v1.0.35 - Clean Up Before She Comes

"Clean Up Before She Comes" - Is a maintenance release to update all core dependencies including llama.cpp

[v1.0.35] - 2026-03-27

Changed

• llama.cpp roll-forward and runtime alignment:

  • upgraded the vendored llama.cpp integration and shipped local llama artifacts to b8547.
  • synchronized version pins across build scripts, packaged libraries, and Docker builder images so local, CI, and container builds use the same llama baseline.

• Container and CI dependency maintenance:

  • rolled forward Docker build dependencies and CI image references used for CPU, CUDA, Vulkan, and Metal packaging.
  • updated the bundled Heimdall and BGE image variants so embeddings are enabled by default in the batteries-included deployment paths.

• UI and build-tool dependency refresh:

  • updated frontend/package dependencies and related build configuration used by the shipped UI bundle.
  • refreshed release/build assumptions tied to current toolchain behavior so maintenance packaging stays reproducible.

Fixed

• Bolt autocommit database executor reuse:

  • fixed the Bolt multi-database autocommit path to reuse cached database-scoped executors instead of rebuilding a fresh Cypher executor on each RUN.
  • preserved auth-forwarded remote routing behavior by keeping auth-scoped database executor resolution uncached.

• llama.cpp upgrade compatibility on macOS and Heimdall generation paths:

  • fixed the local llama generation context initialization used by Heimdall after the llama.cpp upgrade, eliminating a macOS CGO crash during generation-model loading.
  • added generation-context normalization so native model initialization stays within safe runtime bounds after dependency roll-forwards.

• llama build portability and packaging behavior:

  • fixed scripts/build-llama.sh to remain compatible with macOS Bash 3.2.
  • updated the ARM64 Metal Heimdall image to prefer locally available models before falling back to network downloads, improving reproducibility in restricted build environments.

Tests

• Added and updated regression coverage for:
  • Bolt database-scoped executor cache reuse for non-auth paths and cache bypass for auth-forwarded routing
  • llama generation context resolution and safe initialization behavior after the llama.cpp upgrade
  • Heimdall CGO generation-model loading on the upgraded llama runtime

Technical Details

• Range covered: v1.0.34..HEAD
• Commits in range: 5 (non-merge)
• Repository delta: 44 files changed, +1,635 / -1,984 lines
• Primary focus areas: dependency roll-forwards, llama.cpp packaging/runtime compatibility, and release engineering maintenance.