- global_stats() executed the 'last N seconds' COUNT twice per page
load (prepare()->execute() result was discarded, then re-run)
- fetch_upstream_version() cached nothing on failure, so an
unreachable git server added a cURL wait (up to 5 s) to every
dashboard load via the 30 s auto-refresh; failures are now cached
for 15 minutes and get a 2 s connect timeout