v0.5.2 - Massive I/O & Concurrency Optimizations 🚀

This release introduces a comprehensive overhaul of the codebase's core indexer, memory handling, and database interaction layers. These changes dramatically reduce indexing time and fix hidden bottlenecks that caused Tokio thread starvation.

⚡ Concurrency & Threading

• Fixed Tokio Runtime Starvation: Moved heavy AI inference (Qwen3/BERT embeddings) out of the main async executor using spawn_blocking, eliminating server freezing.
• Removed Async Inference Locks: Extracted the embedding engine behind an Arc clone, allowing the RwLock read guard to be dropped immediately before CPU-bound inference starts.
• Parallel AST Parsing: The indexer.rs hot loop now uses a bounded JoinSet::spawn_blocking (max 4 concurrent workers). Files are parsed and hashed in parallel, while database writes remain safely sequential.

💽 Disk I/O & Memory Optimization

• fsync Coalescing in Redb: Implemented put_batch() in the EmbeddingStore. Instead of calling fsync() thousands of times per file, embedding disk caches are now flushed in large, single-transaction blocks (100x disk I/O reduction).
• BM25 Memory Bloat Fixed: Added OMIT embedding to SurrealDB chunk fetching queries. BM25 search no longer loads thousands of unused 768-dimensional float arrays into RAM.
• Zero-Copy Hot Loop: Replaced massive content.clone() allocations with move semantics in the parallel AST parser.

🗄️ SurrealDB Query Batching

• Eliminated N+1 DB Queries (Relations): create_symbol_relations now collects all links in-memory and pushes them to the DB using a single batched FOR $r IN $relations { RELATE ... } statement.
• Eliminated N+1 DB Queries (Subgraphs): Replaced sequential db.select() inside for loops with a massive SELECT * FROM id1, id2... query.
• True Symbol Batching: "Fake" batched upserts in create_code_symbols_batch were rewritten into a native INSERT INTO code_symbols ... ON DUPLICATE KEY UPDATE single query.
• HNSW Vector Index Acceleration: Refactored vector::similarity::cosine vector searches to use the <|K,EF|> KNN operator, activating SurrealDB's internal ANN indexes instead of relying on brute-force O(N) table scans.