whisper-cpp + parakeet-cpp: bump to qvac-ext-lib-whisper.cpp@master a34cb6da (whisper KV-cache CPU fallback PR #34, parakeet Adreno-gen regex parse PR #38)

[pratiknarola-t]

Bumps the whisper-cpp and parakeet-cpp ports to tetherto/qvac-ext-lib-whisper.cpp@master HEAD a34cb6da. Both ports source the same repo (whisper at the repo root, parakeet from the parakeet-cpp/ subfolder), so both move to the same REF and share one source archive (SHA512 b0e80b06…).

Ports

• whisper-cpp 1.8.5#1 → 1.8.5#2 — the only whisper-root change since the previous pin (738d2e9e, the v1.8.5 sync, PR Updated llama-cpp to b6435.0.0 #33) is PR QVAC-5810: Update espeak-ng to 1.52.0#2 #34's KV-cache CPU-buffer fallback: when a single K/V tensor exceeds the backend's max single-tensor size, the cache is allocated from the CPU buffer-type so the K/V ops run on CPU while the rest of the model stays on GPU (Adreno-740 whisper-cli fix). whisper's project version is still 1.8.5, so only port-version increments. All other master commits since that pin touch the parakeet-cpp/, tts-cpp/, or vendored ggml/ trees, which this port does not build (WHISPER_USE_SYSTEM_GGML=ON).
• parakeet-cpp 2026-05-26#2 → 2026-06-04#0 — picks up PR QVAC-6293: Updated llama-cpp and qvac-lib-inference-addon-cpp ports #38: parse_adreno_version() rewritten to a std::regex over the lowercased OpenCL device description (ignores the "OpenCL 3.0" API-version noise in the combined description, maps the Snapdragon-X "X" naming to the 800 tier) for robust Adreno-generation detection + backend-select diagnostics.

Scope

tts-cpp (also sourced from this repo) is intentionally left at c1fd35d8 and handled separately.

Dependency floor

ggml-speech version>= 2026-06-02 is unchanged on both ports. Verified neither new source needs the ggml-speech 2026-06-04 Adreno-OpenCL surface (both deltas call only long-standing core ggml APIs / stdlib), and the registry baseline already resolves consumers to 2026-06-04.

Verification

• Shared source archive SHA512 b0e80b06… stable across integrity-checked downloads (valid gzip, size 21108952). CI is the final arbiter.
• whisper-cpp's 0001-move-gnuinstalldirs-before-add-subdirectory-src.patch is unaffected: its only target (top-level CMakeLists.txt) is byte-identical at the old pin and at master.
• vcpkg x-add-version re-run is a clean no-op for both ports → version DB consistent with the committed port trees.
• The old pins are pre-history-rewrite SHAs; forward-progress confirmed by lineage-by-message (whisper PR Updated llama-cpp to b6435.0.0 #33 / parakeet PR QVAC-5397: Add and fix ports for TTS addon #32 both present in master history) + additive-only diffs.

[gianni-cor]

need to bump ggml-speech version