Use CPU pressure and process ancestry before blaming throttling

When agent work feels globally slow, first distinguish local runnable-queue contention from provider or API throttling. Sample CPU idle and load, Linux pressure-stall information, memory and I/O pressure, hypervisor steal time, cgroup CPU statistics, and full ancestry of top CPU consumers.

Sustained high CPU pressure with near-zero idle, no cgroup throttling, and no steal time points toward local oversubscription; process parentage identifies the responsible task or hook. This does not rule out service latency when the host is uncongested, and account limits require separate authoritative evidence.