…build

- Fix sqlite-vec compilation on Alpine/musl by guarding BSD type aliases with __GLIBC__
- Dockerfile: switch to CPU-only llama.cpp build (Vulkan shaders fail on Alpine)
- Dockerfile: add llama-go go.mod copy for replace directive support
- CI workflows: add GGUF model caching and download steps
- CI workflows: add llama.cpp build steps (CPU-only for tests, GPU for desktop releases)
- CI workflows: add LIBRARY_PATH/C_INCLUDE_PATH env vars for CGO linking
- ci-setup action: add Vulkan SDK and llama.cpp build per platform

Replace the vendored backend/util/llama-go directory (~1200 C/C++ files,
500K+ lines) with a git submodule pointing to seed-hypermedia/llama-go.

Changes:
- Remove vendored llama-go and add as git submodule
- Fix go.mod: use upstream tcpipuk/llama-go module path with replace
  directive pointing to ./backend/util/llama-go
- Update import in llamacpp.go to use upstream module path
- Add submodule init guard to .envrc (before mise activation)
- Add submodule existence check to mise.toml ensure-llama-libs task
- Remove sync_llama_go() and generate_gpu_build_files() from ./dev script
- Add submodules: recursive to 12 CI checkout steps across 10 workflows
- Fix wrapper.cpp in fork: use common_chat_parser_params matching pinned
  llama.cpp version (commit 2eee6c866)

- Use HTTPS URL in .gitmodules so cloning works without SSH keys
- Add ensure-submodule mise task to auto-init submodules
- Make ensure-llama-libs depend on ensure-submodule
- Move setup orchestration from mise enter hook (unreliable with
  direnv) to explicit mise run calls in .envrc
- Result: git clone + cd into repo does everything automatically

The llama-go submodule includes the full llama.cpp source tree (~2500
files, 148MB). The previous glob copied all of them into the Please
sandbox temp dir before building, causing massive disk I/O and memory
pressure that could freeze the machine.

Build in $WORKSPACE in-place (like seed-daemon already does) and copy
only the ~9 output .a files back to the sandbox. The Makefile is kept
as a src for change tracking.

Eliminate the SEED_CPU_ONLY / SEED_USE_GPU toggle that caused build
conflicts when ensure-llama-libs (CPU-only) and plz (GPU) built into
the same directory with different modes.

Now each platform always uses the same GPU mode everywhere:
- macOS: always Metal (built-in, zero deps)
- Linux: always CPU-only for local dev (no Vulkan packages needed)
- CI: handles per-platform GPU builds in ci-setup/action.yml

Changes:
- mise.toml: ensure-llama-libs detects OS and builds Metal on macOS,
  CPU-only on Linux. Detects stale CPU builds on macOS via missing
  libggml-blas.a and forces rebuild.
- backend/BUILD.plz: llama-cpp and seed-daemon genrules use OS
  detection instead of SEED_CPU_ONLY env var.
- dev: remove setup_gpu_build(), --cpu/--gpu flags from all commands.
- .plzconfig: remove SEED_USE_GPU/SEED_CPU_ONLY PassUnsafeEnv.
- Fork Makefile: add Metal mismatch detection alongside existing
  Vulkan detection in CMake cache checks.

The seed-daemon genrule's glob(**/*.c, **/*.h, **/*.cpp, **/*.hpp)
captures ~2500 files from the llama.cpp nested submodule. Please
hashes and copies all of them into the sandbox, causing 10+ minute
builds and extreme CPU/memory usage.

Exclude util/llama-go/llama.cpp/** since the seed-daemon genrule
only needs the compiled .a libraries (via :llama-cpp dependency),
not the C/C++ source files.

plz build takes 10+ minutes for seed-daemon due to sandbox overhead
(copying files, hashing dependencies). go build directly takes ~12s
from cold cache, ~3s incremental.

Replace plz build //backend:seed-daemon with direct go build in all
./dev commands (build-desktop, test-desktop, run-backend, build-backend).
The BUILD.plz genrule is still used by CI workflows.

Also fix build-backend which still referenced the removed
setup_gpu_build() function.

This reverts commit 34208d7.

…ove dead setup_gpu_build from dev script

The test waited for embedCalls==2 then immediately checked the DB,
but the INSERT transaction could still be in-flight. Now also waits
for runOnce to fully complete (task deleted from taskMgr) before
checking DB state.