[Jordan Olshevski (jveski)]

Racer-dataplane's idle buffer pools (the flow recycler and the slab staging pool) now size themselves from recent demand instead of always holding up to 8 buffers per worker: they grow when requests miss a full pool and decay back down after 10 idle seconds, and the recycler caps idle retention at half of each class quota. A new cgroup v2 monitor watches memory usage against memory.high/memory.max and PSI, and on Elevated pressure shrinks pools, releases retained buffers and idle connections, and trims the glibc heap; on Critical pressure it also evicts idle cached pages. It can be disabled with RACER_MEMORY_PRESSURE=false. The binary also pins glibc's mmap threshold to 1 MiB so large freed buffers return to the OS, the operator now requests 1Gi of memory for the dataplane container (no default limit), and new metrics cover RSS, cgroup memory, pressure level, pool sizes, and per-class admission usage. The reference docs are corrected where they described per-worker defaults as node-wide.