fix(ci): let check-issue-queue accept honest Refs #N partial work

[BIMvoice]

Summary

• closingIssuesReferences is populated only by a closing keyword, so a PR that is deliberately a slice of a larger issue could not honestly link it without also claiming (and auto-closing) work it did not finish. Five PRs hit exactly this in one evening, all correctly writing Refs #N instead of lying with Closes #N, and all needing a manual unqueued waiver whose meaning ("did not need to wait for the queue") is the opposite of the truth.
• Adds a second, narrower accepted shape in scripts/lib/issue-refs.mjs, consulted only when closingIssuesReferences closes nothing: the PR body may name an issue with a non-closing keyword (Refs, References, Part of, Towards, case-insensitive). That issue must independently be OPEN and carry ready from an authorised applier — the exact same check a closing link already goes through — so:
  • a PR referencing an issue that is not ready still fails (NO_LINKED_ISSUE)
  • a PR referencing a ready issue that is CLOSED still fails (a closed issue is not a queue entry)
  • a self-applied ready label on the referenced issue still fails (SELF_APPLIED_LABEL, same authority check)
  • a PR referencing no issue at all still fails, and unqueued still works unchanged
  • Closes #N on a ready issue is unaffected — it passes before this new code path is ever reached
• closingIssuesReferences stays the primary, preferred signal; nothing here weakens it or adds a self-service bypass.
• A match now requires intent, not just the word. The body regex originally matched Refs #N anywhere in the text, including inside a fenced code block, an inline code span, or a blockquote, and as an ordinary verb mid-sentence ("this function refs Compare codebase with IFC libraries #12 in a loop") — none of which name a queue entry on purpose. Since this module can only widen a fail into a pass, an intentless match was a live bypass: any PR could pass the gate by quoting a ready issue number in a code sample or a quoted reply. extractRefIssueNumbers now strips fenced code blocks, inline code spans, and blockquote lines before scanning, and requires the keyword to start its line (optionally after a list marker: -, *, +, 1., 1)) — which is how a real Refs #N is actually written, and which also rejects the mid-sentence-verb case without a separate rule for it. Refs #N, References #N, Part of #N, Towards #N, and refs: #N all keep working in their ordinary positions.

Implementation notes

• New verdict: PARTIAL_WORK (pass), printed with the same detail as READY_ISSUE.
• Live mode does a second, cheap gh api graphql round trip for the referenced issue(s), but only when the first read already closed nothing and the body names a candidate — an ordinary PR that closes its issue pays no extra cost and cannot be failed by this path.
• All new logic lives in scripts/lib/issue-refs.mjs (kept check-issue-queue.mjs at its existing 957-line module-size budget without raising it — the file's own header carried a genuine pre-existing formatting defect, a NO_LABELS/NO_TIMELINE/NO_CLOSING_ISSUES paragraph spliced mid-sentence into an unrelated story, which is also fixed here as a byproduct of reflowing the prose to make room).
• Related issue Add a gate for a PR whose body admits partial coverage under a bare Closes/Fixes/Resolves keyword #4154 proposes the opposite-direction check (a PR claiming Closes while admitting partial coverage); not implemented here.
• Not a markdown parser: the pre-pass recognises fenced/inline/blockquote by their line-level markers, not full CommonMark (a markdown link's display text, for example, is left alone) — scoped enough to close the confirmed hole without a new dependency, matching this repo's dependency-light gates.

Test plan

• node --test scripts/check-issue-queue.test.mjs scripts/lib/issue-refs.test.mjs — 95/95 pass, including end-to-end cases and unit tests covering the regex, the query builder, the fail-closed live-fetch wrapper, the verdict logic, and the intent-scoping pre-pass.
• Mutation: reverted the change — 3 top-level test() cases in check-issue-queue.test.mjs go RED (one of them loops over 4 keyword variants), everything else stays green.
• Mutation: dropped the state === 'OPEN' filter — the CLOSED-issue test (both the integration test and its lib/issue-refs.test.mjs unit twin) goes RED; reverted.
• Mutation: disabled the fenced/inline-span/blockquote strip — the fenced-code-block test goes RED (the inline-span and blockquote cases are additionally covered by the line-start anchor itself, so they stay green under this specific mutation; both mechanisms independently reject those two shapes); reverted.
• Mutation: removed list-marker support from the keyword regex — the "after a list marker" honest-form test goes RED; reverted.
• node scripts/check-module-size.mjs — OK, 0 new over budget.
• node scripts/check-test-wiring.mjs — OK.

Closes #4147

[coderabbitai]

Warning

Review limit reached

• Run on-demand review

_{This review includes 5 billable files and costs up to $1.25.}

Or wait 59 minutes for your next included review.

Check out review usage here.

View limit details

Limit details: You’ve used all 2 included reviews currently available.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Advanced

Run ID: 8bbd6770-359f-46d3-bee5-8633d672739b

📥 Commits

Reviewing files that changed from the base of the PR and between 1791e64 and c763fba.

📒 Files selected for processing (5)

• .github/workflows/issue-queue.yml
• scripts/check-issue-queue.mjs
• scripts/check-issue-queue.test.mjs
• scripts/lib/issue-refs.mjs
• scripts/lib/issue-refs.test.mjs

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

Claude review - no findings for 5a7694d9e

Reviewed this diff and found nothing to flag.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

2 Skipped Deployments

Project	Deployment	Actions	Updated
ifc-lite-dev	Ignored		Sep 8, 2026 11:41am UTC
ifc-lite-viewer-embed	Ignored		Sep 8, 2026 11:41am UTC

[github-actions]

Claude review - no findings for 910d38fb5

Reviewed this diff and found nothing to flag.

[github-actions]

Claude review - no findings for 4ce585df2

Reviewed this diff and found nothing to flag.

[BIMvoice]

Ran check-issue-queue.mjs from this PR's head (4ce585df2, plus one later commit) live against every open PR whose Issue queue check is currently failing (4/22 red: #4159, #4160, #4151, #4152). Verdicts below are actual gate output, not prediction.

2 of 4 clear once #4161 merges:

PR	Refs	Verdict here	Why
#4159	#4153	✅ PARTIAL_WORK	line-start Refs #4153; #4153 is OPEN, ready applied by louistrue 2026-09-08T09:44:58Z
#4160	#4153	✅ PARTIAL_WORK	same — first line of the body is Refs #4153.

2 stay red — need a manual unqueued, not this fix:

PR	Refs	Verdict here	Why not
#4151	#3612	❌ NO_LINKED_ISSUE	body has `Refs #3612` mid-sentence, backtick-wrapped ("...the actual trigger. Refs #3612, and requesting..."), not at line start — the tightened matcher doesn't count it even though #3612 is OPEN with ready (louistrue, 2026-09-02T09:16:43Z)
#4152	#3357, #4018	❌ NO_LINKED_ISSUE	same shape defect (mid-sentence, "...classification gaps). Refs #3357 #4018...") and both #3357 and #4018 are chore: version packages PRs, not issues (#3357 merged, #4018 open) — wouldn't qualify even with a fixed reference form

Also checked #4167, which the same Refs-#4153 shape would clear (verified: PARTIAL_WORK) — but its Issue queue check is currently queued, not fail, on GitHub, so it isn't part of the 4-PR red count above.

#4151 and #4152 still need unqueued from you regardless of this PR merging.

[github-actions]

Claude review - no findings for c763fba40

Reviewed this diff and found nothing to flag.