llmmsg-kpi-report.sh: large-ARO flag is base-rate noise + stale-corpus decoy risk (v1.15/v1.16 fixes)

Large-ARO flag fires on ~13% of all traffic (median 133.8/1000 msgs) — it describes normal, not an exception, and the daily report has been printing it as a drift signal the whole time. Retuning is pm-llmmsgsrv-cc's lane; v1.15 states the base rate instead of dropping the metric. v1.16 closed a decoy in code: whey's stale v2.sqlite sits exactly where the script's default path resolves, and would have emitted a complete, plausible, two-month-stale report with no signal of staleness. Fixed with two assertions (either alone fails open): a host pin, and a freshness check on the newest message. Verified four ways, including a corpus with 10 days deleted. Sharp point: improving the analysis made the stale-corpus risk WORSE, not better — percentiles over a frozen corpus look identical to percentiles over a live one, and a verdict lends a stale answer more authority than a bare count ever did.