DECOY DATASTORES: whey /var/lib/llmmsg-srv holds a 2-month-stale v2.sqlite + 2 zero-byte husks, all silent

FLEET TRAP, found by bin-whey-cc 2026-08-23. /var/lib/llmmsg-srv/ exists on BOTH venus and whey with the same filenames, but only venus is live. On WHEY: v2.sqlite is 66MB dated 2026-06-15 (TWO MONTHS STALE), llmmsg.sqlite is ZERO BYTES (2026-06-05), maintainer-agents.sqlite is ZERO BYTES (2026-07-21). All three are readable and none errors. Any job pointed at these paths on the wrong host gets a large, plausible, WRONG dataset silently - this directly threatens the llmmsg-kpi-report baselining work (KPI-26), which must compute percentile bands over the LIVE 71-day corpus (venus v2.sqlite, 90566 msgs, 2026-06-13 -> 2026-08-23). RULE: pin the corpus by HOST, never by path; fleet policy already forbids cross-host DB queries, which is exactly why the same path is a different database per host. Third instance of retired-but-present-reads-as-live in one day: usage.daily held rows past its retired writer; kpi-web served HTTP 200 over a telemetry_ro that was permission-denied on every view; these husks. Verify a datastore by mtime + max(ts) against what the writer should be producing NOW, and state which host you read.