0.4.0 publication repair

The implementation merge and successful dry runs did not publish 0.4.0. At the start of this repair, PyPI returned 0.3.5 as latest and 404 for 0.4.0; GitHub had only the historical v0.3.5 release, without package assets. That release's install command incorrectly named the unrelated aire package. The July failed tag job also showed a 403 uploading that wrong distribution; a subsequent July run had successfully published aire-ai 0.3.5.

The current release workflow was missing GitHub Release creation. Repository metadata also retained README 0.3.x, citation 0.3.5 and an outdated citation date, an unreleased migration heading and a 0.3.5 bug-report example. These are corrected before tagging. The old GitHub release's install command and repository homepage were corrected without moving old tags or replacing published package files.

Required publication contract

• Version, citation/date, changelog, README series, docs and issue-template example agree. Published distribution metadata is aire-ai at the same version.
• The matching tag points to current clean main. CI, CodeQL, Docs and Benchmarks must have successful main-push runs for that exact commit; check again just before uploading. Environment release remains restricted to v* tags.
• Build tools and Hatchling come from the lock without build isolation, with the commit timestamp as SOURCE_DATE_EPOCH. Two consecutive local builds produced byte-identical wheel and sdist.
• A bundle contains exactly the wheel/sdist, source SHA, SHA-256 checksums and integrity-checked changelog-derived notes. A separate job verifies the downloaded bundle and installs the wheel outside the checkout.
• Existing PyPI files must match these hashes. Matching files permit resumable publication; a conflicting file fails. Verify both published downloads and a fresh published-wheel installation before creating the GitHub Release.
• GitHub assets include the same wheel/sdist, checksums and source provenance. Existing drafts can resume; public assets must match on a repeated run, and verified notes/title are reapplied so stale wrong-package instructions cannot survive a green rerun.
• Dry runs explicitly skip publication. Main is merged and all checks plus its dry run pass before the version tag is pushed. The actual publishing job, public indexes, release assets and a fresh index install establish completion.

Verification before release

The full local suite passes 331 tests, with 11 live tests deselected; the 23 new release regressions cover stale source/tag/checks, distribution metadata, artifact/notes integrity, partial/matching/conflicting published files and published-download host/hash validation. make all passes in a separate locked base environment, including strict docs and release metadata checks. Ruff, formatting, strict types, lock validation, actionlint, twine and isolated installed wheel QA pass. Existing optional/service coverage is rerun by the full GitHub matrix before merge and on main. Publication evidence is recorded by the tag workflow, GitHub release assets and PyPI rather than predicted in this document.

Independent review disposition

A fresh reviewer independently ran the initial 21 release tests and reproduced two Important defects. Both received failing regressions before repair: isolated builds resolved an unlocked backend, and an existing public GitHub Release could keep incorrect notes despite matching assets. Hatchling 1.32.4 and dependencies are now locked in the release group; canonical builds use --no-isolation. The exact GitHub shell step now refreshes verified title/notes after public asset comparison. There were no Critical or deferred Minor findings.

The review's external/unsupported cases were evaluated as follows:

Case set aside by reviewer	Disposition / cost if unmet
Live OIDC, PyPI and GitHub publication	Mandatory actual tag/public-index/install verification after all preflight gates; no completion claim from dry runs
Environment/ruleset configuration	Verified release environment permits v* tags; preserve it and existing publisher identity; configuration drift can block publication
Forced tag movement/deletion during a run	Unsupported and prohibited; operator mutation can invalidate or stop a release
Resuming after main advances	Intentionally rejected to honor current-main requirement; a later change requires fresh QA/version decisions
Public releases with missing/conflicting assets	Fail without replacing public assets; maintainer repair may be needed
Unchanged runtime/optional/service behavior	Full local and hosted matrix verifies existing contracts; GPU/provider limits remain explicit
Universal long-term/cross-platform reproducibility	Evidence is scoped to the locked release profile and repeated builds; other toolchains/platforms need separate byte checks

Actual 0.4.0 publication and 0.4.1 follow-up

Tag run 38681349559 uploaded 0.4.0 successfully, then failed before verification because the publisher created *.publish.attestation files in the original distribution folder. The strict two-distribution inventory correctly rejected the changed folder. Both published package hashes matched the pre-tag dry-run bundle exactly. Rerunning the failed job reused those matching files, skipped upload, verified published bytes and a fresh installation, and created the complete GitHub v0.4.0 release. No tag was moved and no PyPI artifact was replaced.

Version 0.4.1 fixes the initial-publication path by copying verified distributions to a separate release-bundle/upload directory. Attestations remain enabled; strict payload/hash validation and GitHub asset selection still use the untouched dist directory. A regression executes the workflow staging step, simulates the publisher's exact sidecar filenames and verifies the original bundle afterward; it failed with the original live error before this repair. Repository version, citation, docs, issue template and changelog are updated before the patch tag. The version-drift test now follows the actual source version across patch bumps.

Patch QA: 332 local tests pass (24 release regressions), with 11 live tests selected separately. The rebuilt 0.4.1 wheel passes twine and isolated installed package checks; locked make all, strict types/docs/lint, actionlint and lock validation pass. Hosted matrix and exact-main dry run are required before tagging.