[apoursh]

Summary

Adds scripts/check_pinned_edits.py, a standard-library static check for one class of PR intake failure: a branch rewrites bytes that other files pin by SHA-256. Existing guards (verify_selected_result.py, check_selected_ci_inputs.py) cover the selected record and its replay inputs. Neither covers the many research packages whose SOURCES.json, receipts or READMEs pin a file.

For each file the branch modifies or deletes, the script hashes the merge-base bytes and runs one git grep -F over the merge-base tree for the full digests. It then classifies each pinning file at the head:

• stale: the old digest is still present, so that check fails or describes bytes that no longer exist
• repinned: the pinner was rewritten, so byte checks pass, but frozen evidence may have changed
• pinner_removed: the pinner was deleted

Exit 1 only for a stale pin on a file under research/ or upstream/, the paths CONTRIBUTING/AGENTS declare frozen. Living records (selected record, task catalogue, docs, shared scripts) are routinely updated while older packages keep their digest as context, so their stale pins are advisory unless --strict. A short pointer is added to docs/research/pr-intake.md.

Evidence (macOS, Python 3.14.8, main 2485805a6)

Diff	Result	Exit
#6318 at c3fdf0b8	4 modified files, 42 stale pins on frozen files across 26 pinners, incl. certificates/selected-result.json source_certificate_sha256, certificates/frontier6030-source.json, pr6030-audit/{INPUTS,SOURCE-REUSE}.json	1
main~30..main	44 modified files, 0 stale frozen pins; 1 frozen README edited with its manifest → REPINNED; 1,281 advisory living-record pins	0
#6314, #6316, #6317 (additive packages)	0 modified files	0
this branch	1 modified doc, no pins	0

Independently of this script, main's verify_selected_result.check() against the #6318 tree fails with Selected package inventory or bytes changed, and passes on main. About 10 s on the full checkout.

tests/test_check_pinned_edits.py (6 tests, temporary git repos) covers: an unpinned edit, an in-place edit with three stale pinners, repinned/removed pinners, deletion beside a binary file, a living-record update (advisory vs --strict), and merge-base use after main moves. Passes under python3 -B and python3 -O -B.

Scope and limits

Static byte check of a single diff. It executes no contribution code and is not a replay, review or mathematical check. Only full lowercase 64-hex SHA-256 digests in text files are found. Abbreviated digests, other hashes and pins over decompressed or canonical payloads are missed, so an empty report does not prove frozen evidence is unchanged. The frozen-path rule is a prefix heuristic taken from repository policy. Not wired into CI; a maintainer could add it to the plan job if wanted.

Open next questions

• CI integration. Running it in research-packages.yml's plan job on pull_request would have flagged Retune P104 twin-rung-removal to the certified tuning ceiling (kappa 0.003125811556925) #6318, which has no CI runs. First step: decide whether advisory output belongs in the job summary.
• Digest formats. Extend the search to sha256: prefixed or uppercase forms if any package uses them. First step: git grep -c -E '[0-9A-F]{64}'.

Related: review of #6318 posted there. Predecessor guards: scripts/check_selected_ci_inputs.py (Sagar Thakrar), scripts/check_research_commit.py (ywz65).

Prepared with substantial AI assistance (Claude Code, Anthropic). Apache-2.0.

🤖 Generated with Claude Code