fix(prefill): keep MLX worker compute thread-affine

[FluffyAIcode]

Summary

• construct and warm factory-backed prefill engines on the same dedicated thread that executes jobs
• keep one MLX model per worker process and reject multi-threaded MLX job configurations
• preserve existing injected engines for unit/service users
• chmod generated LaunchAgent plists to 0644 before bootstrap so upgrades do not fail with launchd error 5

Root cause

The first real two-Mac request reached allens (remote_jobs +1) but MLX 0.31 rejected inference because the model/stream was created on the asyncio main thread and used by ThreadPoolExecutor: There is no Stream(gpu, 1) in current thread.

Tests

• focused worker/orchestrator/deploy tests — 13 passed
• platform-neutral CI selection — 816 passed, 1 skipped
• regression proves factory construction and compute use the same non-main thread
• Python compile, shell syntax, and diff checks pass

Live verification after merge

Redeploy allens worker, then rerun scripts/verify_remote_prefill_e2e.py; acceptance requires remote_jobs, remote_hits, and tokens_reused to increase with no fallback.

Made with Cursor