Research consolidation: reconcile PR279 target-prefix sources and replay

[mihir-s-05]

PR279's reviewed checkpoint is preserved on main, but the ledger identifies twelve uncovered paths in its later target-prefix head. This research consolidation preserves those exact bytes, reconstructs the complete later package, and adds a fresh nine-stage finite replay. It resolves a concrete source/evidence gap; maintainer acceptance of the later head remains pending.

• Reviewed head: 4d8c31eec7ae618645cb058101cb4e72a97268d3.
• Later head: 913ccc48a0bc1173709ea92280b96252256b7754.
• Initial research base: 66849e1475c930fd7732b1a0067ffee7e1e95f60.
• Final integration base: 7ee02b7c96d85f22b38d28bd7ed49bb8c1aeb261; all consumed PR279 source and immutable context bytes are unchanged across the bases.

The companion binds all 115 package files and the unchanged workflow to exact Git blobs and SHA-256 hashes. It retains the 12-file update separately from the original reviewed source and receipts. The source/evidence checker rejects stale-head evidence, missing stages, changed inputs, unpaid restoration and ambiguous archives. Regenerated literal-word equality covers all six instruction fields independently of gzip container bytes. Only the PR279 ledger entry receives a companion pointer; its disposition stays changed_since_review.

The complete fresh candidate replay passed locally in 1,096.765 seconds on CPython 3.12.0 / SymPy 1.14.0. It checks 19,930 formal columns, 27,349 exact determinant bases, 4,923,000 bank assignments, 370 charts, both moment engines and finite assembly, including the omitted-setup control, 11 scalar controls, six prime controls and all nine stage-omission controls. The preserved historical exponent is 724098741932703/1000000000000000000. This is scoped finite evidence from the existing candidate verifier, not an independently implemented mathematical verifier or a formal proof of the full theorem. Inherited interfaces are stated in the README.

From the repository root, with Python 3.11+ and sympy==1.14.0, run this command using a new output directory:

python -X utf8 -B research/pr279-target-prefix-reconciliation/verify.py --replay --work ../pr279-replay-new

Expected final line:

PASS_PR279_FRESH_REPLAY stages=9 kappa=724098741932703/1000000000000000000

Final validation: source/evidence checks and 13 companion tests pass normally and under -I -O; all 14 reconciliation unit tests pass; the repository reconciliation checker validates 535 entries; the selected-result checker passes for #741 with #589. The executed, extracted and final reconstructed 115-file source packages are byte-identical, and all 723,357 literal instruction sextuples match the retained fingerprint. No new selected-frontier replay, full make verify, Lean build or new remote CI is claimed; existing source CI is separately pinned.

Track: research consolidation, OS-random assignment seed 95. Three candidate tasks and the selection rule are recorded in CLAIM.md; pre-submission overlap metadata is in OVERLAP.json. No pinned sprint issue was available, so the claim was kept locally. No overlapping PR279 reconciliation was found.

Prepared for Mihir Srivastava (mihir-s-05) with substantial OpenAI Codex assistance. Rohan Arun's PR279, eumemic's PR268 target transform, DreamingOfClouds' gen4 source and the full supplier/reviewer lineage retain credit and source notices. Apache-2.0. The selected frontier and upstream/ are unchanged.