[1337isnot1337]

The fixed-scalar screens leave changing the scalar dependency graph open. This package checks corrective ADD expansions in Cao Yifan's frozen round-three native word, including nonadjacent pairs at raw-event distance at most 16 with fixed intervening contacts. It establishes no higher κ.

In each orientation, 381 adjacent and 1,512 nonadjacent pairs produce 5,679 corrective orders. The exact binary flag reduction covers every feasible intermediate subspace: 4,621 dimension triples, of which 4,400 preserve the complete paid histogram and 221 are strictly worse for every power price r^t, 0<t<1. Another 1,281 orders are infeasible. Integer truncated-rank inequalities prove the comparison for every concave rank tariff with f(0)=0. Every affected transition, including the bypass and final exits, is charged; all 408 productive boundaries and eight center jobs remain fixed.

This is a scoped source-specific result, not a whole-word optimum. Native-map changes, helper-boundary changes, interacting rewrites and retiming the intervening contacts remain outside it. Task centre-path-sharing is explicitly refined to this supporting scalar-graph screen; the original sharing construction gate is not claimed. No selected source, frozen predecessor, shared code or selected frontier changes.

Credit CubeJerry's PR589 corrective/flag/concavity screen, Ian Libânio/PR933, Dugongue's Steinberg and transported-frame work, Cao Yifan's round-three native word, Cinna Davis/#4470, zxv, Jacob Alan Sussman and the retained source lineage. Substantial OpenAI Codex assistance; no subagents, independent human review or proof-kernel acceptance claimed.

Validation at contribution head 5f3e088621e0a510f5b07ece590409e94ae45897, CPython 3.12.3:

• Registered normal and optimized warning-strict replays pass, including complete source invoices, 108 rational identity fixtures, all 355 nested three-bit boundary flags and seven failure controls. An actual improving flag is detected as a positive control.
• Committed-byte preflight, all source Git-blob/SHA256 pins at the immediate base parent, and direct base ancestry pass. The tool's full-source-history mode was unavailable because this checkout is shallow; no full-history pass is claimed.
• Official committed export and inventory checks pass. Git-free normal/optimized replay from the read-only export passes, with unchanged bytes and read-only modes.
• git diff --check passes. No Lean, optional solver, historical full supplier or outer theorem replay ran. Both hosted checks pass at the exact contribution head: planning and the registered package replay, run 39947855264. Run metadata confirms that head and success.

Reproduce with python3 -B scripts/research_package.py --package research/native-scalar-chain-screen.

Canonical main was refreshed before submission at a5df48c55ea806d7952b65acc9c76d878ff09b13, priority revision 7, with unchanged selected κ 0.00302015. The refreshed overlap report covered 331 open PRs, 4,340 discussion records and the complete 6,918-note HEAD inventory; it found the credited BIT predecessor and this local package for the specific corrective-chain queries. This is metadata coverage, not a novelty certificate. README and TASK retain the next-question gate; increasing this window without a new cost lever is not proposed.