R439 helper-star: exact dirty-cut replacement contract audit

[Arham-Tahir64]

DONE — PR1580, head 2490dac98eacc294bac29f136f0715343611d55e.

Exact finite obstruction: no nonempty subset of the final helper-star restorations in main's PR967 221-edge word, PR1062 221-edge composition, or PR1062 236-edge candidate can move unchanged to the end of fixed C. The target defect is K(S-I)h; distinct leaf columns prevent cancellations and every actual center column is nonzero. The written argument covers all 2^221-1 / 2^236-1 subsets, without pretending to enumerate them. Every dirty leaf witness changes at least eight target outputs although all sources/helpers are eventually restored.

Local CPython 3.12.14 / Windows x86_64 registered replay passed: all 678 singletons, all 661 formal columns forward and reverse (1356 full-matrix replays), a separate direct Boolean witness for each, and all-edge replays. The unchanged complete local ADD/frame bill is checked. Six focused tests pass, including five on-disk corruption controls and isolated optimized minimal export with every input byte preserved. A nontrivial kernel positive fixture accepts deferred restoration on all 16 dirty inputs. Clean-only and two-leaf all-ones fixtures exhibit false acceptance.

This adds a fixed complete-word boundary census to woklerx's PR1017 D-boundary omission atlas; it finds no defect in that atlas. The general reader/stabilizer principle is credited to mightyOil/Philipp PR1203 and Sagar Thakrar PR1199. PR1194/1311's restored semantic-D argument is not repeated. Full PR967/1062/915/926/1015/923 and DreamingOfClouds/eumemic lineage credit, exact pins, scope change, commands and open question are in the package.

Research base c5108f0f68d482e5712a8e26a5978d47762dfc16; rebased on ddb4371c16a5a5de9e363afa8c0d6d7c27429e3c with all seven pinned blobs and guidance unchanged. Final complete overlap refresh at 03:15 UTC covered 728 open PRs and 225 Discussion records; cited pending heads were unchanged. Body/README/ledger search limits remain explicit; no novelty certificate is claimed.

No new kappa, supplier adoption or selected-frontier change. Native multistage/whole-hosted emission, physical compiler/common-label, precision/tape, analytic and Lean obligations remain unrun. The concrete next question changes the excluded hypothesis: derive and fully price one edge's conjugated restoration and target repair with legal frames. Current target/helper final frames and lack of paid-profile headroom are blockers, not asserted opportunities.

Prepared by Arham Tahir with substantial OpenAI Codex assistance; one agent, no subagents or independent human review. Apache-2.0. The advisory claim is complete.

[Arham-Tahir64]

DONE — PR1592, head 20588690897133e9a811f9ba434b0cce8d156d04. The claim is complete.

The post-C handoff from PR1580 is settled under the existing frame contract: no arbitrary appended elementary-XOR compensation can repair any nonempty omitted-restoration subset on the pinned PR967 or either PR1062 word, even with arbitrary finite restored dirty extra wires. All original sources/helpers are already full rank eight, and each output target has a distinct rank-seven final hyperplane. The exact original-input frame cone is {the target itself}; a monotone-frame dependency proof blocks the required full-frame leaf coordinate, including via extra-wire relays. PR1580 owns the defect/subset argument; Ali Romman/PR1128 and Cinna Davis/PR1393 own related general boundary reasoning. This is its new terminal-target application, not a claim to a new general boundary lemma.

The frame-free one-star terminal repair has exact minimum w+1 XORs for one leaf and w+3 for two leaves, where w>=8 is the number of affected targets. Keeping the original early restoration is cheaper by w or w+1. The scalar witnesses deliberately violate the actual helper-to-target final-frame constraint and do not provide a legal supplier or paid-profile gain.

Evidence: 333 exact rational target-cone checks; all 880 within-star repair cases on all 661 dirty columns forward and reverse; separate Boolean witnesses and dot-product frame oracle; exhaustive finite legal-suffix fixture with a dirty extra, short-word exclusion, dense scalar fixtures and real source/receipt corruption controls. Registered CPython 3.12.14 replay and all five package tests pass; optimized read-only export and separate Python 3.14.7 replay pass; 20 repository research-workflow tests run successfully with one skip. Final overlap at 03:36 UTC covered 727 open PRs and 233 Discussion records without warnings. Relevant predecessor heads were unchanged; scope and search limits are explicit.

Reproduce: python3.12 -I -B scripts/research_package.py --package research/r439-terminal-repair-barrier.

HANDOFF: a mechanism must change the continuation before an affected target reaches its final hyperplane, or supply a different paid transport/frame interface. Derive one exact earlier-cut conjugation and price all carrier climbs before constructing a word. Arbitrarily enlarging a terminal XOR suffix cannot help in this model. No native/whole-hosted replay, changed precision/tape, all-size realization, analytic transfer, Lean proof, selected supplier or new kappa is claimed. Hosted CI and maintainer review remain separate.

Credit PR1580/1017, PR1128/1393 and related complex PR1517; PR1194/1311; PR967/1062/915/926/1015/923; DreamingOfClouds/eumemic PR589/555/543 and PR425 lineage. Prepared for Arham Tahir with substantial OpenAI Codex assistance; one agent, no subagents or independent human review. Apache-2.0.