[xiaozongyang]

Expose process_available_cpu_cores as an unlabeled gauge through WriteProcMetrics, WriteProcessMetrics, and WritePrometheus(w, true) so process CPU usage can be compared with its cgroup CPU bandwidth limit, including fractional cores (e.g. 150000 / 100000 reports 1.5).

The collector supports cgroup v1 and v2, resolves membership against mountinfo (including mount roots and escaped paths), and uses the smallest readable positive quota among visible ancestors. Quotas are read on each collection; mountinfo is cached after one bounded read. Unlimited, missing or invalid limits fall back to runtime.NumCPU(), as do non-Linux platforms. Collection does not change GOMAXPROCS.

This reports CPU quota, not CPU reservation or Go scheduler parallelism; ancestor quotas can be shared with sibling cgroups. It does not combine CPU affinity/cpuset restrictions with the quota, and cannot discover ancestors hidden by the cgroup namespace.

Validation:

• Tests cover v1/v2 fractional quotas, unlimited limits, ancestor limits, mount roots, hybrid hierarchies, malformed/missing data, quota changes, export opt-in/out, gauge metadata, and non-Linux fallback.
• go test -race ./... and go vet ./... pass on macOS.
• Linux amd64 and 386 test binaries cross-compile; Linux execution is left to CI because no local Linux container daemon is running.
• golangci-lint run --new-from-rev=origin/master passes. Default full lint reports pre-existing issues in unchanged files.

[cubic-dev-ai]

No issues found across 3 files

_{Re-trigger cubic}

[xiaozongyang]

hi @jiekun would you mind reviewing this PR?