feat: confidence-based Trust Score — more usage = higher potential score

Redesigned from penalty-based to confidence-based scoring.

Old problem: agents that work MORE get lower scores (more chances for errors).
New design: agents EARN trust through proven work over time.

Three layers:
  Proven Reliability (70%): Bayesian estimate (successes+5)/(total+10)
    - 0 interactions → 0.5 (unknown)
    - 100 interactions, 0 errors → 0.95 (proven)
    - 1000 interactions, 50 errors → 0.95 (still excellent, huge sample)
    - 10 interactions, 5 errors → 0.5 (concerning)

  Evidence Integrity (15%): chain ratio (continuous 0-1)

  Activity Confidence (15%): log10(interactions)/3
    - More data = higher confidence = higher score
    - Rewards volume instead of penalizing it

Key changes:
  - New agent starts at ~500 (unknown), not ~950 (fake high)
  - high_latency, hedge_rate, incomplete do NOT affect score
  - Only agent's own errors count
  - 1000 tasks with 95% success > 10 tasks with 100% success

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: 2 people