[Backport release-mainnet-1.2.0-rc] fix(process-metrics): build on non-Linux platforms

[github-actions]

Description

Backport of #4435 to release-mainnet-1.2.0-rc.

[claude]

Claude finished @sveitser's task in 1m 44s —— View job

---

Code Review: Backport fix(process-metrics) build on non-Linux platforms

• Read repository guidelines
• Read existing review comments
• Analyze PR diff against base branch
• Review changed files in detail
• Post review findings

Summary

This is a clean backport of #4435 to release-mainnet-1.2.0-rc. The change refactors the process-metrics crate to compile on non-Linux platforms by splitting the single-file implementation into three modules:

File	Purpose
lib.rs	Module root with cfg-gated ext module — Linux gets linux.rs, others get an inline no-op stub
portable.rs	Cross-platform metrics (RSS, virtual memory, uptime, CPU count) via sysinfo
linux.rs	Linux-only metrics (/proc, cgroup v2, PSI pressure) via procfs

procfs is moved to a cfg(target_os = "linux")-only dependency in Cargo.toml, which is the core fix.

Findings

The refactor is well-structured and the backport is clean. The public API (ProcessMetrics::new() / ProcessMetrics::run()) is unchanged, making this a safe backport for a release branch. A few observations:

Bug fix: cpu_count now actually works

The original code called sysinfo::System::new().cpus().len(), but System::new() without refresh_cpu_all() returns 0 CPUs — so node_cpu_count was always 0. The new std::thread::available_parallelism() approach in portable.rs:69-73 correctly reports the available CPU count. The test at portable.rs:90 documents this by asserting System::new().cpus().len() == 0.

Clean cfg architecture

The #[path = "linux.rs"] + #[cfg(target_os = "linux")] pattern in lib.rs:11-13 paired with the inline stub at lib.rs:17-34 keeps all platform cfg in one place. The stub's interface (new, init, sample) mirrors the real LinuxMetrics exactly, so portable.rs compiles identically on all platforms.

State ownership refactor

LinuxMetrics now owns env: Option<Env> and prev: Previous as fields (linux.rs:139-140) rather than passing them through the call chain. This is cleaner and makes the init()/sample() split natural — init() detects sources once, sample() early-returns if env is None.

Graceful degradation on non-Linux

On non-Linux, the stub LinuxMetrics emits a single info-level log during init() and sample() is a no-op. The portable metrics (RSS, virtual memory, uptime, CPU count) are still collected. This is the right behavior — operators get what the platform can provide, with clear logging about what's absent.

Verdict

LGTM. The refactor is clean, the API is unchanged, tests are preserved and enhanced (new cpu_count_is_at_least_one and process_memory_reports_current_process tests in portable.rs), and the cfg architecture is well-contained. No issues found.