fix(reranker): score passages with tokenizer text pairs

.github/workflows/publish-npm-on-release.yml

@@ -8,9 +8,9 @@ on:
   workflow_dispatch:
     inputs:
       tag:
-        description: 'Tag to publish (e.g. v2.2.0)'
+        description: 'Tag to publish (e.g. v2.3.0)'
         required: true
-        default: 'v2.2.0'
+        default: 'v2.3.0'
 
 permissions:
   contents: read

src/core/reranker.ts

@@ -21,9 +21,15 @@ const AMBIGUITY_THRESHOLD = 0.08;
 
 interface CrossEncoderTokenizer {
   (
-    query: string,
-    passage: string,
-    options: { padding: boolean; truncation: boolean; max_length: number }
+    text: string,
+    options?: {
+      text_pair?: string | null;
+      padding?: boolean | 'max_length';
+      truncation?: boolean | null;
+      max_length?: number | null;
+      return_tensor?: boolean;
+      return_token_type_ids?: boolean | null;
+    }
   ): unknown;
 }
 
@@ -152,7 +158,8 @@ async function scorePair(query: string, passage: string): Promise<number> {
     throw new Error('[reranker] Model not loaded — call ensureModelLoaded() first');
   }
 
-  const inputs = cachedTokenizer(query, passage, {
+  const inputs = cachedTokenizer(query, {
+    text_pair: passage,
     padding: true,
     truncation: true,
     max_length: 512

src/embeddings/transformers.ts

@@ -70,7 +70,7 @@ export class TransformersEmbeddingProvider implements EmbeddingProvider {
       const { pipeline } = await import('@huggingface/transformers');
 
       // TS2590: pipeline() resolves AllTasks[T] — a union too complex for TSC to represent.
-      // Cast to a simpler signature; the actual return type IS FeatureExtractionPipelineType.
+      // Cast to a simpler feature-extraction signature for both v3 and v4-compatible types.
       type PipelineFn = (
         task: 'feature-extraction',
         model: string,

tests/contextbench-baseline-schema-gate.test.ts

@@ -1,5 +1,12 @@
 import { execFileSync } from 'node:child_process';
-import { chmodSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from 'node:fs';
+import {
+  chmodSync,
+  mkdtempSync,
+  readFileSync,
+  realpathSync,
+  rmSync,
+  writeFileSync
+} from 'node:fs';
 import { tmpdir } from 'node:os';
 import path from 'node:path';
 import { describe, expect, it, vi } from 'vitest';
@@ -55,6 +62,10 @@ function cleanupSessionRoot(sessionRoot: string): void {
   }
 }
 
+function normalizeFilesystemPath(filePath: string): string {
+  return realpathSync(filePath);
+}
+
 function tempSessionRoot(phase: 'phase40' | 'phase41' = 'phase40'): string {
   return path.join(
     mkdtempSync(path.join(tmpdir(), `contextbench-${phase}-schema-gate-`)),
@@ -398,7 +409,9 @@ describe('ContextBench Phase 40 schema gate', () => {
         repoCheckoutPath: repoPath,
         verificationStrict: false
       });
-      expect(readFileSync(cwdCapturePath, 'utf8')).toBe(repoPath);
+      expect(normalizeFilesystemPath(readFileSync(cwdCapturePath, 'utf8'))).toBe(
+        normalizeFilesystemPath(repoPath)
+      );
       const stdin = readFileSync(stdinCapturePath, 'utf8');
       expect(stdin).toContain('Problem statement:');
       expect(stdin).toContain('Fix the failing ContextBench task');
@@ -584,7 +597,9 @@ describe('ContextBench Phase 40 schema gate', () => {
           ],
           { encoding: 'utf8', env }
         );
-        expect(readFileSync(cwdPath, 'utf8')).toBe(repoPath);
+        expect(normalizeFilesystemPath(readFileSync(cwdPath, 'utf8'))).toBe(
+          normalizeFilesystemPath(repoPath)
+        );
       }
       execFileSync(
         'node',
@@ -681,7 +696,9 @@ describe('ContextBench Phase 40 schema gate', () => {
         (candidate) => candidate.scoring && 'baselineArmId' in candidate.scoring
       );
       expect(row).toMatchObject({ status: 'completed' });
-      expect(readFileSync(cwdCapturePath, 'utf8')).toBe(repoPath);
+      expect(normalizeFilesystemPath(readFileSync(cwdCapturePath, 'utf8'))).toBe(
+        normalizeFilesystemPath(repoPath)
+      );
       const stdin = readFileSync(stdinCapturePath, 'utf8');
       expect(stdin).toContain('Problem statement:');
       expect(stdin).toContain('Run the diagnostic arm with materialized task text.');

