QVAC-18995 feat[api]: bump transcription-whispercpp to 0.9.0 and surface backend/GPU stats

[ishanvohra2]

🎯 What problem does this PR solve?

• The SDK was pinned to @qvac/transcription-whispercpp@^0.8.0, missing the 0.9.0 release.
• whisper.cpp 0.9.0 adds backend/GPU runtime stats (active device class, GPU backend family, GPU memory) that the SDK did not expose to consumers.

📝 How does it solve it?

• Bump @qvac/transcription-whispercpp ^0.8.0 → ^0.9.0.
• Thread the four new runtimeStats fields — backendDevice, backendId, gpuMemTotalMb, gpuMemFreeMb — through all three stats layers: the addon read shape (addon-responses.ts), the transcribe op mapping, and the public transcribeStatsSchema.
• All four are optional: CPU runs and the parakeet engine simply omit the ones they don't populate; -1 is the no-accounting sentinel for the GPU memory fields.

🧪 How was it tested?

• New unit test transcription-stats-schema.test.ts: round-trips the GPU fields, asserts they're optional (CPU/parakeet path), accepts the -1 sentinel, and rejects non-numeric backend values (4/4 tests, 11/11 asserts).
• bun run typecheck and eslint on changed files pass.
• bare-sdk parity (check:deps-vs-sdk) verified — whispercpp is an excluded plugin addon, no sync needed.

🔌 API Changes

// transcribe(...) now surfaces whisper backend/GPU info in the final stats frame
for await (const ev of sdk.transcribe({ modelId, audioChunk, metadata: true })) {
  if (ev.done && ev.stats) {
    console.log(ev.stats.backendDevice);  // 0 = CPU, 1 = GPU
    console.log(ev.stats.backendId);      // GPU backend family (BackendId enum)
    console.log(ev.stats.gpuMemTotalMb);  // -1 when no memory accounting
    console.log(ev.stats.gpuMemFreeMb);
  }
}

---

• To see the specific tasks where the Asana app for GitHub is being used, see below:
  • https://app.asana.com/0/0/1215009851268635

[github-actions]

QVAC E2E — windows — ✅ all tests passed (95/95, 586s)

Config: suite=smoke · filter=(none) · exclude=(none)
View run · Artifacts: reports

[github-actions]

QVAC E2E — ios — ✅ all tests passed (82/95, 931s)

Config: suite=smoke · filter=(none) · exclude=(none)
View run · Artifacts: reports · Device Farm logs

[github-actions]

QVAC E2E — android — ❌ failed

Totals: 0/95 passed · 95 failed · 0.0% · 130s
Config: suite=smoke · filter=(none) · exclude=(none)
View run · Artifacts: reports · Device Farm logs

Results by section

• addon-logging: 0/2 ❌
• cancellation: 0/1 ❌
• completion: 0/8 ❌
• config: 0/1 ❌
• config-reload: 0/2 ❌
• delegated-inference: 0/4 ❌
• diffusion: 0/3 ❌
• download: 0/1 ❌
• embedding: 0/2 ❌
• error: 0/5 ❌
• finetune: 0/1 ❌
• http: 0/2 ❌
• kv-cache: 0/4 ❌
• lifecycle: 0/2 ❌
• model: 0/8 ❌
• model-info: 0/4 ❌
• multi-gpu: 0/2 ❌
• ocr: 0/4 ❌
• parakeet: 0/4 ❌
• rag: 0/3 ❌
• registry: 0/4 ❌
• sharded-model: 0/3 ❌
• tools: 0/2 ❌
• transcription: 0/5 ❌
• translation-afriquegemma: 0/1 ❌
• translation-bergamot: 0/3 ❌
• translation-indictrans: 0/2 ❌
• translation-llm: 0/2 ❌
• translation-salamandra: 0/2 ❌
• tts: 0/2 ❌
• video: 0/2 ❌
• vision: 0/3 ❌
• wrong-model: 0/1 ❌

Failed tests

• 95 tests: Consumer died before test could be executed
  Affected: model-load-llm, model-load-embedding, model-load-ocr, model-load-invalid, model-unload, and 90 more

[github-actions]

QVAC E2E — linux — ✅ all tests passed (95/95, 447s)

Config: suite=smoke · filter=(none) · exclude=(none)
View run · Artifacts: reports

[github-actions]

QVAC E2E — macos — ✅ all tests passed (93/95, 268s)

Config: suite=smoke · filter=(none) · exclude=(none)
View run · Artifacts: reports

[simon-iribarren]

CI Status

Build/validation are green, but the PR-attached QVAC Tests workflow is still not green: Android run-producer failed after the consumer registered but never bootstrapped, and macOS/Windows desktop jobs were still pending when I checked. The Android failure log shows all 95 tests marked failed after the consumer was declared unresponsive.

API Surface & Tagging

This is correctly tagged [api]: it adds optional public transcription stats fields (backendDevice, backendId, gpuMemTotalMb, gpuMemFreeMb) and the PR body includes a usage example.

Code Review

No blocking code issues found. The addon response shape, server stats mapping, public schema, and unit coverage line up with the stated API change and preserve existing CPU/parakeet behavior.

Recommendation

Codewise LGTM, but holding approval until the PR-attached tests are green or the Android consumer failure is rerun/acknowledged as infrastructure.

[simon-iribarren]

Codewise LGTM.

[ishanvohra2]

/review

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (2/1)
- 1 Team Lead OR Management approval ✅ (1/1)



---
*This comment is automatically updated when reviews change.*