fix(engine): preserve explicit zero pairwise calibration weights (#7477)

Closes #7443

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: RealDiligent

packages/loopover-engine/src/pairwise-calibration.ts

@@ -55,6 +55,10 @@ function finiteNonNegative(value: number | undefined, fallback: number): number
   return value;
 }
 
+function isInvalidWeight(value: number | undefined): boolean {
+  return value !== undefined && (!Number.isFinite(value) || value < 0);
+}
+
 function normalizePairwiseWeights(weights: PairwiseCalibrationWeights | undefined): {
   objectiveAnchor: number;
   pairwiseJudge: number;
@@ -64,7 +68,17 @@ function normalizePairwiseWeights(weights: PairwiseCalibrationWeights | undefine
     pairwiseJudge: finiteNonNegative(weights?.pairwiseJudge, DEFAULT_PAIRWISE_WEIGHTS.pairwiseJudge),
   };
   const total = raw.objectiveAnchor + raw.pairwiseJudge;
-  if (total <= 0) return DEFAULT_PAIRWISE_WEIGHTS;
+  // Preserve explicitly-zeroed weights rather than substituting the defaults: a caller that zeroes every
+  // component must reach the objective-only fallback in computePairwiseCalibrationScore, not silently get
+  // the default 50/50 blend (converges with reviewer-consensus-calibration.ts / #6170; #7443).
+  // NaN/negative inputs still recover to DEFAULT_PAIRWISE_WEIGHTS when the clamped total is empty — same as
+  // pre-#7443 — so the invalid-weight suite keeps asserting the 50/50 default.
+  if (total <= 0) {
+    if (isInvalidWeight(weights?.objectiveAnchor) || isInvalidWeight(weights?.pairwiseJudge)) {
+      return DEFAULT_PAIRWISE_WEIGHTS;
+    }
+    return { objectiveAnchor: 0, pairwiseJudge: 0 };
+  }
   return {
     objectiveAnchor: raw.objectiveAnchor / total,
     pairwiseJudge: raw.pairwiseJudge / total,
@@ -135,11 +149,25 @@ export function computePairwiseCalibrationScore(input: {
     .filter((score): score is number => score !== null);
   const pairwiseJudgeScore =
     stableScores.length === 0 ? null : roundScore(stableScores.reduce((sum, score) => sum + score, 0) / stableScores.length);
-  const weights = normalizePairwiseWeights(input.weights);
-  const compositeScore =
-    pairwiseJudgeScore === null
-      ? objectiveAnchorScore
-      : roundScore(objectiveAnchorScore * weights.objectiveAnchor + pairwiseJudgeScore * weights.pairwiseJudge);
+  const rawWeights = normalizePairwiseWeights(input.weights);
+  // Second-stage usable-weights pass mirrors reviewer-consensus-calibration.ts (#6170 / #7443): zero out any
+  // component whose own signal is unavailable, then fall back to objective-only when that usable total is empty
+  // (covers explicit all-zero weights even when pairwiseJudgeScore is present).
+  const usableWeights = {
+    objectiveAnchor: rawWeights.objectiveAnchor,
+    pairwiseJudge: pairwiseJudgeScore === null ? 0 : rawWeights.pairwiseJudge,
+  };
+  const usableTotal = usableWeights.objectiveAnchor + usableWeights.pairwiseJudge;
+  const weights =
+    usableTotal <= 0
+      ? { objectiveAnchor: 1, pairwiseJudge: 0 }
+      : {
+          objectiveAnchor: usableWeights.objectiveAnchor / usableTotal,
+          pairwiseJudge: usableWeights.pairwiseJudge / usableTotal,
+        };
+  const compositeScore = roundScore(
+    objectiveAnchorScore * weights.objectiveAnchor + (pairwiseJudgeScore ?? 0) * weights.pairwiseJudge,
+  );
   const unstableSamples = samples.filter((sample) => !sample.stable).length;
   const exhaustedSamples = samples.filter((sample) => sample.exhausted).length;
   return {

packages/loopover-engine/test/pairwise-calibration.test.ts

@@ -153,3 +153,17 @@ test("computePairwiseCalibrationScore normalizes invalid weights without produci
   assert.deepEqual(result.weights, { objectiveAnchor: 0.5, pairwiseJudge: 0.5 });
   assert.equal(result.compositeScore, 0.5);
 });
+
+test("computePairwiseCalibrationScore falls back to objective-only when all weights are explicitly zero (#7443)", () => {
+  const result = computePairwiseCalibrationScore({
+    objectiveAnchor: 0.42,
+    samples: [{ attempts: [{ replayFirst: "replay_better", revealedFirst: "revealed_better" }] }],
+    weights: { objectiveAnchor: 0, pairwiseJudge: 0 },
+  });
+
+  // Explicit zeros must not silently restore the 50/50 default, and must not collapse compositeScore to 0
+  // when a real pairwiseJudgeScore is present — fall back to objective-anchor only (#6170 sibling pattern).
+  assert.equal(result.pairwiseJudgeScore, 1);
+  assert.deepEqual(result.weights, { objectiveAnchor: 1, pairwiseJudge: 0 });
+  assert.equal(result.compositeScore, 0.42);
+});

