ELX-30: Bundle maintenance runbook + staleness early warning

[cdevarenne]

A canonical layer nobody refreshes is worse than none: it answers exact questions with
confident stale facts, which is the failure this project exists to prevent.

The runbook is written — see docs/maintenance.md: re-verifying a concept against its
source, moving stale_after, regenerating the compatibility matrix, syncing the JVM repo's
copy of knowledge/, and refreshing the corpus.

What is still open is the early warning. is_stale is evaluated per answer, so the first
notice you get is a citation block already printing STALE. Correct behaviour, late notice.

Done when

• a scan over knowledge/ reports concepts already past stale_after
• --as-of answers "how many will be stale on a future date", so a governance cliff is
  visible before it arrives
• it runs with no cluster, and projects to Elasticsearch only when configured
• a scheduled run is documented, even if not built

Depends on the corpus-state snapshot work.

[cdevarenne]

An initial version of the runbook is available now as mentioned in the description above: docs/maintenance.md
The early warning is next.