…lidation)

DO-NOT-MERGE — overlay-only PR to get an empirical AWS Device Farm signal on
whether the latest speech stack drives Parakeet on Android GPUs (Pixel 9/Mali +
S25/Adreno 830). This is the inverse of the CPU-only workaround in #2525.

Changes (packages/transcription-parakeet):
- ParakeetModel::load — remove the __ANDROID__ guard that forced useGPU=false.
- CMakeLists — widen the Android backend-staging glob from
  libqvac-speech-ggml-cpu-*.so to libqvac-speech-ggml-*.so so the Vulkan/OpenCL
  MODULE libs ship in the prebuild (reverses the [0.7.2] CPU-only packaging);
  refresh the now-stale "intentionally CPU-only" comments.
- gpu-smoke.test.js — drop the four Android early-pass skips so the strict
  assertGpuBackend (backendDevice=1, backendId Vulkan/OpenCL) runs on device.
- vcpkg overlay ports (in-package) — ggml-speech@44fd4817 (speech HEAD) +
  parakeet-cpp@ed749556 (whisper.cpp master), wired via the overlay-ports entry
  in vcpkg-configuration.json. Registry baseline and registry version>= pins are
  unchanged; the registry PR is deferred.
- vcpkg.json — bump parakeet-cpp version>= to the overlay version-date.

Local device finding (Adreno 740 / iQOO 11), TDT q4_0, recorded for reviewers:
- CPU: correct transcript, backendDevice=0.
- GPU OpenCL (engine auto-selects this on Adreno>700): aborts in graph-compute —
  "op not supported joint.token_argmax (ARGMAX)" -> GGML_ASSERT (SIGABRT).
- GPU Vulkan (forced by withholding the OpenCL module): runs (backendId=3) but
  output is degraded vs CPU (dropped words) and ~2x slower; NOT the byte-identical
  result ggml-speech 8bf760f4 reported. Expect the Device Farm Adreno (S25) leg to
  hit the OpenCL ARGMAX abort and the Mali leg to exercise the Vulkan path.
  Do not merge — this is a measurement vehicle.

Bump the parakeet-cpp overlay to 06cef8e7 (off ed749556). The TDT
transducer's per-step GPU graphs do an in-place read-and-write of the LSTM
persistent state; Adreno OpenCL drops those aliased ggml_cpy writes, so the
prediction state never advances and the decode emits one constant token per
frame. The fix routes the TDT decode to the host scalar path on OpenCL while
the encoder still runs on the GPU (stats.backendId stays OpenCL). enc_proj is
write-only so it's fine; EOU/Sortformer don't use this persistent-state
pattern, so they already ran correctly on OpenCL.

Verified on-device (Adreno 740 / iQOO 11): TDT-OpenCL now matches the CPU
baseline byte-for-byte; TDT-Vulkan/CPU and EOU/Sortformer-OpenCL unchanged.

- vcpkg-overlay-ports/parakeet-cpp: REF/SHA512 -> 06cef8e7, version-date 2026-06-15
- vcpkg.json: parakeet-cpp version>= 2026-06-15

Bump the parakeet-cpp overlay to bb585eb1: ARM Mali (Valhall) Vulkan
mis-computes every parakeet model (its narrow subgroup width breaks the
ggml-vulkan shaders), so the engine guards Mali by name and routes it to CPU;
Adreno OpenCL and Samsung Xclipse Vulkan are correct and run on the GPU. TDT
host-decode on Adreno OpenCL is unchanged.

The addon surfaces engine gpu_unsupported() as stats.gpuUnsupported, and the
GPU smoke test treats a CPU backend with gpuUnsupported=1 on Android as the
expected, correct result instead of a GPU regression.