interpreter/native harness are reused with attribution; the unchanged mathematical engines are not described as newly independent code. The new package is research/r439-deferred-native/.

This completes one explicit native handoff, not hosted emission, all-size compiler/physical scratch admission, a Lean proof, supplier adoption or a new kappa. PR2279/PR2356 address different scatter optimization questions. Canonical main and predecessor head were rechecked unchanged; final live overlap covered 578 open PRs and 1104 Discussion records with complete pagination and no warnings. Claim: #2362.

Prepared by Gareth Davidson (@bitplane), with substantial OpenAI Codex assistance. Apache-2.0; no subagents or independent human review. Full predecessor credit is retained.