ci(release): key release notes on the real type labels, not just dependencies

[JarryShaw]

Please follow the guide below

• You will be asked some questions, please read them carefully and answer honestly

• Put an x into all the boxes [ ] relevant to your pull request (like that [x])

• Use Preview tab to see how your pull request will actually look like

• Searched for similar pull requests

• Followed the coding style (make pylint, make mypy, make isort) — N/A, no Python touched

• make test passes, and a test case covers the change

• Added a changelog entry under docs/source/changelog/ and regenerated CHANGELOG.md, if the change is user-visible — N/A — changelog centralised in docs(changelog): shared 1.5.0 changelog — long-lived, merges last (#610, #616, #617, #618, #620) #657

What is the purpose of your pull request?

• fix — corrects a defect
• feat — adds a feature
• perf — changes performance, not behaviour
• refactor — changes neither behaviour nor performance
• test — tests only
• docs — documentation only
• ci — workflows or build tooling
• chore — anything else

---

Description of your pull request and other information

Replaces the two-category .github/release.yml (Dependencies / catch-all) with one category
per type label, now that all 499 merged PRs carry a reliable type label. Order is derived from
a 330-PR survey of label overlaps: breaking first so it never gets buried under fix/release/
ci; release second so old "New distribution" PRs land there and not under ci/fix/
dependencies; dependencies moved from first to second-to-last (judgement call, reasoning in
the file's own comments — flip it if wrong). review:* labels documented as never eligible for
a category or an exclude.labels entry, since either would drop reviewed PRs from the notes.

Pure YAML config, no code path — no coverage delta, no test to add. Validated: YAML parses,
schema checked against GitHub's documented release.yml keys (category title+labels only,
no invented keys), and label combinations traced by hand against the new ordering (see PR
description / session report for the traces, e.g. #691 breaking+fix → Breaking, #21 release+ci
→ Release). actionlint was not available in this environment to also check it.

[JarryShaw]

✅ GOOD TO MERGE @ c1517a8aa — no exclude.labels risk exists; ordering, schema, and labels all check out.

[JarryShaw]

✅ GOOD TO MERGE @ c1517a8aa — no exclude.labels risk exists; ordering, schema, and labels all check out.

Claim	Verdict	Evidence I obtained
review:*/wip/blocked in any exclude.labels	Not present	grep -n exclude on the PR-head file → 2 hits, both in prose (lines 77, 79); Ruby YAML.load_file parse shows zero exclude key anywhere in the structure
YAML valid / schema correct	Valid	Parsed PR-head file: 13 categories, each only {title, labels}, no invented keys, * last
perf+ci PRs (#731, #725) land right	Correct	Both carry perf+ci (gh pr view --json labels); order has ci (pos 6) ahead of perf (pos 7) → both file under "CI & Build" — same rule as the documented #727 precedent, and both titles are ci:-prefixed, so that's the right call
All 12 category labels exist	Confirmed	gh label list: breaking, release, feat, fix, refactor, ci, perf, docs, chore, test, const, dependencies all present — refactor/release are not missing, contra the brief
Existing label with no category and no * fallback	None	* sits last, catches everything unmatched
Stale "only Dependabot labels anything" prose replaced	Mostly accurate	Spot-checked 15 PRs the comment cites (#3,4,6,9,11,12,14,15,21,22,25,41,340,637,691,697,703,727) against real labels — 15/15 label sets match; the ordering conclusions drawn from them hold
"499/499 merged PRs carry a type label"	Number wrong, substance true	Actual merged count = 334 (search API and gh pr list agree); all-state PR count = 505 — neither is 499. Independently verified 0 merged PRs (and 0 of all 505) lack a type label, so the underlying claim is true, just mislabeled
"refactor's only overlap is with breaking (1x)"	Undercounted	Actual breaking+refactor = #3, #4, #6 (3, not 1); #4 also carries feat. Doesn't change output — breaking and feat both already outrank refactor regardless of the count
breaking+fix "38x+5" (43)	Minor drift	Measured 44 now; consistent with a PR or two merging since the 330-PR sample was taken

Nothing found changes where any PR actually files, and the flagged risk (review labels in exclude.labels) does not exist in this file at either scope. The numeric slips above are comment-only and cosmetic.

✅ GOOD TO MERGE @ c1517a8aa — no exclude.labels risk exists; ordering, schema, and labels all check out.

[JarryShaw]

Re-measured the comment's numbers directly (gh pr list --state merged --limit 1000 + gh search prs, cross-checked per-PR via gh pr view):

• Merged PRs: 334 (not 499). Zero lack a type label — reworded to "zero exceptions" instead of a hardcoded N/N that goes stale on the next merge.
• breaking+fix: 44 (39 without test, +5 with) — was 38+5=43, drifted since the original sample.
• breaking+refactor: still 1× (Regular update #6) among merged PRs — the original text was correct. The review's "3×, test/rc/abc #3/Regular update [test/rc/abc] #4/Regular update #6" counted test/rc/abc #3 and Regular update [test/rc/abc] #4, which are CLOSED (never merged), outside the comment's own stated "sampling merged pull requests" scope.
• Also refreshed drifted totals in the frequency table (fix 123→124, ci 39→41, perf 4→7) and the multi-label-overlap count (67→69) while in there.

Config itself (12 categories, order, no exclude.labels) untouched — diff is comment-only. New head: ecf57f9f8. Flipped label to review: pending since head moved past c1517a8aa.

[JarryShaw]

✅ GOOD TO MERGE @ ecf57f9f8 — comment-only delta, every corrected figure independently re-derived and correct, still no exclude key at any scope.

[JarryShaw]

✅ GOOD TO MERGE @ ecf57f9f8 — comment-only delta, every corrected figure independently re-derived and correct, still no exclude key at any scope.

Narrow re-check; supersedes the c1517a8aa verdict. Diff c1517a8aa..ecf57f9f8 is 8+/8−, every line a # line in the L14–57 block — the non-comment lines are sha256-identical (c6ce0db4…), so changelog, categories, titles, labels and order are untouched; the PR still changes only .github/release.yml.

New head parses: 13 categories, each exactly {title, labels} with one label, * last and nowhere else, all 12 keyed labels present in the repo, and no exclude key at any scope — the only two hits (L77, L79) are prose describing what not to do.

claim in the comment	measured independently
sample 334 merged, 0 untyped ("zero exceptions")	334 merged, 0 without a type label	✅
breaking+refactor = 1×	#6 only — #3/#4 are CLOSED, never merged	✅
breaking+fix 39× (+5 with test)	44 total = 39 + 5	✅
fix / ci / perf = 124 / 41 / 7	124 / 41 / 7	✅
more than one type label = 69	69	✅

Commit message now agrees with the file (zero exceptions, 334-PR, 13 categories); the stale 499/499 and 330 are gone.

Two caveats, neither blocking: L43 cites #727 for ci+perf, but three merged PRs carry both (#725, #727, #731) — the sentence claims an example rather than a count, and all three are genuinely CI-first, so the ordering ruling holds. And release's "ci (3×)" is a residual count — 7 raw, less #25/#26/#28 already enumerated under breaking — which is correct, and is exactly the scope discipline the earlier refactor correction lacked.

No test run, deliberately: a comment-only change to a config file with no import path, so the coverage rule genuinely does not apply here rather than being quietly skipped.

✅ GOOD TO MERGE @ ecf57f9f8 — comment-only delta, every corrected figure independently re-derived and correct, still no exclude key at any scope.