Add performance features: caching, cost tracking, retry, compaction, classification, scrubbing

Inspired by zeroclaw's lightweight patterns for slow hardware:
- Response cache (SQLite + SHA-256 keyed) to skip redundant LLM calls
- History compaction — LLM-summarize old messages when history exceeds 50
- Query classifier routes simple/research queries to cheaper models
- Credential scrubbing removes secrets from tool output before sending to LLM
- Cost tracker with daily/monthly budget enforcement (SQLite)
- Resilient provider with retry + exponential backoff + fallback provider
- Approval engine gains session "always allow" and audit log

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

config.yaml

@@ -43,3 +43,12 @@ agents:
 orchestrator:
   max_concurrent: 5
   delegation_timeout: 120
+
+performance:
+  cache_ttl: 3600
+  daily_budget_usd: 5.0
+  monthly_budget_usd: 50.0
+  fallback_model: nvidia_nim/deepseek-ai/deepseek-v3.1
+  model_routing:
+    simple: nvidia_nim/deepseek-ai/deepseek-v3.1
+    research: nvidia_nim/moonshotai/kimi-k2.5