feat: persist OS tuning across reboots (sysctl.d + boot tuning unit)

[emrebulutlar]

Closes #48 (P1.4, v0.3.0-beta).

Problem

system-tuning.sh / make optimize-os wrote runtime sysctls only. After the 2026-07-03 reboot, net.core.rmem_max/wmem_max reverted to 212992, the C media drivers crash-looped on SO_RCVBUF validation, the process manager kept starting nodes against absent drivers, and the kill-mid-write cycle corrupted node0's archive segment (the #35 replay hot-loop). Every reboot recreates this unless someone remembers to re-tune.

Changes

• system-tuning.sh --persist (new make tune-persist):
  • writes the full sysctl set (transport-critical + the broader optimize-os set) to /etc/sysctl.d/99-openexchange.conf, commenting out keys absent on the running kernel (e.g. the pre-EEVDF kernel.sched_* knobs on 6.x), and applies it immediately so runtime == boot state
  • installs + enables openexchange-tuning.service, a oneshot boot unit for the two non-sysctl knobs: THP=never and performance governor
  • idempotent: skips when file/unit already match; degrades to WARN without sudo
• --report drift detection: shows each persisted key that diverges from the runtime value, plus the boot unit's enablement, so make tune-report catches kernel upgrades / manual changes
• make optimize-os now runs the persistence step at the end
• Runbook (.claude/CLAUDE.md) + production checklist (docs/kernel-bypass.md) updated

Verification

• Conf generation verified against the live 6.17 kernel (sched knobs correctly commented out, tcp values normalized)
• Drift detection verified with a fixture file (value mismatch + unknown key both flagged; tab-vs-space normalization for tcp_rmem confirmed)
• No-sudo path degrades to WARN, exit 0
• Root-only steps (install to /etc, systemctl enable) need a privileged run on the box: sudo make tune-persist. Exit criterion (clean reboot with zero manual steps) verifiable at the next reboot via make tune-report.

🤖 Generated with Claude Code