revert: "QVAC-19255 tts-ggml: enable Supertonic GPU via tts-cpp 2026-06-05 (#2473)"

[Zbig9000]

🎯 What problem does this PR solve?

• All Android e2e runs have failed since Jun 9, 2026 (0/431 on main, 0/95 on PRs QVAC-16528 chore[skiplog]: prune unused e2e test assets #2487 / QVAC-20451 fix[skiplog]: stabilize Android logging reload e2e #2479). The consumer app aborts ~1 s into bootstrap, which looks like a bootstrap timeout but is actually a native crash.
• @qvac/tts-ggml@0.2.1 fails to load on Android ARM64. Device-farm logcat:

  AddonError: ADDON_NOT_FOUND: Cannot find addon '.' imported from '.../@qvac/tts-ggml/binding.js'
    Candidates: - linked:libqvac__tts-ggml.0.2.1.so
    [cause]: Error: dlopen failed
  F libc: Fatal signal 6 (SIGABRT) in tid ... (mqt_v_js)
  

• iOS and desktop (Linux / macOS / Windows) are unaffected — they statically link the ggml CPU backend, so the missing symbols resolve there.

📝 How does it solve it?

Reverts #2473 (QVAC-19255) in full and pins tts-cpp back to the last-known-good 2026-06-03#1 (the revision 0.2.0 shipped).

Root cause (confirmed against the qvac-registry-vcpkg ports):

• QVAC-19255 tts-ggml: enable Supertonic GPU via tts-cpp 2026-06-05 #2473 bumped tts-cpp 2026-06-03#1 → 2026-06-05. The two portfiles are identical except the upstream pin: qvac-ext-lib-whisper.cpp@c1fd35d8 → @128dae42 — i.e. the QVAC-19254 "sched + cpu_backend refactor."
• That refactor makes direct ggml_backend_is_cpu / ggml_get_type_traits_cpu calls inside the statically-linked tts-cpp library.
• On Android the shared ggml-speech port builds the CPU backend as runtime-dlopen'd per-microarch MODULE .so variants (GGML_CPU_ALL_VARIANTS=ON + GGML_BACKEND_DL=ON; no static CPU archive). Those two symbols are therefore left UND in libqvac__tts-ggml.*.so with no DT_NEEDED able to resolve them — the CPU variant libs are only dlopen'd lazily inside Engine construction, long after Bare loads the addon. The addon can't be linked → unhandled rejection on the mqt_v_js worklet → abort().

Because tts-cpp is reverted, Supertonic is CPU-only again, so this PR also restores (all from #2473): the SupertonicModel.cpp / index.js useGPU / nGpuLayers rejection gates, the C++ unit tests, the gpu-smoke.test.js integration test, and the README / index.d.ts / examples — keeping the package internally consistent.

Released as 0.2.2, not a rollback to 0.2.0: the broken 0.2.1 dev build is already in the package registry the e2e installs from (+ @qvac/tts-ggml@0.2.1 in the producer log), and the SDK depends on ^0.2.0, so the fix must carry a higher version to be selected.

🧪 How was it tested?

• This change is the exact inverse of QVAC-19255 tts-ggml: enable Supertonic GPU via tts-cpp 2026-06-05 #2473 (which merged green) plus a version + changelog bump; the restored sources are byte-identical to the verified-good 0.2.0 tree.
• The team already validated the equivalent pin (tts-ggml 0.2.0 → tts-cpp 2026-06-03#1) on Android: smoke suite all green — actions/runs/27197673590 (branch tmp-pin-tts-ggml-0.2.0-android-fix).
• Android prebuild + e2e are exercised by CI on this PR.

💥 Breaking Changes

• Supertonic GPU execution (added in 0.2.1) is removed again on all platforms; Supertonic is CPU-only as in 0.2.0. useGPU: true / non-zero nGpuLayers on Supertonic throw at construction again (Chatterbox is unchanged). This is a temporary regression to unblock Android.

➡️ Follow-up (permanent fix)

The proper fix belongs upstream — QVAC-19254 follow-up against tts-cpp / ggml-speech: make ggml_backend_is_cpu / ggml_get_type_traits_cpu defined-and-internal on Android (statically link ggml-cpu into the addon the way desktop/iOS already do, keeping the GPU backends dynamic). The Supertonic GPU work can re-land once that is in place.

---

_{Reverts commit 79378e3c. Drafted by a Cursor agent; please review the root-cause analysis and version bump before merging.}

---

• To see the specific tasks where the Asana app for GitHub is being used, see below:
  • https://app.asana.com/0/0/1215539484661519

[github-actions]

Mobile integration tests — @qvac/tts-ggml (Android)

Result: passed

metric	value
Devices passed	2
Devices failed	0
Test cases total	6
Test cases passed	6
Test cases failed	0
Test cases skipped	0

View workflow run

[Zbig9000]

Decision (Jun 10): this revert is the chosen unblock.

Going revert-first to get Android e2e green now; the only thing it gives up is Supertonic GPU, which is re-landed later via the validated upstream fix. Supertonic is heavily CPU-optimised, so interim CPU-only is acceptable.

• Unblock confirmed: the SDK still depends on @qvac/tts-ggml: ^0.2.0 (unchanged by QVAC-20400 chore: Bump addon deps to latest releases for M-RoPE K-shift #2504), so the 0.2.2 bump here supersedes the broken 0.2.1 dev build; tts-cpp is pinned back to the last-known-good 2026-06-03#1.
• Branch is current with main (merged) and mergeable.

Re-land path (parked in #2506). f7d4d6c is the validated upstream fix — built into the addon it loads + does real CPU synthesis on Android (SDK e2e green). To re-land after this merges:

1. Publish f7d4d6c (or its merged successor) to qvac-registry-vcpkg as a new tts-cpp version + baseline bump.
2. Bump this addon's vcpkg.json to it (+ a fresh package version), drop the QVAC-19254 test[skiplog]: overlay tts-cpp@f7d4d6c to validate Android dlopen fix #2506 overlay.
3. Defect B (separate from this crash): Supertonic GPU on Adreno still ggml_aborts during SupertonicModel::synthesize → ggml_backend_graph_compute (useGPU=true). Needs an upstream Adreno fix, or skip Supertonic GPU on Android in gpu-smoke.test.js (mirroring Chatterbox's existing Android skip) before re-enabling Supertonic GPU on Android.

[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