docs: improve Linux tuning guidance

[visoar]

Why

The existing Linux tuning page recommends fixed kernel values, aggressive conntrack timeout reductions, and CentOS 6-era buffer settings without diagnostics, sizing evidence, or rollback steps. Those recommendations can expire valid flows, waste memory, or hide the actual bottleneck when copied into modern hosts and Kubernetes nodes.

What changed

• replace fixed maximum values with a measurement-led capacity and headroom method
• add diagnostics and rollback guidance for conntrack, file descriptors, listen queues, ephemeral ports, TIME-WAIT, keepalive, and socket memory
• distinguish broker, load-generator, NAT/firewall, container, and Kubernetes node limits
• cite upstream Linux kernel and Kubernetes documentation
• connect the guide to BifroMQ metrics, deployment, load-balancing, and benchmark documentation

Validation

• pnpm run check:site
  • production build passed
  • site policy checks passed for 124 indexable pages
  • long-title count remains at the existing 5
• pnpm run typecheck
• git diff --check

Addresses #72. Post-publication discovery/usage measurement remains tracked in the issue and should not block technical review of this change.

[visoar]

@popduke Could you please review this Linux tuning update?

The reason for the change is that the current page gives fixed kernel values, short conntrack timeouts, and CentOS 6-era buffer settings without showing how to diagnose the actual limit, size it for the deployment, or roll it back. Copying those values into a modern host or Kubernetes node can expire valid flows, increase memory use, or conceal an application/load-balancer bottleneck.

The proposed guide therefore moves to measurement-led sizing, separates broker and network-path limits, adds explicit rollback and reproducibility steps, and cites upstream kernel/Kubernetes documentation. In particular, review of the conntrack, file-descriptor, and Kubernetes guidance would be appreciated.