Research consolidation: PR946 frozen-pin inventory and restoration replay (no new kappa)

[woklerx]

PR946's proposed encoding changes alter bytes authenticated by historical packages. The batch952 review already identified eight PR950 conflicts; this research consolidation adds an exact wider inventory and executable evidence for preserving the historical source closure. No new kappa; no approval or application of PR946.

• Scans all 10,707 regular JSON files at base 407d52ff15b2217b6e22bce44728b4c3907b6515, with zero parse failures. The closed literal-record grammar finds 561 old-byte references across 99 files and 148 proposed paths, plus 330 proposed-byte and 42 other-digest records. These counts are provenance references, not claims of 561 runtime failures.
• Binds all 165 proposed paths to PR946 head b6e22dbf548d72f18b722ccb6c2ea8e3d9df6ac1, its actual parent 91a2094f2e022930069e8cfdc810a3ea1b81cd09, the retained maintainer inventory, and a direct Git diff/blob check.
• The unchanged PR950 verifier rejects each of its eight proposed inputs independently. Restoring seven scripts while retaining the proposed certificate still fails. Restoring all eight reproduces the complete 6,760,000-coefficient historical audit and original receipt ffae0f0a22863718e3134a924592dcf4b05e919223323ba536c587c592bfe6a1, with all 46 closure files unchanged. No expected digest is relaxed.
• Records five additional literal source references in the later accepted PR970 package. No new PR970 runtime replay is claimed.

Only research/pr946-pin-migration/ is added: verifier, reproducible inventory, eight exact proposed-byte fixtures, execution receipt, scope/migration guidance, local sprint claim and attribution. Production code, upstream/, selected certificates and frontier stay unchanged.

Validation (Windows, Python 3.14.5, standard library):

python -I -B research/pr946-pin-migration/check.py

Expected final marker: PASS_PR946_PIN_MIGRATION ... single_file_rejections=8 full_audits=2 coefficients_per_audit=6760000 kappa=unchanged. Also passed --check-proposal, six adversarial inventory tests normally and under -O, the selected-record checker, and staged fixture SHA256 checks. Full replay deliberately requires assertions. The separate agent review independently checked all JSON parsing and found no provenance bypass; it is not human peer review. No Linux CI or repository-wide make verify result is claimed.

Seed 95, research-consolidation track. No pinned coordination issue was present; the local claim is completed with its result link. Before submission, upstream refreshed to 8bddbee377e2f3ce1f1247196c4b290d55767d20; the proposal head, selected kappa, and PR950 input/adapter hashes were unchanged. Targeted overlap searches and latest open PRs through #986 found no matching migration contribution; this was not an exhaustive census.

Credits: Aaron Wang/navinor (#946), Ross Richardson (#950), Douglas R. Colkitt and the original verifier lineage, the batch952 maintainers for the original conflict finding, and predecessor reproducibility work named in NOTICE.md. Prepared for woklerx with substantial OpenAI Codex AI assistance. Apache-2.0. General dependency discovery, migration of every consumer, the broad encoding integration, and inherited multiplication-theorem obligations remain outside this contribution.