- Introduce LlmPoolConfig for managing summary, retrieval, and pilot
  client configurations
- Add comprehensive retry, throttle, and fallback mechanisms
- Implement unified metrics configuration with LlmMetricsConfig,
  PilotMetricsConfig, and RetrievalMetricsConfig
- Add cost calculation capabilities for LLM operations
- Consolidate existing LLM-related config types into new modules
- Update main Config struct to include llm and metrics fields

feat(metrics): implement central metrics hub for unified collection

- Create MetricsHub for collecting LLM, Pilot, and Retrieval metrics
- Add thread-safe atomic metric tracking for multi-threaded usage
- Implement comprehensive reporting capabilities
- Add support for token tracking, latency measurement, and cost
  estimation
- Include metrics for rate limits, timeouts, and fallbacks

Add query_hash and path_hash methods to PilotContext for generating
hashes used in feedback learning system.

Introduce feedback learning capabilities to LlmPilot including:
- Add optional PilotLearner field to LlmPilot struct
- Implement with_learner and with_feedback_store builder methods
- Add helper methods has_learner, learner, and record_feedback
- Integrate learner adjustment logic in decision making process
- Apply confidence adjustments based on historical feedback data
- Skip interventions when learner suggests low historical accuracy

The feedback system allows the pilot to learn from past decisions
and improve future recommendations through confidence adjustments
and strategic intervention skipping.

- Add comprehensive design document for Feedback Learning system
  including architecture, core components, and integration details
- Implement detailed documentation covering FeedbackRecord,
  FeedbackStore, and PilotLearner components
- Add usage examples demonstrating feedback collection and learning
- Include system diagrams showing data flow and component interactions
- Document configuration options and learning strategies
- Provide example code showcasing feedback learning implementation

…arning example

- Remove unused DocumentTree import from vectorless::document module
- Reorganize imports for better grouping and readability
- Move PilotConfig import to appropriate location with other pilot-related imports
- Rename persistent_store variable to _persistent_store to indicate it's intentionally unused

Remove the star call-to-action message and update the early development
warning by removing the ⭐ emoji prefix.

docs(architecture): enhance diagram with comprehensive system design

Update architecture.svg to include:
- Expand canvas size from 720px to 800px height
- Add detailed configuration section with TOML settings
- Include LLM executor with throttle control and retry mechanisms
- Add query decomposition and pilot guidance components
- Implement unified metrics hub with LLM, pilot, retrieval, and
feedback statistics
- Add feedback learning loop from user feedback to decision adjustment
- Include design philosophy section highlighting zero vectors,
algorithm+LLM hybrid approach, feedback learning, and multi-turn
support
- Replace keyword/semantic/LLM strategies with keyword-only, BM25,
and hybrid scoring strategies
- Update workspace to include feedback store component