eyetrajectoriespy 1.2 — draft release qualification audit (23 October 2026)¶

Decision: RELEASE HOLD. This is a source-, workflow- and governance-based audit, not an authorization to publish. All fixes here are confined to a draft pull request. Do not merge, tag, invoke the release publisher, publish to TestPyPI or PyPI, or claim that a passing numerical simulation establishes nominal inference coverage.

Source, evidence and publication identity¶

Frozen audit baseline: protected-main source at 6d9bc5396ebf596ea22f9187d87e44b9dd5c03a6. The main GitHub Actions runs inspected on 23 October comprise 24/24 successful current-head workflow runs (includes Linux/macOS/Windows matrix in the tests workflow; source packaging, docs, examples, release-readiness and optional/functional/sparse qualifications).

Important distinction: 24 passing workflow runs ≠ all possible 1.2 release gates. Other source-path or manual-dispatch workflows do not run on documentation-only main commits. In particular, the 1.2 installed W1–W4 product qualification workflow and the PyPI-installed-RC workflow have separate triggers and are not proved on a fresh literal RC2 by a passing documentation build. The installed-RC workflow in main is still pinned to older 1.1.0rc1 and cannot qualify a new 1.2 RC.

Existing workflow and compatibility coverage¶

• tests.yml: source checkout, editable .[dev] install, tests/coverage, compile and Ruff; Ubuntu, Windows and macOS × Python 3.11, 3.12 and 3.13. These tests protect code behavior but are not nine independent released-wheel installs.
• release-readiness.yml: wheel + sdist build, Twine check, isolated installed wheel and sdist smoke on Python 3.12, source version equality and release-hardening contracts. Passing is necessary, not sufficient.
• workflow-product-qualification.yml: module-level workflow contracts on Python 3.11–3.13, installed wheel W1–W4 ten-route product checks on 3.12. It runs when designated workflow/code paths change; not a fresh RC2 acceptance certificate.
• one-dot-zero-api-stability.yml and scripts/freeze_one_dot_zero_api.py: frozen root API signatures/schemas and exact export boundary. WORKFLOW_API_AUDIT.json requires 1.2 workflow candidates to stay module-scoped.
• installed-rc-qualification.yml: external installed 1.1.0rc1 PyPI reference; useful historical baseline but does not install 1.2 RC2.
• release.yml: manual dispatch defaults to build-only; production governance historically required GitHub readiness, exact protected main, an issue and required checks. Publication routes additionally have separate environment gates. Neither a green CI run nor this audit starts a release.

The source tree includes many independently audited draft W3 experiments that are not merged. A large research PR backlog must not be confused with an eligible source/API surface. Do not silently cherry-pick research estimators into a release.

Audit fixes prepared, not merged¶

1. scripts/verify_release_publication_gate.py checks every manually selected release target before building. The non-publishing build-only path remains possible while disarmed. TestPyPI needs a separate explicitly reviewed publication flag plus trusted publisher declaration; production and resume-production require both independent production and GitHub readiness decisions, environment reviewer/trusted publishing and exact-main governance declarations.
2. release.yml runs that target-policy gate before building any distribution. No publishing can occur from this draft by itself; the unchanged readiness flags remain false.
3. scripts/verify_release_governance.py additionally requires production_release_ready=true and treats only completed successful mandatory checks as valid. Neutral, skipped, queued or failed checks no longer satisfy a release requirement.
4. tests/test_release_publication_gate.py and release-governance fixture updates explicitly regression-test every disarmed target, wrongly skipped/neutral mandatory CI statuses, missing approval flags and retained RC1 evidence.
5. one-dot-two-release-qualification-audit.yml attempts a non-publishing Python 3.11–3.13 installed wheel and sdist matrix, importing the stable 1.1.0 PyPI distribution as a separate root-API reference. It checks candidate wheel imports outside the source checkout, retained 1.1 public root exports, unchanged module-scoped ten workflow entry points, isolated wheel/sdist installation plus pip check, and installed-wheel ten-workflow product output. The workflow does not publish or upload its wheel or source distribution; only qualification JSON reports are retained.

These tests are planned evidence at PR creation. Record their GitHub Actions results, failures and platform-specific gaps as actually observed; do not mark them successful before jobs complete.

Remaining scientific and production qualification gates¶

A future narrower 1.2 candidate is possible only if the release scope deliberately excludes all unqualified W3/MNAR confidence-band or missing-gaze correction claims, clearly keeps these research methods out of a supported public API, maintains stable 1.1 root compatibility and gets its own reviewed exact-version release qualification. A fresh literal RC, not current 1.2.0rc2.dev0, needs matching source/version, qualification JSON and new wheel/sdist/install evidence; the source-level W3 negative results and preregistered refusals must remain available.

For any method advertising statistical 90% simultaneous confidence coverage, issue #289 requires an independently designed, predeclared, adequately sized new-source coverage assessment, including high leverage, independent units, covariance/marginal/reference interactions, negative cases and Monte Carlo uncertainty. For any correction of response-dependent lost gaze, #296 requires externally grounded observation/recovery assumptions; random audit assignment does not guarantee actual audit return, and identified-set evidence is not a calibrated interval. Neither issue may be claimed resolved by a website deployment.

No qualified real GP3 human gaze dataset or reproducibly licensed real-sensor observation/missingness validation was established by this GitHub inspection. The documented synthetic evidence is valid as synthetic evidence only.

Exit criteria and decision¶

Before considering any further release PR: (a) finish new audit exact-head CI including installed artifacts; (b) resolve any actual failed contract without lowering thresholds; (c) review compatibility diffs; (d) document a narrow supported 1.2 release API or continue the hold; (e) qualify a fresh literal version; (f) have all exact-source CI and scientific declarations pass on protected main; and (g) require a separate, expressly requested governance-only release authorization. No merge, tag, TestPyPI or production PyPI action is included in this audit.