ci: hold llama.cpp builds at Rust 1.99; Cargo 1.100's build-dir layout breaks them

.github/workflows/release.yml

@@ -147,16 +147,25 @@ jobs:
           # MLX sidecar steps below. Shipping llama.cpp alongside it would mean
           # compiling it from source on every release for an engine that is
           # slower on Apple silicon and that the plug-in no longer offers.
+          #
+          # `rust_toolchain` holds Windows at 1.99: Cargo 1.100 moved build-script
+          # output one directory deeper, and llama-cpp-sys-2's build.rs (up to
+          # 0.1.161) then panics copying its DLLs, while the DLL staging below
+          # relies on where that copy used to put them. The `RUST_VERSION` note
+          # in server-rs/Dockerfile has the details, and what to check before
+          # moving either on.
           - os: windows-latest
             os_name: windows
             arch: x64
             bin_ext: .exe
             cargo_features: --features llamacpp
+            rust_toolchain: '1.99'
           - os: macos-latest
             os_name: macos
             arch: arm64
             bin_ext: ''
             cargo_features: ''
+            rust_toolchain: stable
 
     name: Build LrGeniusAI - ${{ matrix.os_name }}-${{ matrix.arch }}
 
@@ -198,10 +207,14 @@ jobs:
         with:
           repo-token: ${{ secrets.GITHUB_TOKEN }}
 
+      # clippy is named because the action installs with `--profile minimal`.
+      # The runner's preinstalled stable already has it, but a pinned version
+      # is a fresh install, and the build step below lints.
       - name: Install Rust toolchain
         uses: dtolnay/rust-toolchain@4360b52568e2003a75bf9bc1d59f33a8e3fc893c # stable
         with:
-          toolchain: stable
+          toolchain: ${{ matrix.rust_toolchain }}
+          components: clippy
 
       # The `llamacpp` feature compiles llama.cpp from source: cmake for the
       # build itself, libclang for bindgen. Installed explicitly rather than

docs/wiki/Dev-Build-Environment-Setup.md

@@ -77,13 +77,29 @@ build panics without it):
 # Expected: something like C:\VulkanSDK\1.4.xxx.x
 ```
 
+Build this feature with Rust **1.99**, not the current stable. Cargo 1.100
+moved build-script output one directory deeper, and llama-cpp-sys-2's build
+script (up to 0.1.161) then panics while copying its DLLs:
+
+```
+panicked at ...\llama-cpp-sys-2-0.1.156\build.rs:1431:56:
+called `Result::unwrap()` on an `Err` value: Os { code: 3, kind: NotFound, ... }
+```
+
+The release build is held at 1.99 for the same reason (`rust_toolchain` in
+`.github/workflows/release.yml`). The baseline build is not affected.
+
+```powershell
+rustup toolchain install 1.99 --component clippy
+```
+
 Open a **new** terminal so the updated environment variables are picked up,
 then:
 
 ```powershell
 cd server-rs
-cargo build --release -p lrg-server --features llamacpp
-cargo clippy --workspace --all-targets --features llamacpp
+cargo +1.99 build --release -p lrg-server --features llamacpp
+cargo +1.99 clippy --workspace --all-targets --features llamacpp
 ```
 
 This is a from-source llama.cpp compile plus Vulkan shader compilation, so

docs/wiki/Dev-Server-README.md

@@ -653,8 +653,10 @@ the variable and answers only to the API call. See the doc comment on
 ## Docker
 
 `Dockerfile` here builds and ships the compiled binary (multi-stage:
-`rust:1-trixie` builder, `debian:trixie-slim` runtime; trixie because the
-ONNX Runtime libraries need glibc 2.38+). Model files live in the `/models`
+`rust:1.99-trixie` builder, `debian:trixie-slim` runtime; trixie because the
+ONNX Runtime libraries need glibc 2.38+, and Rust held at 1.99 because
+llama.cpp's build breaks on Cargo 1.100's build-dir layout — see the
+`RUST_VERSION` note in the Dockerfile). Model files live in the `/models`
 volume: the Dockerfile points every family's env var above there
 (`/models/siglip2/...`, `/models/face/...`, `/models/bioclip/...`,
 `/models/llm`), which is where the plug-in's *Download AI models* puts them —

server-rs/Dockerfile

@@ -26,8 +26,19 @@
 # Ubuntu 24.04, because the ONNX Runtime libraries ort-sys links need glibc
 # 2.38+ (see the trixie note below); 22.04 has 2.35.
 ARG UBUNTU_VERSION=24.04
+# Held at 1.99 because Cargo 1.100 moved build-script output one directory
+# deeper (target/release/build/<crate>/<hash>/out instead of
+# build/<crate>-<hash>/out), and two things here assume the old layout:
+# llama-cpp-sys-2's build.rs (up to 0.1.161) finds the target dir by walking
+# three levels up from OUT_DIR, and then panics hard-linking its shared libs
+# into a deps/ directory that is not there; and the library collection in
+# the build stage below globs build/llama-cpp-sys-2-*/out. Before moving
+# this on, check that llama-cpp-sys-2's get_cargo_target_dir() copes with the
+# new layout, and update that glob. .github/workflows/release.yml holds the
+# Windows build at the same version for the same reason.
+ARG RUST_VERSION=1.99
 
-FROM rust:1-trixie AS rust-toolchain
+FROM rust:${RUST_VERSION}-trixie AS rust-toolchain
 
 # --- Build environment, one per LLAMACPP value -----------------------------
 
@@ -194,7 +205,7 @@
 # The onnxruntime and llama.cpp shared libraries ship next to the binary (see
 # the build stage above); it has no rpath pointing here, so the loader needs
 # telling. Kept after the base image's own value (the CUDA images set one).
 ENV LD_LIBRARY_PATH=/app${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
 
 # Fail the build, not the first start: every library the server and its
 # llama.cpp modules link has to be in the image. All but the NVIDIA driver's

server-rs/README.md

@@ -649,8 +649,10 @@ the variable and answers only to the API call. See the doc comment on
 ## Docker
 
 `Dockerfile` here builds and ships the compiled binary (multi-stage:
-`rust:1-trixie` builder, `debian:trixie-slim` runtime; trixie because the
-ONNX Runtime libraries need glibc 2.38+). Model files live in the `/models`
+`rust:1.99-trixie` builder, `debian:trixie-slim` runtime; trixie because the
+ONNX Runtime libraries need glibc 2.38+, and Rust held at 1.99 because
+llama.cpp's build breaks on Cargo 1.100's build-dir layout — see the
+`RUST_VERSION` note in the Dockerfile). Model files live in the `/models`
 volume: the Dockerfile points every family's env var above there
 (`/models/siglip2/...`, `/models/face/...`, `/models/bioclip/...`,
 `/models/llm`), which is where the plug-in's *Download AI models* puts them —