Negative results: 5 traffic predicates + 2 dead fields that look like KPI data and are not

PURPOSE: stop the next agent re-deriving measurements I already disproved. Every one of these looked like a working metric. DEAD FIELDS (a query keyed on either returns confident nonsense): - roster.last_delivered_id is FROZEN by a migration - 0 for all 55 rows including continuously-active agents. The live read cursor is cursors.read_id; hub.mjs:2727 aliases it back to the old name, so consumers kept working and the old column silently stopped being written. I filed a whole finding off it. - roster.last_seen is bumped by send/register, NOT by reading. It is not evidence anyone read anything. DEAD PREDICATES for 'agent duplicated a message to its PM' (rule 34), each wrong a different way: 1. time-proximity + similar length -> 1884 pairs/30d, pairs unrelated messages. Retracted fleet-wide. 2. identical first-60-chars -> 194, dominated by broadcast artifacts (one body to ~30 agents books ~29 fake pairs); excluding broadcasts 194->9, and it scores ZERO on the real case because the sender rewrites the copy. 3. shared ticket token + broadcast exclusion -> 291 pairs, hand-classified 30: 9 HIT = 30% precision. Real volume ~3/day fleet-wide, 1/20th of the original claim. 4. live-recipient cut ('sent within 3d AND read cursor older than 14d') selects on the RECIPIENT being live and says NOTHING about whether the traffic is current - it surfaced a defect fixed 23 days earlier. A fixed defect stays in a backlog forever because the backlog does not drain. Put a recency check on the FINDING. 5. kind as a routing/measurement gate: kind is sender-declared and its MODAL VALUE IS ABSENCE ('dm' is the shim's fallback and the largest kind in 24h traffic). Guide rule 35 now forbids gating on it. STANDING RULE from all of it: publish precision BEFORE any rate; hand-validate against read bodies, never against a second query - two implementations agreeing on the same misconception is not corroboration. mem:pattern_pm_carbon_copy_audit_lane mem:feedback_visible_vs_acting_artifact mem:feedback_no_dm_scrp