[mosobande]

Adds references/practices/ci-runner-hosts.md, linked from the Runners section of ci.md. The rules come from reviewing a week of jobs on a small 2-core self-hosted runner host:

• Check when build cache was last used before pruning it. Docker's "reclaimable" figure includes cache used minutes ago.
• Run all of a host's runners as one service user. A runner per user duplicates toolchains, package stores and Docker.
• Install toolchains once on the host through mise. Keep one pnpm store on the same filesystem as the work folders, and skip actions/cache there.
• Send jobs under about a minute to a light runner by label, and fold tiny jobs into their neighbours.
• Use shallow checkouts, and sparse checkouts for jobs that read a few paths.
• Don't run more jobs than a small, already-swapping host can hold.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}