fix(MODEL-AUDIO-PARAKEET-TRANSDUCER): fetch parakeet.cpp at master, the branch it actually has

[phantomic12]

One line in CMakeLists.txt. It is the largest single source of red on the pull request lane, and it is not caused by any open branch.

CMakeLists.txt:1596 fetches parakeet.cpp at GIT_TAG main. That repository has no main branch. git ls-remote --symref reports ref: refs/heads/master HEAD, and the head listing has no refs/heads/main. So FetchContent dies during configure, before a single file is compiled:

fatal: invalid reference: main
CMake Error at parakeet_cpp-subbuild/parakeet_cpp-populate-prefix/tmp/
  parakeet_cpp-populate-gitclone.cmake:61 (message):
  Failed to checkout tag: 'main'
Build step for parakeet_cpp failed: 2

VLLM_CPP_WITH_DIARIZATION is ON by default, so this sits on the default feature path. That is also why it reads as a build failure rather than a fetch failure.

This is what the red on the other open PRs is. build-test-cpu, build-test-vulkan, cuda-fat-build, build-newest-gcc, build-test-cpu-arm64, both sanitize-cpu jobs and both verify jobs all fail on this one line, on every open pull request, because they share the same CMakeLists.txt. The push lane does not show it, which is why it survived: the push lane's build job does not configure diarization.

It is pre-existing. The introducing commit is 2833d6300 ("feat(diarization): add diarization and SAS support via parakeet.cpp (ABI v30)"), an ancestor of both the previous base 1de097c46 and current main d15b1cc09. That commit also carries no FOLLOWING_AGENTS_PROTOCOL trailer at all.

The fix is the ref name only. I checked the surface on master before changing it, because a tag that resolves is not the same as a tag that is compatible:

• all nine parakeet C-API symbols this tree calls are present on master — parakeet_capi_load, parakeet_capi_free, parakeet_capi_diarize_path, parakeet_capi_diarize_pcm, parakeet_capi_transcribe_and_diarize, parakeet_capi_free_sas_results, parakeet_capi_free_string, plus the parakeet_ctx and parakeet_sas_result types
• master carries include/parakeet.h and include/parakeet_capi.h, the include root handed to vllm at CMakeLists.txt:1606
• master defines the parakeet library target linked at CMakeLists.txt:1608, and honours the PARAKEET_BUILD_TESTS / PARAKEET_BUILD_EXAMPLES switches set just above the fetch

Evidence, by running the two clones the configure step would run:

git clone --depth 1 --branch main   ... -> fatal: Remote branch main not found in upstream origin
git clone --depth 1 --branch master ... -> succeeds; include/ holds parakeet.h and parakeet_capi.h

Branch tracking is kept rather than swapped for a commit SHA, because the fetch was written to follow a branch deliberately and GIT_SHALLOW ON is paired with it.

Gates. check-agent-record, check-model-checklist, check-build-runtime-deps, check-gate-commands, check-surface-coverage, check-conflict-markers, check-now-current, check-commit-trailers --range and check-commit-style --range all rc=0. agent-issue-index --check is rc=2 on the frozen-archive byte comparison — the CRLF class this tree already tracks, untouched here.

Not run, and it matters. No C++ compile was performed: this host has no toolchain and the build exceeds the command budget. What is proved is that the clone resolves and that the header and library surface this tree binds to is present on master. Whether the resulting translation units compile is unmeasured — CI is the first place that will say.

ISSUE-LOCAL-01M3MQKE9M5BCGS4N80K2RDWG0

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:codebuff/buffy [freebuff]

[mudler-agent]

Review blocked: changing the default dependency from missing branch main to moving branch master fixes today’s fetch but leaves builds unreproducible, and no fresh configure/build evidence was provided. Pin a reviewed immutable commit or release tag and verify configure/build plus the failing old-ref control.