fix(engine): weight local-scorer totalTokenScore with TEST_FILE_CONTRIBUTION_WEIGHT

packages/loopover-engine/src/local-scorer.ts

@@ -9,6 +9,7 @@
 // moving the large, Node-coupled local-branch.ts; they are structurally identical to local-branch.ts's
 // definitions. isCodeFile/isTestPath are the same portable classifiers local-branch.ts already delegates to.
 
+import { TEST_FILE_CONTRIBUTION_WEIGHT } from "./scoring/model.js";
 import { isCodeFile, isTestPath } from "./signals/test-evidence.js";
 
 export type LocalScorerChangedFile = {
@@ -50,8 +51,13 @@ export function computeLocalScorerTokens(input: { changedFiles: LocalScorerChang
   const files = input.changedFiles.filter((file) => !file.binary);
   const testTokenScore = files.filter((file) => isTestPath(file.path)).reduce((sum, file) => sum + fileLines(file), 0);
   const sourceTokenScore = files.filter((file) => isCodeFile(file.path)).reduce((sum, file) => sum + fileLines(file), 0);
-  const totalTokenScore = files.reduce((sum, file) => sum + fileLines(file), 0);
-  const nonCodeTokenScore = Math.max(0, totalTokenScore - sourceTokenScore - testTokenScore);
+  const rawLineTotal = files.reduce((sum, file) => sum + fileLines(file), 0);
+  const nonCodeTokenScore = Math.max(0, rawLineTotal - sourceTokenScore - testTokenScore);
+  // #8875: weight test-file tokens with TEST_FILE_CONTRIBUTION_WEIGHT (0.05×) so this total mirrors
+  // buildScorePreview's `derivedTotalTokenScore`. Emitting a raw unweighted line sum here let a caller feed it
+  // back into the preview as an explicit `totalTokenScore`, which bypasses the test-file discount preview applies
+  // when it derives the total itself — over-counting a test-heavy diff's test lines in the contribution-bonus ramp.
+  const totalTokenScore = sourceTokenScore + TEST_FILE_CONTRIBUTION_WEIGHT * testTokenScore + nonCodeTokenScore;
   const failed = (input.validation ?? []).some((entry) => entry.status === "failed");
   const warnings = failed ? ["Local validation reported failures — token scores describe the diff, not a passing build."] : [];
   return {

packages/loopover-engine/src/scoring/model.ts

@@ -8,6 +8,11 @@ import type { ScoringModelSnapshotRecord } from "./types.js";
 
 export const DEFAULT_ISSUE_DISCOVERY_SHARE = 0.5;
 
+// Upstream weights test-file tokens at 0.05× relative to source tokens (#808). Exported as a strongly-typed
+// named constant so non-snapshot callers (local-scorer.ts) can weight their own totals without indexing the
+// `Record<string, number>` map — keeping the value single-sourced with DEFAULT_SCORING_CONSTANTS below.
+export const TEST_FILE_CONTRIBUTION_WEIGHT = 0.05;
+
 export const DEFAULT_SCORING_CONSTANTS: Record<string, number> = {
   OSS_EMISSION_SHARE: 0.9,
   // Upstream name is ISSUES_TREASURY_EMISSION_SHARE (plural). The prior singular spelling never matched
@@ -21,7 +26,7 @@ export const DEFAULT_SCORING_CONSTANTS: Record<string, number> = {
   MAX_CONTRIBUTION_BONUS: 5,
   CONTRIBUTION_SCORE_FOR_FULL_BONUS: 1500,
   // Applied in preview.ts when computing totalTokenScore from components (#808).
-  TEST_FILE_CONTRIBUTION_WEIGHT: 0.05,
+  TEST_FILE_CONTRIBUTION_WEIGHT,
   // Upstream-enforced eligibility floors for PR and issue-discovery history (#808).
   // These gate whether a validator counts a contributor's submissions, not the per-PR/issue score itself.
   // Stored here so they sync from upstream and no longer appear as unmodeled drift warnings.

packages/loopover-engine/test/local-scorer.test.ts

@@ -20,7 +20,7 @@ test("classifies source / test / non-code disjointly from changed-file metadata"
   assert.equal(r.sourceTokenScore, 12);
   assert.equal(r.testTokenScore, 6);
   assert.equal(r.nonCodeTokenScore, 4);
-  assert.equal(r.totalTokenScore, 22);
+  assert.equal(r.totalTokenScore, 16.3); // #8875: 12 source + 0.05 × 6 test + 4 non-code
   assert.equal(r.warnings, undefined);
 });
 

test/unit/local-scorer.test.ts

@@ -1,5 +1,22 @@
 import { describe, expect, it } from "vitest";
 import { computeLocalScorerTokens } from "../../src/signals/local-scorer";
+import { buildScorePreview } from "../../src/scoring/preview";
+import type { ScoringModelSnapshotRecord } from "../../src/types";
+
+// Minimal snapshot: empty `constants` means every constant (incl. TEST_FILE_CONTRIBUTION_WEIGHT) falls back to
+// DEFAULT_SCORING_CONSTANTS, so preview's derived total uses the same 0.05× weight the local scorer applies.
+const snapshot: ScoringModelSnapshotRecord = {
+  id: "local-scorer-fixture",
+  sourceKind: "test",
+  sourceUrl: "fixture://constants.py",
+  fetchedAt: "2026-05-23T00:00:00.000Z",
+  activeModel: "current_density_model",
+  constants: {},
+  programmingLanguages: {},
+  registrySnapshotId: "registry-fixture",
+  warnings: [],
+  payload: {},
+};
 
