perf(runtime): Hermes Phase 1-3 — prompt caching + parallel tools + smart retry

Phase 1: Anthropic prompt caching
- Add cache_control ephemeral on system prompt blocks
- Track cache_creation/cache_read tokens in CompletionResponse + StreamChunk

Phase 2A: Parallel tool execution
- Add ToolConcurrency enum (ReadOnly/Exclusive/Interactive)
- JoinSet + Semaphore(3) for bounded parallel tool calls
- 7 tools annotated with correct concurrency level
- AtomicU32 for lock-free failure tracking in ToolErrorMiddleware

Phase 2B: Tool output pruning
- prune_tool_outputs() trims old ToolResult > 2000 chars to 500 chars
- Integrated into CompactionMiddleware before token estimation

Phase 3: Error classification + smart retry
- LlmErrorKind + ClassifiedLlmError for structured error mapping
- RetryDriver decorator with jittered exponential backoff
- Kernel wraps all LLM calls with RetryDriver
- CONTEXT_OVERFLOW recovery triggers emergency compaction in loop_runner

crates/zclaw-runtime/src/driver/mod.rs

@@ -15,11 +15,14 @@ mod anthropic;
 mod openai;
 mod gemini;
 mod local;
+mod error_classifier;
+mod retry_driver;
 
 pub use anthropic::AnthropicDriver;
 pub use openai::OpenAiDriver;
 pub use gemini::GeminiDriver;
 pub use local::LocalDriver;
+pub use retry_driver::{RetryDriver, RetryConfig};
 
 /// LLM Driver trait - unified interface for all providers
 #[async_trait]
@@ -106,6 +109,12 @@ pub struct CompletionResponse {
     pub output_tokens: u32,
     /// Stop reason
     pub stop_reason: StopReason,
+    /// Cache creation input tokens (Anthropic prompt caching)
+    #[serde(default)]
+    pub cache_creation_input_tokens: Option<u32>,
+    /// Cache read input tokens (Anthropic prompt caching)
+    #[serde(default)]
+    pub cache_read_input_tokens: Option<u32>,
 }
 
 /// LLM driver response content block (subset of canonical zclaw_types::ContentBlock).