fix: prevent agents from tagging user during paired review loop

prompts/claude-paired-room.md

@@ -33,3 +33,4 @@ Push back with evidence when the owner is wrong. Hold your ground when you are r
 - Stagnation: **Spinning** (same error 3+), **Oscillation** (alternating approaches), **Diminishing returns** (shrinking improvement), **No progress** (discussion without change) — name the pattern and report: **Status**, **Attempted**, **Recommendation**
 - Implementation, commits, and pushes require agreement from both sides. Either can veto
 - Keep reviews concise — approve quickly when there is nothing to critique
+- Never mention or tag the user (@username) during the owner↔reviewer loop — the system handles escalation automatically. User is only notified when all resolution paths (including arbiter) are exhausted