SIGABRT crash on exit during Metal cleanup (macOS ARM64) #59

Environment

• macOS ARM64 (Apple Silicon)
• bun 1.1.45
• node-llama-cpp 3.15.1

Issue

qmd query crashes with SIGABRT during process exit after successfully completing reranking:

GGML_ASSERT([rsets->data count] == 0) failed

The reranker works correctly — results are output before the crash. The crash happens during Metal device cleanup in llama.cpp's ggml_metal_device_free().

Reproduction

qmd query "winback" -n 3
# Results print successfully, then crash on exit

Root cause

This is an upstream llama.cpp bug in Metal cleanup. Reference: ggml-org/llama.cpp#17869

Workaround

Redirect stderr and force exit 0:

qmd query "$@" 2>/dev/null; exit 0

Impact

• CLI usage: cosmetic (results complete before crash)
• MCP server: may affect long-running server stability

Would be great to either:

1. Update to a llama.cpp version with the fix
2. Add explicit resource cleanup before exit to avoid the race