fix(rag): restore live-validated retrieval order

src/lib/answer-ranking.ts

@@ -249,6 +249,7 @@ export function rankAnswerEvidence(
       (a, b) =>
         b.score - a.score ||
         (b.result.hybrid_score ?? b.result.similarity ?? 0) - (a.result.hybrid_score ?? a.result.similarity ?? 0) ||
+        (b.result.score_explanation?.rankScore ?? 0) - (a.result.score_explanation?.rankScore ?? 0) ||
         a.index - b.index,
     );
 

src/lib/rag.ts

@@ -549,6 +549,42 @@ function recordRetrievalLayer(
   }
 }
 
+/**
+ * Keep distinct results in the current ranked order while upgrading duplicate chunks to the
+ * strongest released-hybrid copy.
+ *
+ * App-layer rank scores remain available to answer evidence ranking and telemetry, but the
+ * live corpus gate has not validated replacing the final distinct-result order with them.
+ */
+export function stabilizeReleasedSearchOrder(results: SearchResult[]) {
+  const compareReleasedSearchStrength = (left: SearchResult, right: SearchResult) => {
+    const leftHybrid = left.hybrid_score ?? left.similarity ?? 0;
+    const rightHybrid = right.hybrid_score ?? right.similarity ?? 0;
+    if (rightHybrid !== leftHybrid) return rightHybrid - leftHybrid;
+    const leftSimilarity = left.similarity ?? 0;
+    const rightSimilarity = right.similarity ?? 0;
+    if (rightSimilarity !== leftSimilarity) return rightSimilarity - leftSimilarity;
+    if (right.relevance?.score !== left.relevance?.score)
+      return (right.relevance?.score ?? 0) - (left.relevance?.score ?? 0);
+    return left.id.localeCompare(right.id);
+  };
+  const strongestById = new Map<string, SearchResult>();
+  for (const result of results) {
+    const current = strongestById.get(result.id);
+    if (!current || compareReleasedSearchStrength(result, current) < 0) strongestById.set(result.id, result);
+  }
+  const deduped: SearchResult[] = [];
+  const seen = new Set<string>();
+  for (const result of results) {
+    if (seen.has(result.id)) continue;
+    seen.add(result.id);
+    deduped.push(strongestById.get(result.id) ?? result);
+  }
+  results.length = 0;
+  results.push(...deduped);
+  return results;
+}
+
 /** Record search score telemetry. */
 function recordSearchScoreTelemetry(telemetry: SearchTelemetry, results: SearchResult[]) {
   if (!results.length) {
@@ -565,17 +601,7 @@ function recordSearchScoreTelemetry(telemetry: SearchTelemetry, results: SearchR
     return;
   }
 
-  // Preserve the rankScore ordering established by clinical/second-stage ranking. Coverage
-  // telemetry still uses the raw hybrid signal, but must never mutate the returned result order.
-  const deduped: SearchResult[] = [];
-  const seen = new Set<string>();
-  for (const result of results) {
-    if (seen.has(result.id)) continue;
-    seen.add(result.id);
-    deduped.push(result);
-  }
-  results.length = 0;
-  results.push(...deduped);
+  stabilizeReleasedSearchOrder(results);
   const coverageScores = results
     .map((result) => Math.max(0, result.hybrid_score ?? result.similarity ?? 0))
     .sort((left, right) => right - left);

tests/answer-ranking.test.ts

@@ -179,6 +179,26 @@ describe("answer evidence ranking", () => {
       ranking.scoresByChunkId.get("low-semantic") ?? 0,
     );
   });
+
+  it("uses the app-layer rank only as a final answer-evidence tie-breaker", () => {
+    const lowerRank = result({
+      id: "lower-rank",
+      content: "Lithium dosing guidance.",
+      hybrid_score: 0.7,
+      score_explanation: { rankScore: 0.8 } as NonNullable<SearchResult["score_explanation"]>,
+    });
+    const higherRank = result({
+      id: "higher-rank",
+      content: "Lithium dosing guidance.",
+      hybrid_score: 0.7,
+      score_explanation: { rankScore: 1.1 } as NonNullable<SearchResult["score_explanation"]>,
+    });
+
+    const ranking = rankAnswerEvidence("lithium dosing guidance", [lowerRank, higherRank]);
+
+    expect(ranking.rankedResults.map((item) => item.id)).toEqual(["higher-rank", "lower-rank"]);
+    expect(ranking.scoresByChunkId.get("higher-rank")).toBe(ranking.scoresByChunkId.get("lower-rank"));
+  });
 });
 
 describe("high-yield answer bolding", () => {

tests/rag-second-stage-ranking.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, it } from "vitest";
-import { applySecondStageRerankIfNeeded } from "../src/lib/rag";
+import { applySecondStageRerankIfNeeded, stabilizeReleasedSearchOrder } from "../src/lib/rag";
 import type { SearchTelemetry } from "../src/lib/rag-contracts";
 import type { SearchResult, SearchScoreExplanation } from "../src/lib/types";
 
@@ -136,4 +136,54 @@ describe("second-stage rank score", () => {
     expect(ranked.map((item) => item.id)).toEqual(["dose-evidence-without-explanation", "supported"]);
     expect(ranked[1].score_explanation?.finalRank).toBe(2);
   });
+
+  it("keeps distinct rank order while upgrading duplicate chunks to the strongest released-hybrid copy", () => {
+    const telemetry = {} as SearchTelemetry;
+    const strongerHybridDuplicate = result({
+      id: "stronger-hybrid",
+      document_id: "older-doc",
+      hybrid_score: 0.8,
+      score_explanation: explanation(0.4),
+      source_metadata: {
+        source_title: "Older guideline",
+        publisher: "Test publisher",
+        jurisdiction: "Australia/WA",
+        version: "1",
+        publication_date: null,
+        review_date: null,
+        uploaded_at: null,
+        indexed_at: null,
+        uploaded_by: null,
+        document_status: "outdated",
+        clinical_validation_status: "locally_reviewed",
+        extraction_quality: "good",
+      },
+    });
+    const higherRankScore = result({
+      id: "higher-rank-score",
+      document_id: "current-doc",
+      hybrid_score: 0.79,
+      score_explanation: explanation(0.5),
+    });
+    const weakerDuplicate = {
+      ...strongerHybridDuplicate,
+      hybrid_score: 0.78,
+      score_explanation: explanation(0.41),
+      content: "Weaker duplicate",
+    };
+
+    const secondStage = applySecondStageRerankIfNeeded({
+      queryClass: "medication_dose_risk",
+      results: [higherRankScore, weakerDuplicate, strongerHybridDuplicate],
+      telemetry,
+      topK: 3,
+    });
+    expect(secondStage.map((item) => item.id)).toEqual(["higher-rank-score", "stronger-hybrid", "stronger-hybrid"]);
+
+    const stabilized = stabilizeReleasedSearchOrder(secondStage);
+
+    expect(stabilized.map((item) => item.id)).toEqual(["higher-rank-score", "stronger-hybrid"]);
+    expect(stabilized[1].hybrid_score).toBe(0.8);
+    expect(stabilized[1].content).toBe(strongerHybridDuplicate.content);
+  });
 });