Repository navigation
Releases
closure_drift 1.11.1
Two fixes (tests/PREREGISTRATION.md §38.1 and §38.2). Reports gain one optional field; no other verdict changed.
Fixed
- The GitHub Action annotates drift found inside a component, not only at the top level (§38.1, proof ACT08).
--at commitsin a shallow clone no longer reportsclean: the verdict isincomplete, reason code
shallow_history, withcoverage.history_shallowtrue and a hint to unshallow (§38.2, proofs SH01–SH03).
--at tagsis unaffected.
Documentation
docs/REPORT.md: wording fixes A1–A3 from the consumer study (§38,tools/consumers/RESULT.md).
Version DOI: 10.5281/zenodo.24305611
closure_drift 1.11.0
An independent verifier of the closure ids (tests/PREREGISTRATION.md §37 and amendments 1–3). Decisions, reports
and exit codes of the detector unchanged.
Added
closure-verify(closure_verify.py): checks a--json --evidencereport on its own. Structural level: the
fields agree with each other. Git level (--repo): every closure in drift recomputed from the Git objects, each
object checked against its name, replace refs, grafts and shallow history refused. Exit 0 verified, 1 the report
is contradicted, 2 cannot verify;reason: CODEon the last stderr line. It does not use the detector's code
and does not claim the repository is the one published. On the regression corpus: 29 drift reports, 225
closures, all verified.--evidence(with--json): addsclosure_commits, the commit where each closure in drift was first seen.
Off by default: a commit id changes with any file, and the report depends on the closure alone.docs/VERDICTS.md: the closure definition a verifier must reproduce.- Proofs: eight edits of a report or of its Git objects, each with its own code; eight mutants of the verifier.
Version DOI: https://doi.org/10.5281/zenodo.24267427
closure_drift 1.10.0
Reason codes (tests/PREREGISTRATION.md §36 and amendments 1–4). Decisions and exit codes unchanged.
Added
- Every refusal and every undetermined verdict carries a stable code:
reason_codein the JSON of an
undetermined verdict, and a last stderr linereason: CODE(a refusal with--jsonstill prints no JSON).
The table is indocs/VERDICTS.md. --tag: when no label is found and a workflow at the tagged commit writes the version from the tag, the
first line names the file and line.- Proofs RC01–RC05.
Not added
- A recipe for monorepos (§36 front 2). Measured on 168 candidate pairs against an independent reference, it
gave the right answer in 3 of 10 repositories; most of the rest set the version at release. Those 168 were
candidates, never shipped; this is not a false-positive rate of any release.
Version DOI: https://doi.org/10.5281/zenodo.24249739
closure_drift 1.9.0
Structural only (tests/PREREGISTRATION.md §35 and amendments 1–7). Same commands, reports, exit codes,
verdicts and refusals as 1.8.2.
Changed
measure()(319 lines) is split intoprepare(),scan(),assemble(),decide()andreport();
it is now 93 lines. The report stamp is built once, inmake_stamp().- Each step was checked against 1.8.2 by a differential harness (
tools/structural/differential/):
every detector call of the batteries run by both, 2,696 calls, zero differences, every call accounted
for. Git calls and run time unchanged on five corpus repositories.
Added
docs/VERDICTS.md: which function decides each verdict.- Proofs ST01–ST03: the text
incompletefor a contradicted source, and the label table cut after 25.
Version DOI: https://doi.org/10.5281/zenodo.24242196
closure_drift 1.8.2
From a second external check, of 1.8.1. The detector is unchanged.
Fixed
- 1.8.1's build provenance named
refs/heads/main, not the tag: the workflow was dispatched from
main.release-assets.ymlnow refuses unless it runs from the tag it builds.
Changed
- The release carries
DEPOSIT.sha256under the attestation and inSHA256SUMS, so the deposit's
manifest has an attested digest. DEPOSIT.sha256andSECURITY.mdsay the deposited copy differs only in the DOI line: compare the
nine hash lines.
