fix(signals): require a real shared file before path-driven PR collisions cluster

src/signals/engine.ts

@@ -33,6 +33,7 @@ import { PREFLIGHT_LIMITS } from "./preflight-limits";
 import type { UnifiedCollapsible } from "../review/unified-comment";
 import { splitAiReviewNits } from "../review/ai-notes";
 import { GITTENSORY_GATE_CHECK_NAME } from "../review/check-names";
+import { diffFilePriority } from "../review/review-diff";
 
 export type ParticipationLane = "direct_pr" | "issue_discovery" | "split" | "inactive" | "unknown";
 export type SignalFinding = AdvisoryFinding;
@@ -862,16 +863,23 @@ export function buildCollisionReport(
       }
       const overlap = termOverlap(itemTerms.get(itemKey(left)) ?? collisionTerms(left), itemTerms.get(itemKey(right)) ?? collisionTerms(right));
       if (overlap.score < 0.58 || overlap.shared < 2) continue;
-      // A contributor iterating on their own work (e.g. a follow-up PR touching the same file as their still-open
-      // prior PR) is not duplicate effort. Title/label overlap between a contributor's own items is today's
-      // established behavior (unchanged, e.g. a self-filed issue and its own PR); what's new here is that
-      // `changedFiles` now also feeds this same heuristic, and two of a contributor's own PRs sharing a file is
-      // exactly the false-positive path-overlap creates. Re-score without paths: if the pair only clears the bar
-      // WITH file-path terms, paths alone drove the match — self-authored, so skip it. If title/label terms alone
-      // already clear the bar, this is pre-existing behavior and still clusters.
-      if (isPullRequestShapedItem(left) && isPullRequestShapedItem(right) && Boolean(left.authorLogin) && sameLogin(left.authorLogin, right.authorLogin ?? "")) {
-        const titleOnlyOverlap = termOverlap(collisionTerms(left, false), collisionTerms(right, false));
-        if (titleOnlyOverlap.score < 0.58 || titleOnlyOverlap.shared < 2) continue;
+      // Re-score without path terms: tells us whether title/label overlap ALONE already clears the bar
+      // (pre-existing behavior, unaffected) or whether changedFiles tokens are what pushed this pair over —
+      // the two false-positive shapes that creates are guarded separately below.
+      const titleOnlyOverlap = termOverlap(collisionTerms(left, false), collisionTerms(right, false));
+      const pathDrivenMatch = titleOnlyOverlap.score < 0.58 || titleOnlyOverlap.shared < 2;
+      if (pathDrivenMatch) {
+        // A contributor iterating on their own work (e.g. a follow-up PR touching the same file as their
+        // still-open prior PR) is not duplicate effort — self-authored path-only overlap is dropped outright.
+        if (isPullRequestShapedItem(left) && isPullRequestShapedItem(right) && Boolean(left.authorLogin) && sameLogin(left.authorLogin, right.authorLogin ?? "")) {
+          continue;
+        }
+        // Different authors: file paths tokenize into directory segments (src, review, test, unit, ...) that
+        // recur across nearly every PR in a consistently-organized repo, so shared TOKENS alone are not
+        // reliable collision evidence — a repo-wide shadow test found this drove the large majority of
+        // path-only matches with zero actual shared files. Require an ACTUAL shared file (ignoring
+        // lockfiles/generated artifacts nobody would call a collision over) before clustering.
+        if (!sharesMeaningfulFile(left.changedFiles, right.changedFiles)) continue;
       }
       const key = [itemKey(left), itemKey(right)].sort().join("--");
       if (clusters.has(key)) continue;
@@ -5406,6 +5414,15 @@ function isPullRequestShapedItem(item: CollisionItem): boolean {
   return item.type === "pull_request" || item.type === "recent_merged_pull_request";
 }
 
+/** True when two changed-file lists share at least one path that isn't a lockfile/generated/vendor artifact
+ *  (diffFilePriority's least-useful-to-review bucket) — a shared package-lock.json or dist/ output is touched
+ *  incidentally by unrelated PRs and is not evidence of a real collision. */
+function sharesMeaningfulFile(left: string[] | undefined, right: string[] | undefined): boolean {
+  if (!left || !right || left.length === 0 || right.length === 0) return false;
+  const rightSet = new Set(right);
+  return left.some((path) => rightSet.has(path) && diffFilePriority(path) < 4);
+}
+
 function sameRepo(left: string | null | undefined, right: string | null | undefined): boolean {
   return Boolean(left && right && left.toLowerCase() === right.toLowerCase());
 }

test/unit/signals-v2.test.ts

@@ -257,6 +257,90 @@ describe("v2 signal builders", () => {
       expect(cluster).toBeDefined();
     });
 
