fix(signals): count scoped related-work as the union of PR-specific and preflight clusters

[philluiz2323]

Summary

The public PR panel comment (buildPublicPrIntelligenceComment, src/signals/engine.ts) derived the "Duplicate risk" row's scoped-overlap count from the maximum of two cluster-source lengths, while the panel's "Review context" details section listed their deduplicated union:

const scopedOverlapCount = Math.max(prCollisionClusters.length, args.preflight.collisions.length);   // before: max
// ...
const overlapDetails = relatedWorkDetails(args.pr, prCollisionClusters, args.preflight.collisions);   // dedup union internally

The two sources are different filtered subsets of the same collision report (shared id scheme): prCollisionClusters are clusters whose items include this PR, and args.preflight.collisions are clusters that term-overlap the planned contribution. A cluster can be in one set and not the other, so max(|A|, |B|) understates |A ∪ B|. The result was a self-inconsistent comment: the "Duplicate risk" row could say "3 scoped related-work signals" while the details block directly below enumerated 4+ items, and the understated count also flipped duplicateRiskAction guidance. Closes #424.

Scope

• src/signals/engine.ts:
  • Compute the deduplicated union of prCollisionClusters and args.preflight.collisions once (scopedOverlapClusters), and use its length for scopedOverlapCount.
  • Refactor relatedWorkDetails to take that same deduped cluster set, so the headline count and the listed items are computed from one source and can never diverge again.
• test/unit/signals-coverage.test.ts:
  • Fail-on-revert: a PR that sits in 2 repo collision clusters disjoint from 3 preflight clusters now renders "5 scoped related-work signals" (the union); the old Math.max(2, 3) = 3 value is asserted absent.

Validation

• npx tsc --noEmit -- clean.
• npx vitest run test/unit/signals-coverage.test.ts test/unit/signals.test.ts test/unit/signals-v2.test.ts -- pass (including the new test).
• Full suite -- 1132 passed, 1 skipped (excluding the two unparseable upstream test files and one unrelated mcp-cli timeout that passes in isolation).
• Branch coverage 97.02% (above the 97% gate); engine.ts 98.05% branch.

Safety

• No output-shape change: the panel renders the same rows/sections; only the (now correct, union) scoped-overlap count and the matching action text change, and only when the two cluster sets differ.
• The existing scoped-overlap test (PR-specific clusters empty) is unaffected because union == max there.

Notes

displayScopedCount still caps the rendered value at 9+; this fix corrects the underlying count below that cap and keeps it consistent with the related-work items the panel lists.

[gittensory]

Note

Gittensory Gate skipped

PR closed before full evaluation. No late first comment was created.

Signal	Result	Evidence	Action
Gate result	⚠️ Skipped	#425 is no longer open.	No action.

---

Checked by Gittensory, a quiet PR intelligence layer for OSS maintainers.

[JSONbored]

@philluiz2323 this is ready from my side.

A few notes:

• The fix matches the actual #424 bug: the public panel count now uses the deduped union of PR-specific and preflight clusters.
• The regression covers the case that used to undercount related work.
• Non-Gittensory checks are green, and I do not have code changes to request.