QVAC-18421 test[skiplog]: add e2e regression tests for Bergamot vocab cache invalidation

packages/sdk/tests-qvac/tests/desktop/consumer.ts

@@ -50,6 +50,7 @@ import { ModelLoadingExecutor } from "../shared/executors/model-loading-executor
 import { CompletionExecutor } from "../shared/executors/completion-executor.js";
 import { ToolsExecutor } from "../shared/executors/tools-executor.js";
 import { TranslationExecutor } from "../shared/executors/translation-executor.js";
+import { TranslationBergamotCacheExecutor } from "../shared/executors/translation-bergamot-cache-executor.js";
 import { ShardedModelExecutor } from "../shared/executors/sharded-model-executor.js";
 import { HttpEmbeddingExecutor } from "../shared/executors/http-embedding-executor.js";
 import { KvCacheExecutor } from "../shared/executors/kv-cache-executor.js";
@@ -390,6 +391,8 @@ export const executor = createExecutor({
     new ErrorExecutor(resources),
     new ToolsExecutor(resources),
 
+    // Must precede TranslationExecutor — patterns overlap, dispatch is first-match-wins.
+    new TranslationBergamotCacheExecutor(),
     new TranslationExecutor(resources),
     new ShardedModelExecutor(resources),
     new OcrExecutor(resources),

packages/sdk/tests-qvac/tests/mobile/consumer.ts

@@ -364,6 +364,10 @@ export const executor = createExecutor({
     new SkipExecutor(/^multi-gpu-/, "Multi-GPU tests disabled on mobile (not supported on single-GPU devices)"),
     new SkipExecutor(/^tools-(?!simple-function$|no-function-match$)/, "Tools test disabled on mobile"),
     new SkipExecutor(/^(diffusion-|addon-logging-diffusion$)/, "SD v2.1 1B Q8_0 cold-load is too heavy for Device Farm devices (iOS variable 5–15min, Android blocks JS thread >300s and trips heartbeat)"),
+    new SkipExecutor(
+      /^translation-bergamot-.+-cache-reload$/,
+      "Server-side Bare code path, identical across platforms — desktop coverage is source of truth",
+    ),
     // suspend() hangs the test runner on mobile (the lifecycle coordinator
     // pauses MQTT/network ops and never resumes within the test timeout).
     // Only resume-idempotent is safe -- it does not call suspend().

packages/sdk/tests-qvac/tests/shared/executors/translation-bergamot-cache-executor.ts

