Now with embedded finite observer, memory-based tracking, and intervention modes from the TRA paper!
Off: No observer tracking
Passive: Track largest component (no memory)
+ Memory: Track with persistence bias (prefers halo overlap)
+ Guided: Memory + interventions targeting halo area
+ Random: Memory + random interventions (same budget)
🔬 Key finding: Guided and Random have the same intervention budget, but only Guided maintains identity. This proves informational coupling, not action magnitude, sustains identity!