Code-scanning audit: feat/v0.0.71#

The 2026-09-25 inventory contained 24 open alerts, all most recently reported against main at 708bf0607d8c21104fcaafae7285defddc06f28e. Alert numbers below refer to the code-scanning dashboard, not issues. Fixes stay on the feature branch; a valid finding on main is not dismissed simply because a feature-branch fix exists.

Branch fixes#

Alerts	Finding	Remediation
#189–#192	Built-in/GPG/registry dependency cycle	A private structural registration interface removes imports of the concrete registry from built-ins. GPG receives its owning registry instead of importing the global singleton. Public plugin protocols and API version are unchanged.
#237, #238	Mutable action references	Pin checkout and setup-python to the reviewed full commit hashes already used by the other workflows.
#239	Unhashed spelling tool installation	Generate a hash-locked codespell requirements file from its source input.
#219, #222, #223, #224	Local installs can resolve build dependencies despite no-deps	Hash-lock and preinstall build backends, build local wheels without isolation and with PIP_NO_INDEX=1, then install them with the hash-locked installer tool, which performs no dependency resolution. The fuzz job also installs hash-locked runtime dependencies.
#215	Generated SDK installation resolves arbitrary dependencies	Hash-lock the generated client’s runtime dependencies; build and install its reviewed local wheel without build isolation or dependency resolution, then run pip check.

Regression tests in tests/test_code_scanning_regressions.py enforce the workflow install policy and action references. GPG tests exercise distinct registries for whole-file and streaming dispatch. CodeQL now also runs on feature-branch pushes, using the existing query suite without exclusions.

The first fresh branch analysis also exposed a scheme-dispatch cycle (#245–#251 and #192), a redundant test import (#252), a type-only import the scanner did not recognize inside a quoted cast (#253), and unparenthesized test-script concatenation (#254). The follow-up separates pure scheme rules from dispatch, preserves public aliases and qualified class names, removes the redundant import, makes the CEL function type explicit in an annotation, and makes the intended string joining explicit. No findings are hidden with query exclusions; see ADR-0007.

Generated dependency locks live beside their .in inputs in .github/requirements/. Regenerate with Python 3.12 and pip-compile --allow-unsafe --generate-hashes --strip-extras --output-file=NAME.txt NAME.in. The base requirements.txt and .github/requirements/api.txt instead derive directly from pyproject.toml, the API file using --extra=api and --constraint=requirements.txt. This replaces the stale duplicate api.in input and the base lock’s former untracked temporary input. Runtime pins must satisfy the package metadata; every wheel installation now runs pip check. The allow-unsafe compiler option includes hashes for build tools such as setuptools; it does not disable pip’s installation-time hash verification.

False-positive review#

Alerts	Evidence
#244	`pain001.main` is a PEP 562 lazy export implemented by module `__getattr__`. A fresh-interpreter regression verifies every exported name resolves and CLI imports remain lazy. Eagerly importing main would regress browser and library startup.
#184–#188	Ellipses are intentionally empty method bodies in PEP 544 Protocol definitions, not forgotten runtime operations. Existing structural compatibility tests cover positive and negative implementations. The published contract stays unchanged.
#193	`_init_attempted` is read both before and inside the initialization lock. Existing tests prove cached initialization and explicit reset; removing it would break once-only initialization when tracing is disabled.
#183	`pytest.raises(RuntimeError)` catches the deliberately raised exception. The subsequent assertion is reached and verifies plaintext cleanup after a failing inner loader. The targeted test passes.

These eight findings were individually dismissed as false positives with the evidence above on 2026-09-25. No directory or query was excluded.

Findings that code on this branch cannot close#

#205 and #221: the SLSA reusable generator is deliberately referenced by v2.1.0. Upstream requires an exact version tag for trusted builder verification and rejects hash references. Retaining this dependency is an explicit maintainer-accepted exception (approved on 2026-09-25). Both alerts are dismissed as “won’t fix”, not false positives or an untested switch that breaks release provenance.
#194: branch protection is repository state, not branch content. The initial 2026-09-25 approval required one approving review and CODEOWNERS approval, with administrator enforcement. Later that day the maintainer explicitly approved a solo-maintainer policy: PRs remain required, approving-review count is zero, and CODEOWNER approval is no longer mandatory. Administrator enforcement remains enabled. Read-back verified that existing status checks, conversation resolution, force-push prohibition and deletion prohibition were preserved. Authors cannot approve their own changes; this is not independent review or a claim of maximum Scorecard protection: stricter review freshness and up-to-date requirements remain separate policy choices. The earlier Scorecard run raised the score from 3 to 8, but still requests two approvals and misreports administrator enforcement; the authenticated protection API confirms enforcement is enabled. The alert was subsequently dismissed with explicit maintainer risk acceptance; that earlier score does not describe the relaxed policy.
#199: historical changesets lack independent approving reviews. A new workflow, a self-review, or a bot dismissal cannot create that evidence. Improving this finding requires genuine independent approval; historical review cannot be retroactively fabricated. The solo-maintainer policy does not satisfy that security criterion. This alert was subsequently dismissed with explicit maintainer risk acceptance, not marked fixed.

All four accepted exceptions above have owner sebastienrousseau, a review deadline of 2026-10-25 and an expiry of 2026-12-24, following acceptance on 2026-09-25. The machine-readable registry is enforced in CI and release preflight. Renewal requires a new explicit maintainer decision; neither governance finding is represented as remediated.

No versions, release tags, public API signatures, generated payment XML or main-branch content are changed by this remediation.