- Add background cleanup thread for periodic garbage collection
  - Configurable via CleanupConfig (interval, retention periods)
  - Cleans deleted rows, old transactions, and previous versions
  - Auto-starts on Database::open(), stops on close()

- Implement arena slot reuse to prevent unbounded memory growth
  - Free list tracks cleared slots for reuse on insert
  - Batch clearing via clear_batch() for efficient cleanup

- Add Index trait _into methods to avoid allocations
  - get_row_ids_equal_into, get_row_ids_in_range_into, get_row_ids_in_into
  - Callers can reuse Vec buffers across multiple calls

- Convert rows to Arc storage immediately in TransactionVersionStore::put()
  - Enables efficient Arc cloning during commit

- Add rollback_all_tables() to clean up txn_version_stores on rollback
  - Prevents memory leaks when transactions are rolled back

- Fix Database::Drop to only remove from registry when last reference
  - Uses Arc::ptr_eq and strong_count check to avoid breaking clones

…zations

- Add committed_row_count atomic counter to VersionStore for O(1) COUNT(*)
  - Updated add_version, add_version_single, add_versions_batch to track counter
  - +1 for INSERT, -1 for DELETE, handles state transitions correctly

- Add get_distinct_count_excluding_null() to Index trait for O(1) COUNT(DISTINCT)
  - Implemented in BTreeIndex and HashIndex
  - Avoids cloning all values just to count them

- Add compiled query cache for COUNT(*) and COUNT(DISTINCT) queries
  - CompiledCountStar and CompiledCountDistinct in query_cache.rs
  - Schema epoch validation for cache invalidation
  - Properly handles FILTER clause (bypasses fast path)

- Introduce RowIdVec pooled vector for index lookups
  - Thread-local buffer pooling with best-fit allocation
  - 256K max cached capacity to prevent unbounded memory
  - Fixed broken IntoIter Drop that couldn't recover buffer

- BTree index optimization: use Borrow trait for lookups
  - No allocation needed for equality checks
  - Only create CompactArc for range operations

- Clean up dead code in try_index_lookup OR expression handling

Performance: COUNT(*) ~22.5µs → ~2.9µs (7.7x faster)

- Add clear_version_map_pools() for TransactionVersions pools
- Add clear_program_cache() for expression bytecode LRU cache
- Add clear_classification_cache() for query classification LRU cache
- Call all cleanup functions from clear_all_thread_local_caches()

Memory at program end: 665 KB → 153.6 KB (77% reduction)

…fficiency (#10)

Major optimizations:

**SmartString** - Custom SSO string replacing CompactString
- Inline storage for strings ≤22 bytes (no heap allocation)
- Owned (Box<str>) for computed values, Shared (Arc<str>) for cloned values
- 24-byte size with O(1) clone for shared variant

**I64Map** - High-performance hashmap for i64 keys
- Uses i64::MIN as empty sentinel (row/txn IDs are always ≥0)
- FxHash-based with linear probing and backward-shift deletion
- ~45% faster lookups than FxHashMap<i64, V>

**Value Interning** - Shared instances for common values
- Interned: NULL (all 7 DataTypes), booleans, integers 0-1000
- Reduces allocations for frequently used values in rows

**Aggregation Fixes & Optimizations**
- Fixed hash collision bug in single-column GROUP BY (was using hash as key)
- Added 2-column tuple optimization (30% faster than Vec<Value>)
- Single-column primitive GROUP BY uses I64Map directly

**SIMD Pattern Matching**
- memchr::memmem for LIKE '%pattern%' substring search
- Pre-compiled Finder stored in CompiledPattern

**Other Improvements**
- CompactVec: new insert(), retain(), drain() methods
- All indexes migrated to I64Map and CompactVec
- Thorough SAFETY documentation on unsafe code