Research consolidation: bind CITATION/CONTRIBUTING/CONTRIBUTORS to the selected record (no new kappa)

[ianlibanio]

Research consolidation + tooling. No new κ. This PR does not change the certificate, construction, selected record or frontier.

Base: 9a16422bf8d0ebd3342c1156c33cb5703c6cb02a. Sprint track 95–99 (consolidation), seed 98.

Discrepancy

certificates/selected-result.json selects PR890 with PR589 at κ = 0.001661957774647029, with replay command make frontier890-verify. README.md and current-status.md agree with it. At base, three other public entry points did not:

• CITATION.cff, which GitHub's "Cite this repository" reads, still said "selected checkpoint is PR626 with PR543" and gave κ = 51110230636731/31250000000000000. That is two selections stale, and it credits the wrong bit supplier.
• CONTRIBUTING.md still pointed to community-frontier741-review.md and "currently make frontier741-verify".
• CONTRIBUTORS.md headline still said "Current reviewed composition: Conditional κ = 0.001661634653669567: four paired read exchanges on PR741's frame-coalesced word #799 with Bit c = 0.001733427039589844: two A-section transports on #555's punctured R439 word (+0.0216%, κ unchanged) #589".

This keeps happening

research/selected-entrypoint-drift/history.py runs the same unchanged binder against all 17 commits that changed the selected record, plus the base. 12 of the 17 selection commits left at least one restatement that contradicted the record committed with it:

File	Number of selection commits where it contradicted the record
CONTRIBUTING.md	9
CITATION.cff	7
CONTRIBUTORS.md	1
current-status.md	1
README.md	0

I checked every contradiction against the text at that commit. Cases where the old wording had no anchor phrase (missing) are not counted as agreement or as drift. One false positive, README's older \times10^{-4} notation, was found during review. The parser now handles it and a test covers it.

Change

• CITATION.cff, CONTRIBUTING.md and the CONTRIBUTORS.md headline now match the selected record, README and the Conditional κ = 0.001661957774647029: T21 whole-owner packing and fixed-family bound #890 review. The existing cumulative contributor table is now labelled "SW4 and X4 rounds through Conditional κ = 0.001661634653669567: four paired read exchanges on PR741's frame-coalesced word #799". No credit text was removed.
• scripts/verify_selected_entrypoints.py checks 18 fields across these files against the record: exact κ and decimal, the primary PR, review and receipt links, and the replay command. It also checks that reproduce is a real Makefile target and that previous_kappa matches its checkpoint. A missing anchor fails, and decimals are compared exactly. It complements scripts/verify_selected_result.py.
• tests/test_selected_entrypoints.py adds 7 tests with negative controls. Each stale base value is rejected on exactly the expected fields. Rounded decimals, missing anchors, an undefined target, a wrong previous κ and an underivable primary PR are also rejected. make verify-tests picks the tests up automatically. No Makefile or workflow edits.

Reproduce

python3 -I -B scripts/verify_selected_entrypoints.py
# PASS 18 selected-result restatements agree with certificates/selected-result.json: kappa 1661957774647029/1000000000000000000

On a checkout of the base commit, the same command (--root <base checkout>) exits 1 with FAIL 8 of 18. Further commands:

python3 -m unittest discover -s tests -p test_selected_entrypoints.py -v   # 7 tests OK
python3 -I -B research/selected-entrypoint-drift/history.py --check        # needs a non-shallow clone

Checks run locally (Fedora, Python 3.14; the new checks also passed on 3.12):

• make selected-record-check passes, and so do the existing test_selected_result.py tests.
• make verify-tests: 93 of 94 modules pass, including the new one. The exception is test_cpp_headers.py, which needs a c++ compiler that this machine lacks. It fails identically on the untouched base. I did not run frontier replays, because no construction input changed.

Scope and open questions

• This is a metadata consistency check only. It does not replay or certify κ, and it does not decide what should be selected; the record is authoritative. Prose credit lists are not mechanically checked.
• The CITATION version/date-released follow the existing community-<audit>-<date> pattern. The maintainer should confirm them, because they cannot be derived from the record.
• The binder is not wired into selected-record-check, so frontier replays don't depend on documentation text. The maintainer may choose to wire it in.
• Noted and left unchanged: make verify still invokes frontier741-verify, not the selected frontier890-verify. That may be intentional for cost. Ledger coverage of the selected source PRs is navinor's open Research consolidation: selected-source provenance gap map + 298-row vendored-source re-verification (no new kappa) #867 and is not duplicated here.

Credit

This builds on the maintainer's verify_selected_result.py and selection records. It is related to #841 (Mihir Srivastava, digest-bound proof pointers) and #867 (navinor, ledger coverage map).

Prepared by Ian Libânio with substantial AI assistance (Claude Opus 5.5, Anthropic, via Claude Code). Apache-2.0. Not independent human review.

🤖 Generated with Claude Code