build: prefer llama.cpp-mainline over sibling llama.cpp

[smanaton]

Summary

When auto-detecting llama.cpp, prefer a sibling/checkout named llama.cpp-mainline over llama.cpp.

Some forks lag upstream ggml and have produced non-finite (NaN) logits on Q1_0 models. An explicit mainline directory name makes the intended tree obvious. WWAMA_LLAMA_CPP_DIR still overrides entirely.

Dependency

On Windows you still need the MSVC/CUDA build PR (or equivalent) to link successfully. Suggested merge order: windows-fixes first, then this.

Test plan

• With this ordering + Windows build fixes, cargo build without WWAMA_LLAMA_CPP_DIR selected llama.cpp-mainline when present.
• Finite baselines observed on Bonsai Q1_0 when building against mainline.

Notes

Fork staging: smanaton#2

[Copilot]

The job was not started because the account is locked due to a billing issue.

[bitnom]

Please address these before merge:

1. build.rs:373-382 changes the candidate order but does not make Cargo reevaluate the resolver when llama.cpp-mainline is added after a fallback llama.cpp was selected. The script only emits rerun-if-changed for the selected directory (build.rs:23-24). Watch stable candidate parent/sentinel paths, or otherwise ensure a newly available preferred checkout invalidates the cached build.

2. Please update the repository-layout section in README.md:80-91, which still says the lookup checks ../../cpp/llama.cpp first. The documented behavior now contradicts the build script.

3. Add focused automated coverage for candidate ordering, explicit WWAMA_LLAMA_CPP_DIR precedence, and the no-candidate diagnostic. There are currently no CI checks on this PR.

The override semantics and candidate ordering are otherwise sensible, and this PR should remain dependent on the Windows link fixes being corrected.

[bitnom]

Superseded by draft replacement PR #4: #4

The corrected implementation is on the upstream prefer-mainline-llama-fix feature branch, based on the corrected Windows feature branch. This fork branch cannot be updated from the maintainer account, so the original PR will be closed to keep the merge candidates unambiguous.

[bitnom]

Closing as superseded by draft PR #4. The corrected work remains on the feature branches pending multi-platform validation.