Release notes: publish the changelog entry, not the commit subjects (#273)

Cutting 1.9.34 produced a release page reading:

    ### Changes
    - chore(release): v1.9.34
    - Changelog: one 1.9.34 entry, and credit where it was missing

while the entry describing single sign-on, the removal of the last native image
dependency, three update failures and every outside contributor sat in
CHANGELOG.md and was never published. The notes on the release page are what most
people actually read; they should be the written ones.

The workflow now takes the section for the version being released and uses it as
the body. Commit subjects remain the fallback for a version with no entry, so a
release never publishes with no notes at all — scripts/bump-version.sh already
warns about a missing heading, and this is the same gap showing up downstream.

awk rather than sed for the extraction: the body contains regex metacharacters and
markdown that a sed range would mangle.

Verified against the real file: 1.9.34 extracts 269 lines and stops at the next
heading with all 13 contributor credits intact, 1.9.33 and 1.9.29 extract cleanly,
and a version with no entry yields nothing and takes the fallback.

v1.9.34's notes were corrected by hand after release; this is so the next one does
not need that.

.github/workflows/release.yml

@@ -100,15 +100,42 @@ jobs:
       - name: Generate release notes
         run: |
           PREV="${{ steps.ver.outputs.prev }}"
+          VERSION="${{ steps.ver.outputs.version }}"
+
+          # Prefer the hand-written CHANGELOG section for this version.
+          #
+          # The generated list is commit SUBJECTS, which describe the work, not the release: cutting
+          # 1.9.34 produced a page reading "chore(release): v1.9.34" and one changelog commit, while
+          # the entry describing single sign-on, the native-dependency removal, the update fixes and
+          # every outside contributor sat in CHANGELOG.md and was never published. The notes on the
+          # release page are what most people actually read, so they should be the written ones.
+          #
+          # awk rather than sed: the body contains regex metacharacters and markdown that a sed range
+          # would mangle. This takes everything between `## <version>` and the next `## ` heading.
+          CHANGELOG_BODY="$(awk -v v="## $VERSION" '
+            $0 == v {found=1; next}
+            found && /^## / {exit}
+            found {print}
+          ' CHANGELOG.md)"
+
           {
             echo "## ScreenTinker ${{ steps.ver.outputs.tag }}"
             echo
-            echo "### Changes"
-            if [ -n "$PREV" ]; then
-              git log --no-merges --pretty='- %s' "${PREV}..${{ steps.ver.outputs.tag }}"
+            if [ -n "$(printf '%s' "$CHANGELOG_BODY" | tr -d '[:space:]')" ]; then
+              echo "$CHANGELOG_BODY"
             else
-              echo "_First tagged release. Most recent changes:_"
-              git log --no-merges --pretty='- %s' -n 30 "${{ steps.ver.outputs.tag }}"
+              # No entry for this version — fall back to commit subjects rather than publish a
+              # release with no notes at all. scripts/bump-version.sh already warns when the
+              # CHANGELOG has no matching heading; this is the same gap showing up downstream.
+              echo "_No CHANGELOG entry for $VERSION; listing commits instead._"
+              echo
+              echo "### Changes"
+              if [ -n "$PREV" ]; then
+                git log --no-merges --pretty='- %s' "${PREV}..${{ steps.ver.outputs.tag }}"
+              else
+                echo "_First tagged release. Most recent changes:_"
+                git log --no-merges --pretty='- %s' -n 30 "${{ steps.ver.outputs.tag }}"
+              fi
             fi
             echo
             echo "### Artifacts"