tests/release-truth-surfaces.test.ts

@@ -73,17 +73,17 @@ function extractMarkdownLinks(markdown: string): string[] {
 describe('release truth surfaces', () => {
   const packageJson = readJson<PackageJson>('package.json');
   const releaseManifest = readJson<ReleaseManifest>('.release-please-manifest.json');
+  const expectedVersion = packageJson.version;
   const changelog = readText('CHANGELOG.md');
   const readme = readText('README.md');
   const workflow = readText('.github/workflows/publish-npm-on-release.yml');
   const todoDoc = readOptionalText('docs/TODO.md');
   const visualsDoc = readOptionalText('docs/visuals.md');
   const packagedPaths = ['README.md', 'LICENSE', ...(packageJson.files ?? [])];
 
-  it('keeps package metadata, release manifest, and changelog on 2.2.0', () => {
-    expect(packageJson.version).toBe('2.2.0');
-    expect(releaseManifest['.']).toBe('2.2.0');
-    expect(changelog).toContain('## [2.2.0]');
+  it('keeps package metadata, release manifest, and changelog aligned', () => {
+    expect(releaseManifest['.']).toBe(expectedVersion);
+    expect(changelog).toContain(`## [${expectedVersion}]`);
     expect(changelog).not.toContain('## Unreleased');
   });
 
@@ -114,8 +114,8 @@ describe('release truth surfaces', () => {
     expect(visualsDoc).toContain('Historical snapshot');
   });
 
-  it('keeps the manual publish fallback aligned to v2.2.0', () => {
-    expect(workflow).toContain("description: 'Tag to publish (e.g. v2.2.0)'");
-    expect(workflow).toContain("default: 'v2.2.0'");
+  it('keeps the manual publish fallback aligned to the package version', () => {
+    expect(workflow).toContain(`description: 'Tag to publish (e.g. v${expectedVersion})'`);
+    expect(workflow).toContain(`default: 'v${expectedVersion}'`);
   });
 });

tests/reranker-recovery.test.ts

@@ -63,7 +63,12 @@ describe('reranker corruption recovery', () => {
     await fs.mkdir(modelCacheDir, { recursive: true });
     await fs.writeFile(path.join(modelCacheDir, 'model.onnx'), 'corrupt');
 
-    const tokenizer = vi.fn((query: string, passage: string) => ({ query, passage }));
+    const tokenizer = vi.fn(
+      (query: string, options?: { text_pair?: string | null }) => ({
+        query,
+        passage: options?.text_pair ?? ''
+      })
+    );
     const model = vi.fn(async (inputs: { passage: string }) => {
       if (inputs.passage.includes('/a.ts')) return { logits: { data: [1] } };
       if (inputs.passage.includes('/b.ts')) return { logits: { data: [3] } };
@@ -89,6 +94,33 @@ describe('reranker corruption recovery', () => {
     const reranked = await rerank('auth token', results);
     expect(transformersMocks.tokenizerFromPretrained).toHaveBeenCalledTimes(2);
     expect(transformersMocks.modelFromPretrained).toHaveBeenCalledTimes(2);
+    expect(tokenizer).toHaveBeenCalledWith(
+      'auth token',
+      expect.objectContaining({
+        text_pair: expect.stringContaining('/a.ts'),
+        padding: true,
+        truncation: true,
+        max_length: 512
+      })
+    );
+    expect(tokenizer).toHaveBeenCalledWith(
+      'auth token',
+      expect.objectContaining({
+        text_pair: expect.stringContaining('/b.ts'),
+        padding: true,
+        truncation: true,
+        max_length: 512
+      })
+    );
+    expect(tokenizer).toHaveBeenCalledWith(
+      'auth token',
+      expect.objectContaining({
+        text_pair: expect.stringContaining('/c.ts'),
+        padding: true,
+        truncation: true,
+        max_length: 512
+      })
+    );
     expect(reranked.map((result) => result.filePath)).toEqual(['/b.ts', '/c.ts', '/a.ts']);
     expect(getRerankerStatus()).toBe('active');
   });