Version DOI: https://doi.org/10.5281/zenodo.24221479
closure_drift 1.8.1
Release hardening after an external check of the 1.8.0 artifacts. The detector is unchanged.
Changed
release-assets.ymlpinsbuildandsetuptools, builds twice from the tag with
SOURCE_DATE_EPOCHand fails if the two differ.BUILD-ENVrecords the tools and is attested
with the wheel and sdist.tools/normalize_sdist.py: the sdist's tar and gzip headers no longer carry the builder's user,
checkout times or build time, so it rebuilds bit for bit (1.8.0's did not).- The Zenodo deposit carries
DEPOSIT.sha256and links to the release tag and the commit; the
script refuses a dirty tree.tools/verify_deposit.pychecks the deposited manifest.
Added
SECURITY.md: how to verify a release (hashes, attestation, rebuild).
Fixed
.claude-plugin/plugin.jsonstill said 1.5.1.
Version DOI: https://doi.org/10.5281/zenodo.24209557
closure_drift 1.8.0
Release integrity and usability. The Action splits args like a shell, so quoted globs reach the detector as written (in 1.7.0 and before the quotes were passed on and --closure 'src/**' could hide a drift). It runs the detector once and tells refused from error; new inputs closure, tags, version-file, version-regex pass each value as written. Every exit-2 answer now ends with a Next: line that says nothing was decided and why. --tag also answers incomplete when the release workflow takes the version from github.event.release.tag_name or git describe --tags. Exit 2 is not a pass. New: docs/START.md.
The wheel and sdist of this release were built from the tag by release-assets.yml, with build provenance (gh attestation verify <file> -R luizfnsilva/closure_drift). SHA256SUMS covers both.
272 proofs, 85 mutants, 315 adversarial cases, 17 properties, on Linux (Python 3.9–3.14), macOS and Windows (tests/RECORD.md). Regression corpus: 99 of 99 checkable match.
closure_drift 1.7.0
Version written from the tag at publish time. When the declared version differs from the tag and a workflow at the tagged commit writes the version from the tag ref (uv version "${GITHUB_REF_NAME#v}" and similar), --tag now answers incomplete (exit 2) instead of tag_mismatch, and names the file and line. This is not a pass. A placeholder value alone is not evidence. Only --tag changes.
252 proofs, 79 mutants, 315 adversarial cases, 17 properties, on Linux (Python 3.9–3.14), macOS and Windows (tests/RECORD.md). Regression corpus: 99 of 99 checkable match. SHA256SUMS covers the wheel and sdist, built from a clone of this tag.
closure_drift 1.6.0
Cargo workspaces. A member Cargo.toml with version.workspace = true (or version = { workspace = true }) now takes the version of its workspace: [workspace.package].version of the file package.workspace names, or of the nearest ancestor Cargo.toml with [workspace]. Before, it gave no label (exit 2). Nothing else in the detector changes; a refusal under --json still writes nothing to standard output, as documented.
247 proofs, 78 mutants, 315 adversarial cases, 17 properties, on Linux (Python 3.9–3.14), macOS and Windows (tests/RECORD.md). Regression corpus: 99 of 99 checkable match. SHA256SUMS covers the wheel and sdist, built from a clone of this tag.
closure_drift 1.5.1
Documentation and a demo. The detector differs from 1.5.0 only in its version string; verdicts and exit codes unchanged.
- Short README with a GIF of
--tag; everything it held is indocs/USAGE.md. RESULTS.md: two measurements by another user (bothdriftunder--at commits, where every commit is a publication).SCOPE.mdandROADMAP.mdcorrected where they contradicted 1.5.0.- CI: mutants in their own job; Python 3.9 on
ubuntu-24.04.
243 proofs, 77 mutants, 315 adversarial cases, 17 properties, on Linux (Python 3.9–3.14), macOS and Windows (tests/RECORD.md). SHA256SUMS attached. DOI: 10.5281/zenodo.24065769