(in prog): add factory, adaptors for indexing strategies

[allen-munsch]

• in the style of vring, explores quantized vector slice as directory

Core Purpose:

The PR introduces a pluggable indexing strategy system that allows switching between different indexing methods, enabling:

1. Existing centroid-based strategy (shard-based with bloom filters)
2. New quantized ring-based strategy (grid/space partitioning)

Key Changes:

1. Strategy Pattern Architecture

• Created Mosaic.Index.Strategy behaviour defining a common interface
• Implemented two strategies:
  • Mosaic.Index.Strategy.Centroid - Wraps existing shard router logic
  • Mosaic.Index.Strategy.Quantized - New ring-based spatial partitioning

2. Quantized Ring Implementation

• Quantization: Converts embeddings to grid coordinates via PathEncoder
• Cells: Manages SQLite databases for each grid cell (Mosaic.Index.Quantized.Cell)
• Cell Registry: Tracks and manages cell instances
• Path Encoding: Uses multi-level binning of embedding dimensions

3. Configuration & Runtime Switching

• Added index_strategy config option ("centroid" or "quantized")
• Dynamic strategy selection at application startup
• Index supervisor that starts the chosen strategy

4. Migration Tools

• Created Mosaic.Index.Migrator for migrating between strategies
• Provides verification and switching capabilities

5. Query Engine Integration

• Modified Mosaic.QueryEngine to use the active strategy
• Added index_state to maintain strategy-specific state
• Updated grounding logic to work with both strategies

6. Refactored Existing Components

• Mosaic.Indexer: Simplified, moved centroid calculation to separate module
• Mosaic.ShardRouter: Implemented new router behaviour interface
• Application Supervision: Replaced direct components with strategy supervisor

Technical Details of Quantized Strategy:

How it works:

1. Vector Normalization: Embeddings are normalized per dimension
2. Multi-level Binning: Each level bins dims_per_level dimensions into bins buckets
3. Path Creation: Creates directory-like paths (e.g., "012/045/123")
4. Cell Management: Each unique path gets a dedicated SQLite database
5. Neighbor Expansion: Queries search within search_radius of target cells

Benefits:

• Scalability: Distributes data across many small databases
• Efficiency: Reduces search space via spatial partitioning
• Parallelism: Cells can be searched independently
• Flexibility: Tunable via bins, dims_per_level, search_radius

What's "in prog" (in progress):

1. Experimental feature: Still being refined and tested
2. Performance exploration: Comparing centroid vs quantized approaches
3. Migration readiness: Building tools to switch between strategies
4. quantized ring-based strategy: Still needs metadata to be wired up

Impact:

• Backward Compatibility: Existing centroid strategy remains default
• Extensibility: New strategies can be added by implementing the behaviour
• Operational Flexibility: Can A/B test strategies or migrate based on workload
• Research/Exploration: Enables experimenting with novel indexing approaches