Decision caching (LRU / content-hash based)

[temp-noob]

Why

Every tool call currently goes through the full evaluation pipeline including potentially an HTTP round-trip to an LLM provider. For agents making 50-200 tool calls per task, many calls are repeated or near-identical (e.g., reading the same set of files).

What

Cache semantic evaluation results to avoid redundant LLM calls.

Acceptance Criteria

• Cache keyed by content hash of (tool_name, sorted arguments, task_context)
• LRU cache with configurable max size (default: 256)
• Configurable TTL (default: 300 seconds)
• Static checks always run (fast, no caching needed)
• Only semantic evaluation results are cached
• Cache hit/miss logged in audit decision
• Cache stats available via metrics (if Add OpenTelemetry spans and decision/provider metrics #4 OTEL is implemented)
• Tests: cache hit, cache miss, TTL expiry, LRU eviction

Related

Partially addresses #23 (latency budget).