, #2300)

Add PruningCache to agent loop state, keyed on (message_content_hash,
tool_list_hash), to avoid redundant LLM calls when the tool list and
user message have not changed between agent loop iterations.

- PruningCache with CachedResult::Ok/Failed enum; single-slot, ~48 bytes
  on stack plus at most max_tools McpTool clones (~3 KB) on hit
- prune_tools_cached() wrapper; negative caching prevents LLM retry
  storms on persistent pruning failures
- tool_list_hash hashes full metadata (name, description, input_schema)
  sorted by qualified name using BTreeMap-backed serde_json::to_vec;
  sentinel byte on schema serialization failure
- Cache wired into rebuild_system_prompt (once per turn); apply_pruned_
  mcp_tools updates shared_tools RwLock without calling sync_mcp_executor_
  tools; both write sites carry cross-referencing ordering contract comments
- Cache reset at: process_user_message_inner start, check_tool_refresh,
  handle_mcp_add, handle_mcp_remove
- Err branch of pruning match in rebuild_system_prompt calls
  sync_mcp_executor_tools to self-heal shared_tools on pruning failure

Add 13 integration tests covering: always_include partition (including
cross-server bare-name matching), max_tools cap, min_tools_to_prune
early-return (replaces tautology), LLM failure propagation, parse error,
positive cache hit, cache miss on message/tool-list change, negative
cache hit (skips LLM on second call), and reset-clears-negative.

Closes #2298, #2300