QVAC-19213 transcription-parakeet: OpenCL-only overlay test PR for ex…

[pratiknarola-t]

…t-ggml #19 + whisper #38

Pin the addon's ggml-speech + parakeet-cpp vcpkg ports to the two open fix-PR heads (qvac-ext-ggml#19 @89bbeeaa, qvac-ext-lib-whisper.cpp#38 @01a71ec6) via OpenCL-only overlay ports, so the Adreno OpenCL fix can be exercised before the PRs merge. Android features narrowed to opencl; the Android useGPU=false guard is removed so OpenCL is actually used. Temporary -- drop the overlays and bump version>= once the PRs land. Registry baseline untouched.

What problem does this PR solve?

How does it solve it?

Breaking changes

[ogad-tether]

Reviewed the overlay-port mechanism and the source change. The approach is sound and well-documented, and it does its job: the mobile integration tests pass on both Android (6/6) and iOS (6/6), so the Adreno OpenCL fix is exercised pre-merge. A few notes below — leaving this as a comment rather than an approval, mainly because cpp-lint is red and this is a deliberately temporary PR.

What looks good

• Overlay ports pin to immutable commits with SHA512 (vcpkg_from_github REF + SHA512), not branch heads — good supply-chain hygiene for a pre-merge pin.
• Version chain is internally consistent: ggml-speech overlay 2026-05-27#1 satisfies parakeet-cpp's version>= 2026-05-27, and parakeet-cpp overlay 2026-05-26#2 satisfies the root vcpkg.json's version>= 2026-05-26#2.
• The Android dep change (features: ["opencl"], default-features: false) correctly makes Android OpenCL-only; consistent with the overlay's Android default-feature.
• The !vcpkg-overlay-ports/**/*.cmake .gitignore whitelist is necessary (the global *.cmake ignore would otherwise drop the new portfile.cmakes) and correctly scoped.
• Clear "drop this overlay once PR #19 / #38 merges" comments in both portfiles and the manifests.

Blocking-ish

• cpp-lint is failing. The errors are all readability-identifier-naming violations in ParakeetModel.cpp/.hpp (e.g. ggml_log_buf_mutex, tmp_dir, is_gguf, on_segment). I confirmed these exist verbatim on main — they are pre-existing and not introduced by this diff (which only removes the __ANDROID__ useGPU guard). They surface because touching the file triggers clang-tidy on it. If any form of this is meant to merge, the check needs to go green — either fix/suppress the flagged names or scope-exempt the file. Worth flagging since it'll block other PRs that touch this file too. (merge-guard / validate-pr is just the verified-label gate, ignore.)

Minor

• Stale SHAs in the PR description. The body cites qvac-ext-ggml#19 @89bbeeaa and whisper.cpp#38 @01a71ec6, but the portfiles actually pin bff2347b… and 1cc29a3b… (the portfile comments agree with the pins). Looks like the PR branches advanced after the description was written — suggest updating the body so it's clear which heads were tested.
• The useGPU guard removal lives in addon source, not in an overlay, so it persists after the overlays are dropped. That's the intended end-state (OpenCL stabilized on Android, which is now OpenCL-only), but note it also removes the Mali/Vulkan fallback comment/guard entirely — fine as long as Android stays OpenCL-only.
• Couldn't diff the overlay portfiles against the actual registry ggml-speech/parakeet-cpp ports (separate registry repo) to confirm "mirrors the registry port" line-by-line — trusting the description there.

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ❌ PENDING

**Requirements:**
- 1 Team Member approval ❌ (0/1)
- 1 Team Lead OR Management approval ❌ (0/1)



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