PyTorch caps clang-tidy parallelism to prevent CI OOM crashes

PyTorch's CI team resolved a persistent out-of-memory (OOM) issue affecting clang-tidy and clang-format linting jobs on Kubernetes ARC runners. The root cause: the clang adapters used os.cpu_count() to determine parallelism, but on k8s pods this returns the host node's physical core count — often significantly higher than the pod's allocated CPU limit. For an "--all-files" run, this resulted in far more memory-hungry clang-tidy processes being launched than the pod could handle, leading to OOM kills and broken CI.

To fix it, they introduced a -j/--num-workers flag for the adapters that defaults to the MAX_JOBS environment variable (with os.cpu_count() as fallback). In the clang lint job, they set MAX_JOBS="$(nproc --ignore=2)" — exactly as _linux-build.yml does — so nproc correctly reports the pod's cgroup-constrained CPU budget. This caps the number of concurrent clang-tidy processes to a safe limit, preventing memory exhaustion. The patch (PR #187535, approved by @atalman and @huydhn) was authored with assistance from Claude Code.