perf(daemon): arm the processed-map prune timer only while it has entries

[leinardi]

Summary

The coordinator pruned its processed map on a 60 s ticker for the daemon's whole life. The map only gets entries during a pass (startup, SIGHUP, systemd reload, reconnect), and they expire processedTTL (60 s) later. So on a quiet node the daemon woke every minute to scan an empty map.

The prune timer is now a one-shot time.After, armed by armPrune only while the map has entries, and kept (not restarted) once armed. After a pass it fires once or twice, then stays off until the next pass.

The saving is small: about 1 of the daemon's ~16 wakeups per minute at idle, measured on a production node. The daemon now only wakes at idle for its own lifecycle sweep, the Go runtime and incoming events.

• Tests: the new TestCoordinator_PruneArmedOnlyWithEntries covers four cases:
  • no timer is armed for an empty map;
  • a timer is armed once an entry exists;
  • an armed timer is kept;
  • no timer is re-armed after a prune empties the map.
• Docs: docs/architecture.md notes that the timer is armed only while needed.
• Release: perf bumps no version by itself; it ships with the next release (1.0.1, via fix(processor): log symlinks leading outside /dev at debug unless allowed #27).

Pull request checklist

• I am targeting the master branch
• I have rebased this branch on top of the destination branch
• I have executed make check locally before creating the commit and it has run successfully
• I have performed a self-review of my own code
• There are no WIP commits in this PR

Type of changes

• 🐛 Bug fix
• ✨ New feature
• 🔧 Refactoring
• 📜 Docs
• 🧰 CI / tooling / infra
• Other (describe in Summary): performance, fewer idle wakeups