Each path test opens its own device and compiles the path pipeline. On a cold driver cache that takes minutes and gigabytes per compile, and libtest ran one per CPU thread side by side: on NVIDIA the solidcraft-gpu tests ran for hours and committed ~80 GB. Executor::new now holds a process-wide lock while it compiles; the first compile fills the driver's cache and the waiting ones get the pipeline from there. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>