claude-pr-review quality-gate miscounts "commits since last tag" on release PRs → false-blocking CHANGELOG-mismatch finding

[cbeaulieu-gt]

Symptom

The reusable claude-pr-review.yml@v2 quality-gate false-blocks release PRs. It reported a critical / blocking finding claiming a CHANGELOG/commit-count mismatch that does not exist.

Observed on glitchwerks/claude-configs PR #987 (release v0.27.0):

• Quality-gate finding: "PR/CHANGELOG claim 3 user-facing commits since v0.26.0, but git history shows only 1 commit (#836)." → marked 🔴 critical, quality-gate FAIL, quality-gate-shadow agree:blocking.

• Actual git history (verified):

  $ git rev-list -n1 v0.26.0
  f56481b...
  
  $ git log --oneline v0.26.0..origin/main
  4b5a77a fix(scan): disambiguate source_id ... (#836)
  ed60ad1 fix(test-implementer): format authored test files ... (#973)
  0d7737f feat(router): consume mixed_content decision ... (#704)
  
  $ git merge-base --is-ancestor ed60ad1 v0.26.0 && echo IN || echo NOT-IN
  NOT-IN     # #973 is genuinely unreleased
  $ git merge-base --is-ancestor 0d7737f v0.26.0 && echo IN || echo NOT-IN
  NOT-IN     # #704 is genuinely unreleased
  

  CHANGELOG was correct: 3 unreleased commits. Gate saw only 1.

Likely root cause

On a release PR, the head branch (release/vX.Y.Z) is cut from main and already contains the unreleased commits. The gate appears to compute "commits since last tag" against the wrong base — candidates:

• diffing only the PR's own commit(s) (the single chore(release) commit) instead of git log <last-tag>..HEAD,
• using the squash-merge title of the latest already-merged PR as the entire history,
• a shallow fetch-depth that truncates history so only the most recent commit/tag relationship is visible.

Suggested fix

• Enumerate release commits with git log <last-annotated-tag>..HEAD (after git fetch --tags --unshallow or sufficient fetch-depth), not against the PR diff or a single merge commit.
• For the CHANGELOG cross-check specifically: resolve the prior tag via git describe --tags --abbrev=0 HEAD^ (or the tag named in the CHANGELOG link refs) and count the real range.
• Add a regression fixture: a release-style PR whose branch contains N>1 unreleased commits must not be flagged as a count mismatch.

Impact

Every future claude-configs release PR will hit this and require a manual UI override to merge. Workaround in the meantime: verify the count via git log <tag>..HEAD + merge-base and override-merge if the finding is the stale-base false positive.

Reported from claude-configs PR #987 (merged via UI override after verification).

🤖 Generated by Claude Code on behalf of @cbeaulieu-gt