@@ -0,0 +1,100 @@
+import {
+  loadModel,
+  unloadModel,
+  BERGAMOT_FR_EN,
+  BERGAMOT_EN_FR,
+} from "@qvac/sdk";
+import {
+  BaseExecutor,
+  type TestResult,
+} from "@tetherto/qvac-test-suite";
+import { translationBergamotCacheTests } from "../../translation-bergamot-cache-tests.js";
+
+interface BergamotCacheParams {
+  pair: string;
+}
+
+interface ProgressEvent {
+  downloaded: number;
+  total: number;
+  percentage: number;
+  downloadKey: string;
+}
+
+const PAIRS = {
+  "fr-en": { descriptor: BERGAMOT_FR_EN, from: "fr", to: "en" },
+  "en-fr": { descriptor: BERGAMOT_EN_FR, from: "en", to: "fr" },
+} as const;
+
+type Pair = (typeof PAIRS)[keyof typeof PAIRS];
+
+async function loadAndUnload(
+  pair: Pair,
+  onProgress?: (p: ProgressEvent) => void,
+): Promise<void> {
+  const id = await loadModel({
+    modelSrc: pair.descriptor as never,
+    modelType: "nmt",
+    modelConfig: { engine: "Bergamot", from: pair.from, to: pair.to },
+    ...(onProgress && { onProgress }),
+  });
+  await unloadModel({ modelId: id });
+}
+
+function summarizePartials(events: ProgressEvent[]) {
+  const partials = events.filter((e) => e.total > 0 && e.downloaded < e.total);
+  return { partials, touchedKeys: new Set(partials.map((e) => e.downloadKey)) };
+}
+
+export class TranslationBergamotCacheExecutor extends BaseExecutor<
+  typeof translationBergamotCacheTests
+> {
+  pattern = /^translation-bergamot-.+-cache-reload$/;
+
+  protected handlers = Object.fromEntries(
+    translationBergamotCacheTests.map((t) => [t.testId, this.run.bind(this)]),
+  ) as never;
+
+  // Round 1 warms the cache; Round 2 must be a pure cache hit. We detect
+  // re-downloads cross-platform via onProgress: a true cache hit emits at
+  // most one final 100% event per file, while a real download emits many
+  // partial-percentage events.
+  async run(params: BergamotCacheParams): Promise<TestResult> {
+    const pair = PAIRS[params.pair as keyof typeof PAIRS];
+    if (!pair) {
+      return { passed: false, output: `Unknown pair "${params.pair}"` };
+    }
+
+    try {
+      await loadAndUnload(pair);
+    } catch (error) {
+      const msg = error instanceof Error ? error.message : String(error);
+      return { passed: false, output: `Round 1 (warm cache) failed: ${msg}` };
+    }
+
+    const round2: ProgressEvent[] = [];
+    try {
+      await loadAndUnload(pair, (p) => round2.push(p));
+    } catch (error) {
+      const msg = error instanceof Error ? error.message : String(error);
+      return { passed: false, output: `Round 2 (cache hit) failed: ${msg}` };
+    }
+
+    const { partials, touchedKeys } = summarizePartials(round2);
+    if (touchedKeys.size > 0) {
+      const sample = partials
+        .slice(0, 3)
+        .map((e) => `${e.downloadKey}@${e.percentage.toFixed(0)}%`)
+        .join(", ");
+      return {
+        passed: false,
+        output: `Cache invalidation regression: Round 2 re-downloaded ${touchedKeys.size} file(s), ${partials.length} partial events. First: ${sample}`,
+      };
+    }
+
+    return {
+      passed: true,
+      output: `Cache hit on Round 2 — ${round2.length} cache-hit notification(s), no partial downloads`,
+    };
+  }
+}

packages/sdk/tests-qvac/tests/test-definitions.ts

@@ -7,6 +7,7 @@ import { embeddingTests } from "./embedding-tests.js";
 import { ragTests } from "./rag-tests.js";
 import { translationIndicTransTests } from "./translation-indictrans-tests.js";
 import { translationBergamotTests } from "./translation-bergamot-tests.js";
+import { translationBergamotCacheTests } from "./translation-bergamot-cache-tests.js";
 import { translationLlmTests } from "./translation-llm-tests.js";
 import { translationSalamandraTests } from "./translation-salamandra-tests.js";
 import { translationAfriquegemmaTests } from "./translation-afriquegemma-tests.js";
@@ -210,6 +211,9 @@ export const tests = [
   // Translation: Bergamot (EN→FR, EN→ES)
   ...translationBergamotTests,
 
+  // Translation: Bergamot cache reload regression
+  ...translationBergamotCacheTests,
+
   // Translation: LLM (open-vocabulary via from/to)
   ...translationLlmTests,
 

packages/sdk/tests-qvac/tests/translation-bergamot-cache-tests.ts

@@ -0,0 +1,21 @@
+import type { TestDefinition } from "@tetherto/qvac-test-suite";
+
+interface BergamotCacheParams {
+  pair: string;
+}
+
+const cacheReloadTest = (pair: string): TestDefinition => ({
+  testId: `translation-bergamot-${pair}-cache-reload`,
+  params: { pair } satisfies BergamotCacheParams,
+  expectation: { validation: "function", fn: () => true },
+  metadata: {
+    category: "translation-bergamot-cache",
+    dependency: "none",
+    estimatedDurationMs: 180000,
+  },
+});
+
+export const translationBergamotCacheTests: TestDefinition[] = [
+  cacheReloadTest("fr-en"),
+  cacheReloadTest("en-fr"),
+];