+    it("does not flag two open PRs from different authors whose only overlap is generic directory-segment tokens, not a real shared file", () => {
+      // Both touch src/review/*.ts + test/unit/*.ts, so "src"/"review"/"test"/"unit" tokenize as shared terms
+      // even though the two PRs share zero actual file paths — a repo-wide shadow test found this drove most
+      // path-only false positives once changedFiles fed the same bag-of-words scorer as titles/labels.
+      const ivanPr: PullRequestRecord = {
+        repoFullName: repo.fullName,
+        number: 223,
+        title: "CLA check",
+        state: "open",
+        authorLogin: "ivan",
+        authorAssociation: "NONE",
+        labels: [],
+        linkedIssues: [],
+        changedFiles: ["src/review/cla-check.ts", "test/unit/cla-check.test.ts"],
+      };
+      const judyPr: PullRequestRecord = {
+        repoFullName: repo.fullName,
+        number: 224,
+        title: "RAG builder",
+        state: "open",
+        authorLogin: "judy",
+        authorAssociation: "NONE",
+        labels: [],
+        linkedIssues: [],
+        changedFiles: ["src/review/rag.ts", "test/unit/rag.test.ts"],
+      };
+      const report = buildCollisionReport(repo.fullName, [], [ivanPr, judyPr]);
+      expect(findCluster(report, 223, 224)).toBeUndefined();
+    });
+
+    it("does not flag two open PRs whose only shared file is a lockfile (touched incidentally, not real overlap)", () => {
+      const kenPr: PullRequestRecord = {
+        repoFullName: repo.fullName,
+        number: 225,
+        title: "Rotate secrets",
+        state: "open",
+        authorLogin: "ken",
+        authorAssociation: "NONE",
+        labels: [],
+        linkedIssues: [],
+        changedFiles: ["package-lock.json"],
+      };
+      const lisaPr: PullRequestRecord = {
+        repoFullName: repo.fullName,
+        number: 226,
+        title: "Trim caches",
+        state: "open",
+        authorLogin: "lisa",
+        authorAssociation: "NONE",
+        labels: [],
+        linkedIssues: [],
+        changedFiles: ["package-lock.json"],
+      };
+      const report = buildCollisionReport(repo.fullName, [], [kenPr, lisaPr]);
+      expect(findCluster(report, 225, 226)).toBeUndefined();
+    });
+
+    it("flags two open PRs from different authors once a real (non-lockfile) shared file joins otherwise-generic path tokens", () => {
+      const ivanPr: PullRequestRecord = {
+        repoFullName: repo.fullName,
+        number: 227,
+        title: "CLA check",
+        state: "open",
+        authorLogin: "ivan",
+        authorAssociation: "NONE",
+        labels: [],
+        linkedIssues: [],
+        changedFiles: ["src/review/cla-check.ts", "test/unit/cla-check.test.ts", "src/shared/helper.ts"],
+      };
+      const judyPr: PullRequestRecord = {
+        repoFullName: repo.fullName,
+        number: 228,
+        title: "RAG builder",
+        state: "open",
+        authorLogin: "judy",
+        authorAssociation: "NONE",
+        labels: [],
+        linkedIssues: [],
+        changedFiles: ["src/review/rag.ts", "test/unit/rag.test.ts", "src/shared/helper.ts"],
+      };
+      const report = buildCollisionReport(repo.fullName, [], [ivanPr, judyPr]);
+      expect(findCluster(report, 227, 228)).toBeDefined();
+    });
+
     it("flags an open PR against a recently-merged PR from a different author sharing a file (extends to merged history)", () => {
       const openPr: PullRequestRecord = {
         repoFullName: repo.fullName,

worker-configuration.d.ts

@@ -1,5 +1,5 @@
 /* eslint-disable */
-// Generated by Wrangler by running `wrangler types` (hash: ed445a7a260a431e4c5988e63c82a35a)
+// Generated by Wrangler by running `wrangler types` (hash: 11efd919605a2b914db02ea113814228)
 // Runtime types generated with workerd@1.20260617.1 2026-05-28 nodejs_compat
 interface __BaseEnv_Env {
 	DB: D1Database;
@@ -35,7 +35,7 @@ interface __BaseEnv_Env {
 	GITTENSORY_PUBLIC_STATS: "true";
 	GITTENSORY_PUBLIC_STATS_REPOS: "JSONbored/gittensory,JSONbored/awesome-claude,JSONbored/metagraphed";
 	GITTENSORY_DUPLICATE_WINNER: "true";
-	GITTENSORY_OPEN_PR_FILE_COLLISION: "false";
+	GITTENSORY_OPEN_PR_FILE_COLLISION: "true";
 	RATE_LIMITER: DurableObjectNamespace<import("./src/index").RateLimiter>;
 }
 declare namespace Cloudflare {

wrangler.jsonc

@@ -139,9 +139,13 @@
     "GITTENSORY_DUPLICATE_WINNER": "true",
     // Open-PR file-path collision (#2653): enrich changedFiles on the reviewed PR and its open siblings from
     // the pull_request_files cache before building the collision report, so two independently-open PRs on the
-    // same file get flagged the way two title-similar PRs already are. Default OFF — unset/false leaves every
-    // PullRequestRecord's changedFiles unset (byte-identical to today, no extra D1 reads).
-    "GITTENSORY_OPEN_PR_FILE_COLLISION": "false",
+    // same file get flagged the way two title-similar PRs already are. ENABLED: a repo-wide shadow test against
+    // the live open-PR queue found path-only term overlap alone drove 12 of 17 newly-flagged pairs to false
+    // positives (shared generic directory-segment tokens like src/review/test/unit, zero actual shared files) —
+    // fixed by requiring an actual shared, non-lockfile file path before a path-driven match clusters (see the
+    // sharesMeaningfulFile guard in buildCollisionReport). Re-validated post-fix: every remaining flagged pair is
+    // backed by a real shared file.
+    "GITTENSORY_OPEN_PR_FILE_COLLISION": "true",
   },
   "routes": [
     {