test/unit/engine-calibration-convergence.test.ts

@@ -2,6 +2,7 @@ import { describe, expect, it } from "vitest";
 import {
   computeGateVerdictCompositeCalibrationScore,
   computeFindingSeverityCompositeCalibrationScore,
+  computePairwiseCalibrationScore,
 } from "../../packages/loopover-engine/src/index";
 
 // Converges gate-verdict + finding-severity calibration with reviewer-consensus-calibration.ts's already-correct
@@ -75,3 +76,63 @@ describe("gate-verdict/finding-severity calibration convergence (#6170)", () =>
     ]);
   });
 });
+
+// Extends the #6170 all-zero-weight pattern to pairwise-calibration.ts (#7443). Vitest coverage is what
+// Codecov grades; the engine package's node:test suite mirrors the same assertions.
+describe("pairwise calibration zero-weight convergence (#7443)", () => {
+  it("explicit all-zero weights fall back to objective-only even when pairwiseJudgeScore is present", () => {
+    const result = computePairwiseCalibrationScore({
+      objectiveAnchor: 0.42,
+      samples: [{ attempts: [{ replayFirst: "replay_better", revealedFirst: "revealed_better" }] }],
+      weights: { objectiveAnchor: 0, pairwiseJudge: 0 },
+    });
+    expect(result.pairwiseJudgeScore).toBe(1);
+    expect(result.weights).toEqual({ objectiveAnchor: 1, pairwiseJudge: 0 });
+    expect(result.compositeScore).toBe(0.42);
+  });
+
+  it("NaN/negative weights still recover to the 50/50 default (not the objective-only fallback)", () => {
+    const result = computePairwiseCalibrationScore({
+      objectiveAnchor: 1,
+      samples: [{ attempts: [{ replayFirst: "revealed_better", revealedFirst: "replay_better" }] }],
+      weights: { objectiveAnchor: Number.NaN, pairwiseJudge: -1 },
+    });
+    expect(result.weights).toEqual({ objectiveAnchor: 0.5, pairwiseJudge: 0.5 });
+    expect(result.compositeScore).toBe(0.5);
+  });
+
+  it("non-zero weights take the normalized usable path (covers usableTotal > 0)", () => {
+    const result = computePairwiseCalibrationScore({
+      objectiveAnchor: 0.55,
+      samples: [
+        { attempts: [{ replayFirst: "replay_better", revealedFirst: "revealed_better" }] },
+        { attempts: [{ replayFirst: "tie", revealedFirst: "tie" }] },
+      ],
+      weights: { objectiveAnchor: 1, pairwiseJudge: 3 },
+    });
+    expect(result.weights).toEqual({ objectiveAnchor: 0.25, pairwiseJudge: 0.75 });
+    expect(result.compositeScore).toBe(0.7);
+  });
+
+  it("missing pairwise signal zeros that component then falls back to objective-only when usable total is empty", () => {
+    const result = computePairwiseCalibrationScore({
+      objectiveAnchor: 0.42,
+      samples: [{ attempts: [{ replayFirst: "incomparable", revealedFirst: "incomparable" }] }],
+      weights: { objectiveAnchor: 0, pairwiseJudge: 0 },
+    });
+    expect(result.pairwiseJudgeScore).toBeNull();
+    expect(result.weights).toEqual({ objectiveAnchor: 1, pairwiseJudge: 0 });
+    expect(result.compositeScore).toBe(0.42);
+  });
+
+  it("missing pairwise signal with non-zero weights renormalizes to objective-only (covers usableTotal > 0 + null pairwise)", () => {
+    const result = computePairwiseCalibrationScore({
+      objectiveAnchor: 0.42,
+      samples: [{ attempts: [{ replayFirst: "incomparable", revealedFirst: "incomparable" }] }],
+      weights: { objectiveAnchor: 1, pairwiseJudge: 1 },
+    });
+    expect(result.pairwiseJudgeScore).toBeNull();
+    expect(result.weights).toEqual({ objectiveAnchor: 1, pairwiseJudge: 0 });
+    expect(result.compositeScore).toBe(0.42);
+  });
+});