llama.cpp decision runner (D41): reproducible build pipeline + smoke-tested release binary

Start with forks/llama.cpp/build.sh, then read forks/llama.cpp/README.md, docs/FORK_LLAMA_SERVER.md, and the just build-llama-fork recipe. Run the documented CPU build from a fresh clone and record its output; completion requires a provenance manifest and a smoke-tested llama-server that returns a per-choice distribution from a real GGUF at /v1/decision.

Context

D41 established branch-only tracking of a llama.cpp decision-inference fork: base `thecodacus/llama.cpp` @ `parallel-decision` (tree-mode `POST /v1/decision`), plus fast-LoRA kernels when a CUDA port exists (RESEARCH §28; 1Cat-vLLM recorded as V100/SM70 informatics reference only — serving path stays llama.cpp static binary, no Python at runtime).

The skeleton landed on main (c2fec9a) but the tracked pipeline never produced a binary:

• `forks/llama.cpp/build.sh` defaulted to `github.com/opencodifier/llama.cpp` — 404 (fork never promoted); real base is `thecodacus/llama.cpp`
• CMake flags were invented names (`LLAMA_SERVER`, `LLAMA_EXAMPLES`, `GGML_STRIP`) that silently defaulted everything ON — correct names are `LLAMA_BUILD_*`
• `LLAMA_BUILD_TOOLS=OFF` silently removes the `llama-server` target (server lives under `tools/server/`)
• Historical ad129b0/38de7eb builds were ad hoc, never through this script — nothing reproducible to release
• No branch: skeleton work went straight to main

Scope (this ticket, branch `fork/llama-decision-2026-10`)

Working worktree: `/nas/Temp/repos/OpenCodifier-fork` (separate from main working tree).

• `build.sh`: correct repo URL, correct `LLAMA_BUILD_*` flags, `BACKEND=cpu|cuda|vulkan`, provenance manifest (`BUILD-INFO.txt`: base/final commit, backend, host, UTC)
• `forks/llama.cpp/.gitignore` so `build/`, clone, and logs never pollute commits
• Working compiled binary: `BACKEND=cpu` build completes; binary produced at `forks/llama.cpp/build/bin/llama-server`
• Smoke test: built binary serves `POST /v1/decision` against a real GGUF and returns a per-choice distribution; commands + output recorded in the ticket
• Docs aligned: `forks/llama.cpp/README.md` + `docs/FORK_LLAMA_SERVER.md` real repo URL; `just build-llama-fork` recipe verified against the new script
• `BACKEND=cuda` path documented for T5500/V100 builds (build executed there later)
• PR through the DevOps pipeline (CI = `.gitforge.yml` gates)

Explicitly out of scope (linked, other sessions)

• Fast-LoRA kernel A/B (woct0rdho CUDA port does not exist yet — §28) → follow-up after this lands
• Fork CI smoke-build job → follow-up once the build recipe is stable here
• D42 e2e test fix, main-tree CI gates, E1-G eval wrap → main-session work, untouched by this branch

Acceptance

A fresh clone of this branch + `cd forks/llama.cpp && BACKEND=cpu ./build.sh` yields a working `llama-server` whose `/v1/decision` answers a real request, with the build provenance in `BUILD-INFO.txt`.