feat(tts-ggml): Android dynamic ggml backends

[GustavoA1604]

What problem does this PR solve?

• No dynamic-backend packaging in @qvac/tts-ggml: Android builds need libqvac-speech-ggml-{vulkan,opencl,cpu-android_armv*_*.so} next to the .bare module so ggml_backend_load_all_from_path() can discover backends at runtime. The addon had no BACKENDS_SUBDIR staging, no backendsDir / openclCacheDir JS surface, and no forwarding into tts_cpp::EngineOptions — unlike @qvac/transcription-parakeet and @qvac/llm-llamacpp.
• Upstream tts-cpp not consumable yet: Registry-based backend selection landed in qvac-ext-lib-whisper.cpp#29 (EngineOptions::backends_dir, Adreno tier policy, init_gpu_backend()). This package must pin tts-cpp >= 2026-05-20 and depend on qvac-registry-vcpkg#159 (ggml-speech#4 + tts-cpp port bump).
• Chatterbox useGPU: true default OOMs on Android: GPU backends mirror large f16 S3Gen weights (~1 GB) on top of the mmap'd CPU copy; 8 GB devices hit lmkd SIGKILL. Default should be CPU with explicit opt-in on capable hosts.
• Mobile / CI test friction: iOS bundled reference wav paths were not readable from native code; integration tests relied on local HF→GGUF conversion in CI instead of the QVAC model registry; Chatterbox mobile tests used oversized f16 GGUFs.

How does it solve it?

CMake / prebuilds (packages/tts-ggml/CMakeLists.txt)

• find_package(ggml) for GGML_AVAILABLE_BACKENDS + loose .so glob from vcpkg lib/.
• bare_target + bare_module_target → BACKENDS_SUBDIR (android-arm64/qvac__tts-ggml, …).
• add_bare_module(... EXPORTS INSTALL TARGET ggml::<backend>) + install(FILES libqvac-speech-ggml-*.so) for Vulkan/OpenCL MODULE backends not exposed as IMPORTED targets.
• Android 16 KB page-size link flags (-Wl,-z,max-page-size=16384) — same as parakeet (Pixel 9 class devices).
• Apple compiler-rt force_load for @available → __isPlatformVersionAtLeast (iOS Metal stability on unload/reload).

JS + C++ wiring

• backendsDir / openclCacheDir on constructor options and TTSGgmlRuntimeConfig; default backendsDir → path.join(__dirname, 'prebuilds').
• ChatterboxModel / SupertonicModel compose backendsDir / BACKENDS_SUBDIR into opts.backends_dir; forward opencl_cache_dir.
• JSAdapter reads both fields for Chatterbox and Supertonic configs.

GPU policy

• useGPU defaults to false for Chatterbox (was true in 0.1.1). Opt in with config: { useGPU: true } on Metal / Vulkan / OpenCL hosts.
• #ifdef __ANDROID__ in loadLocked(): forces n_gpu_layers = 0 and logs a warning if the host requested GPU — Vulkan (Mali) and OpenCL (Adreno) paths for Chatterbox/Supertonic graphs are not validated yet. Dynamic backend .so staging still matters on Android for per-arch CPU dlopen even while GPU stays off.

Tests & CI

• @qvac/registry-client devDependency + downloadModel.js registry fetch (q4_0 Chatterbox T3, f16 S3Gen, Supertonic q4_0) with min/max size bands to reject stale caches.
• resolveRefWavPath() — mobile global.assetPaths / Library/Caches/jfk.wav before in-bundle path (fixes iOS ModelFileNotFound).
• Remove iOS workflow HF→GGUF conversion step; models fetched via registry in mobile/integration runs.
• GPU smoke tests skip Android; lifecycle tests use CPU defaults.

Docs & version

• README updated: CPU-by-default, useGPU table, backendsDir / openclCacheDir knobs (mirrors parakeet).
• CHANGELOG 0.1.2, package.json 0.1.2.
• vcpkg.json: tts-cpp >= 2026-05-20.

Breaking changes

Change	Migration
Chatterbox useGPU default true → false	Pass config: { useGPU: true } where you previously relied on the implicit GPU default (macOS Metal, CUDA desktop, etc.).
Android GPU requests ignored at addon boundary	Expected until Vulkan/Mali + OpenCL/Adreno validation completes; CPU + dynamic CPU backends still work.

Supertonic remains CPU-only at construction time when useGPU: true is passed.

How was it tested

Manual CI run here

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

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



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

[GustavoA1604]

/review