 describe("computeLocalScorerTokens (#782)", () => {
   it("classifies source / test / non-code from metadata and sums additions + deletions", () => {
@@ -16,7 +33,7 @@ describe("computeLocalScorerTokens (#782)", () => {
       sourceTokenScore: 12,
       testTokenScore: 8,
       nonCodeTokenScore: 6,
-      totalTokenScore: 26,
+      totalTokenScore: 18.4, // #8875: 12 source + 0.05 × 8 test + 6 non-code (test lines weighted, not raw-summed)
       sourceLines: 12,
     });
     expect(scorer.warnings).toBeUndefined();
@@ -71,4 +88,37 @@ describe("computeLocalScorerTokens (#782)", () => {
     expect(computeLocalScorerTokens({ changedFiles: [{ path: "src/a.ts", additions: 1 }], validation: [{ command: "t", status: "passed" }] }).warnings).toBeUndefined();
     expect(computeLocalScorerTokens({ changedFiles: [{ path: "src/a.ts", additions: 1 }] }).warnings).toBeUndefined();
   });
+
+  it("test-heavy diff: totalTokenScore weights test lines and agrees with buildScorePreview's derived total (#8875)", () => {
+    const scorer = computeLocalScorerTokens({
+      changedFiles: [
+        { path: "src/widget.ts", additions: 20, deletions: 5 }, // source: 25
+        { path: "test/widget.test.ts", additions: 180, deletions: 20 }, // test: 200 (test-heavy)
+        { path: "docs/widget.md", additions: 10, deletions: 0 }, // non-code: 10
+      ],
+    });
+    expect(scorer).toMatchObject({ sourceTokenScore: 25, testTokenScore: 200, nonCodeTokenScore: 10 });
+    // 25 source + 0.05 × 200 test + 10 non-code = 45 — NOT the raw unweighted 235.
+    expect(scorer.totalTokenScore).toBe(45);
+
+    const baseInput = {
+      repoFullName: "octo/demo",
+      sourceTokenScore: scorer.sourceTokenScore,
+      testTokenScore: scorer.testTokenScore,
+      nonCodeTokenScore: scorer.nonCodeTokenScore,
+      sourceLines: scorer.sourceLines,
+      openPrCount: 0,
+      credibility: 1,
+    };
+    // Preview deriving its own total from components, vs. being handed the local scorer's explicit total.
+    const derived = buildScorePreview({ repo: null, snapshot, input: baseInput });
+    const explicit = buildScorePreview({ repo: null, snapshot, input: { ...baseInput, totalTokenScore: scorer.totalTokenScore } });
+    // The two now agree — the explicit total no longer bypasses the test-file discount.
+    expect(explicit.scoreEstimate.contributionBonus).toBe(derived.scoreEstimate.contributionBonus);
+    expect(explicit.scoreEstimate.estimatedMergedScore).toBe(derived.scoreEstimate.estimatedMergedScore);
+
+    // Regression guard: the pre-fix raw unweighted total (25 + 200 + 10 = 235) over-counts test lines in the ramp.
+    const rawUnweighted = buildScorePreview({ repo: null, snapshot, input: { ...baseInput, totalTokenScore: 235 } });
+    expect(rawUnweighted.scoreEstimate.contributionBonus).toBeGreaterThan(explicit.scoreEstimate.contributionBonus);
+  });
 });

test/unit/mcp-cli-plan-scorer-tools.test.ts

@@ -64,7 +64,7 @@ describe("loopover-mcp plan-DAG + local-scorer + predict-gate tools (#6150) —
     expect(data.sourceTokenScore).toBe(14);
     expect(data.testTokenScore).toBe(8);
     expect(data.nonCodeTokenScore).toBe(3);
-    expect(data.totalTokenScore).toBe(25);
+    expect(data.totalTokenScore).toBe(17.4); // #8875: 14 source + 0.05 × 8 test + 3 non-code (test lines weighted)
   });
 
   it("loopover_run_local_scorer surfaces a validation-failure warning without changing the scores", async () => {

test/unit/mcp-run-local-scorer.test.ts

@@ -28,7 +28,7 @@ describe("MCP loopover_run_local_scorer (#782)", () => {
     });
     expect(result.isError).toBeFalsy();
     const data = result.structuredContent as { tokenScores: { mode: string; sourceTokenScore: number; testTokenScore: number; nonCodeTokenScore: number; totalTokenScore: number }; usage: string };
-    expect(data.tokenScores).toMatchObject({ mode: "external_command", sourceTokenScore: 12, testTokenScore: 8, nonCodeTokenScore: 5, totalTokenScore: 25 });
+    expect(data.tokenScores).toMatchObject({ mode: "external_command", sourceTokenScore: 12, testTokenScore: 8, nonCodeTokenScore: 5, totalTokenScore: 17.4 }); // #8875: 12 + 0.05 × 8 + 5
     expect(data.usage).toMatch(/localScorer/);
     expect(JSON.stringify(data)).not.toMatch(/wallet|hotkey|reward|payout|trust score/i);
   });