[aliromman]

This evidence-only PR retains 29 accepted review outputs that were absent from the original publication commits. Reviewers can now inspect the recorded rejection tests and final cold-run receipts at their stated repository paths. This changes no mathematical result or complete kappa calculation.

The exact file list, independent retention review, and structured review bind 47,576 original evidence bytes:

• Four encoder control logs record rejection of a false requirement and an outside-file read, in normal and optimized Python. They complete evidence from PR #4717 and its pinned review.
• The complete 21-file original-clean output set contains cold replay receipts, console output, and eight guard-control logs. It completes evidence from PR #4727 and its pinned review.
• Four distinct query receipts match hashes already recorded in the accepted replay summary. They complete evidence from PR #4720 and its pinned review.

The base is 624de9dc66858af0945e5f8a0942488d62280d43, after upstream merged those contributions. The addition preserves every existing file. The original fifteen publication heads remain fixed in the manifest. The separate byte audit verified the accepted source hashes, the expected control rejections, and the receipt-summary bindings. Other local review files either duplicate retained bytes, have their complete console text embedded in a retained receipt, belong to temporary inputs and earlier runs, or record later checks of PR descriptions. They remain preserved locally. This supplement does not rerun the mathematical suites or establish the parent packages' open caller, hosting, precision, or supplier obligations.

From the repository root, use Python 3.11 or later and the standard library. The following check was run on the committed supplement:

python3 - <<'PY'
import hashlib
import json
from pathlib import Path

root = Path('research/publication-reviews/centre-evidence-retention')
manifest = json.loads((root / 'SUPPLEMENT-FILES.json').read_text())
count = 0
size = 0
for item in manifest['files']:
    data = Path(item['destination']).read_bytes()
    if len(data) != item['bytes'] or hashlib.sha256(data).hexdigest() != item['sha256']:
        raise SystemExit('FAIL_RETAINED_REVIEW_OUTPUT: ' + item['destination'])
    count += 1
    size += len(data)
if count != 29 or size != 47576:
    raise SystemExit('FAIL_RETAINED_REVIEW_TOTALS')
print('PASS_RETAINED_REVIEW_OUTPUTS: 29 files, 47576 bytes')
PY

Expected output:

PASS_RETAINED_REVIEW_OUTPUTS: 29 files, 47576 bytes

The source packages retain their predecessor notices and Apache-2.0 terms: encoder notice, original-clean notice, and query notice. OpenAI Codex gave substantial assistance in the original reviews and this retention audit. The independent retention review is a separate agent check, not human peer review.