Studio whisper: pair slim bundles on the ggml commit, not the full llama tag

[danielhanchen]

Summary

Fixes curated whisper.cpp dictation going unavailable whenever the llama.cpp prebuilt advances to a newer build that keeps the same ggml commit.

The slim whisper bundle is ggml-less: whisper-server links the ggml runtime (libggml.so.0, libggml-base.so.0, backend modules) out of the installed llama.cpp prebuilt, so each whisper release pins a paired llama release. The installer required that pin to match the installed llama tag exactly (slim_pairing_for_artifact, _slim_release_incompatibility, resolve_selection).

llama fork tags are b<upstream_build>-mix-<ggml_commit>. The build number tracks upstream llama and fork PRs that live in llama-server, outside ggml; the -mix-<ggml_commit> suffix is what fixes the ggml ABI the slim bundle links against. The llama installer always installs the newest llama, so as soon as llama republishes a newer build with the same ggml (a frequent event), the installed tag advances past the whisper pin and the exact-match gate rejects an ABI-identical pairing:

whisper.cpp    no compatible prebuilt (installed llama.cpp b10079-mix-fb3d4ca;
               whisper requires b10069-mix-fb3d4ca)

Both tags carry the same ggml commit fb3d4ca; only the build number differs, so the runtime .so.0 files are byte-for-byte ABI-identical.

Fix

Key the pairing gate on the ggml commit after -mix- instead of the full tag, via a small llama_runtime_pairs helper used in all three comparison sites. requires_ggml_sonames remains the real per-file ABI gate, so a runtime that is actually missing a library still fails closed, and a genuine ggml skew (different -mix-<commit>) still reports incompatible. Tags without a -mix- marker fall back to exact matching, so nothing regresses for non-fork tags. This is first-party-to-first-party pairing (our llama.cpp and whisper.cpp forks), so there is no security or authenticity implication.

Tests

• test_llama_runtime_pairs_keys_on_ggml_commit: exact tag, newer-build-same-ggml, same-build-different-ggml, missing requirement, and non--mix- tags.
• test_slim_pairs_across_llama_build_bump_with_same_ggml: the live failure case (installed b10079-mix-fb3d4ca, pinned b10069-mix-fb3d4ca) now selects the slim bundle instead of degrading to CPU.
• test_slim_build_bump_same_ggml_is_not_a_compatibility_error: a same-ggml build bump does not surface as a release incompatibility.
• Existing skew tests still hold: b10068-mix-old and b99999-mix-0000000 are genuine ggml skews and still reject.
• tests/studio/install/ full suite: 1307 passed, 2 skipped.

[danielhanchen]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Bravo.

Reviewed commit: 67f9b16b5d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".