changelog-gate: package-root testing-conventions.toml demands a fragment it can never warrant #974

What

internals/checks/src/checks/changelog_gate/decide.py::_is_exempt treats every non-test file under packages/<pkg>/ as source. Since #933 both SDK packages carry a gate config at their root:

• packages/python/testing-conventions.toml
• packages/ts/testing-conventions.toml

Neither is exempt, so any edit to them — narrowing an [e2e] scope, nudging a coverage floor, adding an exemption — fails changelog-gate unless the author remembers a skip-changelog: line. #971 needed exactly that on both of its commits.

Why these are not source

The distinguishing test the maintainer already applied to READMEs in #932 item 4 is does it ship in the published artifact:

READMEs ship in all three published artifacts (PyPI long-description, npm tarball, crate), which is why release-ci.yml deliberately does not exclude them while it does exclude fragment dirs.

Gate config fails that test in both packages. The maturin wheel carries only the dirsql package, and packages/ts/.npmignore keeps the tarball to the built output — a testing-conventions.toml at the package root reaches no user, so no user-facing change can ever follow from editing one. That puts it with changelog.d/ and e2e-attestations/, both already exempt, rather than with READMEs.

The escape hatch works, but it is the wrong instrument here: skip-changelog: is meant for the occasional no-observable-change PR, and this makes it a standing requirement for a file class that is never changelog-worthy. A forgotten trailer produces a red gate whose fix is a commit-message amend, which is exactly the friction the gate's actionable-error rule exists to avoid.

Proposed change

Add one clause to _is_exempt, alongside the fragment-dir and receipt-dir entries:

or re.fullmatch(rf"packages/{p}/testing-conventions\.toml", path)

Package-root only — a hypothetical testing-conventions.toml nested deeper under a package would not match, and should not.

Worth deciding at the same time whether the sibling repo-tooling configs deserve the same treatment for consistency; they sit outside packages/, so the gate never sees them today and nothing changes either way.

Work involved

Behavior change, so normal red/green: a red colocated test in decide_test.py asserting the path is exempt, pushed and observed red in CI, then the one-line implementation. unit mutation is diff-scoped and will mutate the new clause, so the test needs to assert both directions (the package-root path exempt, a nested or differently-named path not).

Notes

Surfaced in #971 (#933), which introduced the two files. Related but distinct: #944 covers the same files matching putitoutthere's publish globs, which is a release-cost question rather than a gate question — the two want separate fixes in separate places.