fix: prevent macOS Metal JSON query aborts after successful output

[sk-holmes]

Problem

On macOS Metal, qmd query --json and qmd deep-search --json can produce valid JSON and then abort during native llama.cpp cleanup. That leaves callers with a non-zero exit even though the command already completed successfully.

This mainly breaks machine-readable JSON workflows. The output is valid, but the process lifecycle reports failure.

Fix

This adds a macOS-only supervised exit path for GPU-backed JSON query and deep-search runs.

The parent process runs the query in a child process and accepts success only when both conditions are true:

• stdout parses as valid JSON
• the child writes a success sentinel on a side-channel fd after output is flushed

The child then terminates before native Metal finalizers can abort after successful output. Real query failures still return through the normal failure path.

I also tightened the GPU env handling so the supervisor only runs when GPU use is actually enabled, and the CLI reuses the same GPU-disabled semantics as the LLM layer.

Verification

• node scripts/build.mjs
• node ./node_modules/typescript/bin/tsc -p tsconfig.build.json --noEmit
• full Node/Vitest suite: 24 files passed, 824 passed, 73 skipped
• real macOS fixture checks with QMD_FORCE_CPU=0 QMD_LLAMA_GPU=metal:
  • qmd query "metal cleanup" --json -c docs
  • qmd deep-search "metal cleanup" --json -c docs

Notes

The workaround is intentionally narrow: Darwin only, JSON only, query/deep-search only, non-child process only, and disabled for explicit CPU/no-GPU runs.