[SHIN09-git]

The selected PR4470 raw invoice has variable target and paid moment sums, so the older fixed-target screening model cannot be carried over unchanged. This package freezes the published invoice and tests certified alternating-Taylor bounds on their signed difference. It adds a sign-only adaptive screen and a rejection-only gate that retains the original full interval calculation for accepted candidates.

Seven fixtures agree with two other exponential enclosures, including the unchanged upstream arithmetic engine. In 252 fresh-process measurements, the rejection-only gate saves a median paired 32–35% on the two rejecting fixtures and costs about 11% more on the accepting fixture. These are three fixed test rates, not a candidate population or a complete replay speedup. The first roughly 10 seconds overlapped with a verification run; the timing file and README disclose this and uncontrolled host contention.

Scope and credit:

• No new kappa, supplier, root, native replay, uniform compiler, outer inequality or multiplication theorem is claimed.
• This is a migration probe using generic Taylor bounds, not an unchanged replay of Independent verification: quadratic screening brackets sharpen selected D/E diagnostics #1669. The fixed-target false-acceptance control tests an incorrect migration assumption, not a reported bug in Independent verification: quadratic screening brackets sharpen selected D/E diagnostics #1669.
• Retains the PR4470 invoice and zxv's unchanged higher-moment arithmetic engine, source hashes and Apache-2.0 attribution. Prepared for SHIN09-git with substantial OpenAI Codex assistance.
• The next question is an actual caller/candidate batch with the same complete invoice; the sign-only result cannot replace a downstream quantitative interval.

Validation on Windows 11 / CPython 3.12.10:

• Registered package replay passes: seven fixtures, eleven arithmetic/control groups and complete deterministic receipt equality.
• Source-only optimized replay passes; modified invoice, engine and expected receipt are rejected.
• Metadata/source pins and committed publication bytes pass.
• The official exporter and offline integrity check pass on commit 785ecde; a fresh extraction passes the same optimized source-only and tamper checks.
• The optional full-history preflight is unavailable in this shallow checkout. The declared source base is the publication commit's direct parent and its ancestry check passes.
• Final overlap refresh covers 305 open PRs, 4,090 Discussion records and the complete 6,690-note tracked inventory. Closed work is represented by local notes and the ledger; absence of overlap is not a novelty claim.

Reproduce from the repository root:

python -B scripts/research_package.py --package research/mixed-moment-screen

Or download only the directory and run python -I -O -B test_integrity.py. No Git, network, GPU or third-party dependency is needed for that direct replay.