- JSON_TYPE now accepts optional path argument: JSON_TYPE(json, '$.path')
- JSON_TYPEOF updated to support same signature
- Added JSON_OBJECT, JSON_ARRAY, JSON_ARRAY_LENGTH to documentation
- Fixed incorrect limitations section (these functions already exist)
- Updated introduction and best practices

- Add tests for JSON_TYPE with two-argument form (path extraction)
- Add tests for JSON_TYPE with path on table columns
- Add tests for JSON_TYPEOF with two arguments
- Add tests for JSON_TYPE with NULL path argument
- Add tests for JSON_OBJECT function
- Add tests for JSON_ARRAY function
- Add tests for JSON_ARRAY_LENGTH function
- Add tests for JSON_ARRAY_LENGTH with path (two arguments)

Improves test coverage for codecov patch check.

- Replace per-query cache clearing with table-based invalidation for
  scalar and IN subquery caches
- Cache entries now track referenced tables for selective invalidation
  on INSERT, UPDATE, DELETE, TRUNCATE
- Optimize JoinType::parse with zero-allocation byte-level matching
- Use extend_from_slice instead of extend(iter.copied()) in index code

Nested subquery benchmark: 2390μs → 447μs (5.3x faster)

- Add I64Set with custom pre-mixing hash for i64 keys
- Use WyMix pre-mixing in Value::hash for better FxHash distribution
- Add StringMap/StringSet type aliases (AHash-backed) for String keys
- Convert unbounded thread-local caches to LRU-bounded:
  - Scalar/IN subquery caches: 128 entries
  - Semi-join cache: 256 entries
  - Regex caches: 128 entries
- Add DDL cache invalidation for DROP TABLE/VIEW and ALTER TABLE
- Switch Value-keyed maps from AHash to FxHash (with WyMix pre-mixing)

- Add ValueMap<V> and ValueSet type aliases using AHash in core/mod.rs
- Migrate FxHashMap<Value, _> and FxHashSet<Value> to ValueMap/ValueSet:
  - executor/context.rs: semi-join cache, batch aggregate cache
  - executor/subquery.rs: IN expression sets, batch aggregates
  - executor/aggregation.rs: single-column GROUP BY
  - executor/dml.rs: UPDATE precomputed values
  - executor/expression/*: IN operation types and compilation
  - executor/cte.rs: distinct value tracking
  - executor/index_optimizer.rs: hash set extraction
  - parser/ast.rs: InHashSetExpression
  - functions/aggregate/mod.rs: DistinctTracker
- Add GroupKeyMap<V> type alias in version_store.rs for GroupKey maps
- Migrate bitmap_index.rs to AHash for CompactArc<Value> keys
- Keep FxHash for Vec<Value> GROUP BY (raw_entry_mut compatibility)
- Use AHash for (Value, Value) tuple GROUP BY

AHash provides HashDoS resistance via randomized seeds, important for
user-controlled data like GROUP BY columns and indexed values. Value::hash()
retains WyMix pre-mixing for universal hasher compatibility.

Storage Simplification:
- Remove per-value CompactArc wrapper from Row storage
- Simplify from 3 variants (Shared/Inline/Owned) to 2 (Shared/Owned)
- Store Arc<[Value]> directly instead of Arc<[CompactArc<Value>]>
- Add CompactVec::extend_clone() for optimized slice extension
- String sharing preserved via SmartString::from_string_shared()

Bug Fixes:
- Fix dirty read vulnerability: Database::clone() now creates independent
  executor with its own transaction state instead of sharing
- Fix join projection column ordering: use ColumnSource enum to preserve
  SELECT order regardless of join swap optimization
- Fix ORDER BY + LIMIT with Hash index: return None when index doesn't
  support ordered iteration to fall back to regular query path
- Fix FULL OUTER JOIN null row handling in nested loop join

Parser Improvements:
- Better error messages for NULL bytes in string literals
- Integer overflow now falls back to float parsing
- Handle Error tokens gracefully in expression parser

Performance:
- Reduced memory overhead by removing per-value Arc wrapper (~16 bytes/value)
- extend_clone() is 7x faster than extend(iter().cloned())

Major refactor achieving 33% memory reduction for Value type through
sophisticated niche optimization techniques.

Key changes:

SmartString (24→16 bytes):
- Redesigned from enum to packed struct with tag byte
- Tag values 0-16 used, leaving 17-255 as niches for Value discriminant
- Inline capacity reduced from 22 to 15 bytes (tradeoff for size)

CompactArc DST support:
- Extended to support str and [T] with thin pointers (8 bytes)
- Unified Header with dropper function pointer for type-erased cleanup
- 50% smaller than std::Arc for DSTs

Value (24→16 bytes):
- Now uses CompactArc<str> for Json instead of Arc<str>
- Niche optimization places discriminant in SmartString's unused tag values
- Option<Value> also 16 bytes (no discriminant overhead)

Performance fixes:
- row.rs: Use into_vec() instead of to_vec() to avoid cloning Values
- Index constructors now accept expected_rows for pre-allocation

Panic safety fixes in CompactVec:
- clone(): Added CloneGuard RAII for proper cleanup on T::clone() panic
- extend(): Added ExtendGuard RAII for iterator panic safety
- extend_clone(): Added ExtendCloneGuard RAII for clone panic safety
- from_iter(): Added FromIterGuard RAII for iterator panic safety
- Added comprehensive panic safety tests for all fixed methods

- Introduce Copy-on-Write B+ tree with structural sharing for version store
- Replace RwLock<BTreeMap> with CowBTreeMap enabling O(1) snapshot cloning
- Readers clone tree root (atomic increment) then iterate without holding locks
- Add metadata storage to CompactArc header for O(1) tree size tracking
- Add extend_copy to CompactVec for fast bulk memcpy of Copy types
- Optimize transaction commit to avoid Vec<Box<dyn Table>> for read-only txns
- Rename int_maps → maps and Int64Map → I64Map for consistency

CowBTree improvements:
- Introduce dual refcount system (CompactArc count + drop_count) for
  correct concurrent drop coordination
- Add drop_count (AtomicU32) to NodeHeader for thread-safe V drops
- Implement rightmost split optimization for sequential inserts
- Add Entry API with O(1) get via cached path
- Improve panic safety in deep_clone with incremental set_len
- Add comprehensive SAFETY comments to all unsafe blocks

Performance optimizations:
- Remove unnecessary .clone() after versions.read() in VersionStore
- CowBTree supports lock-free reads without cloning

Documentation:
- Add SAFETY comments to compact_arc.rs, compact_vec.rs, i64_map.rs
- Document thread safety, memory layout, and lifetime guarantees

Fixed a bug in merge_with_left where moving children out of an internal
node via ptr::read would still cause Drop to try dropping child 0
(since for internal nodes, children = len + 1, so len=0 means 1 child).

The fix sets drop_count to 2 before the node is dropped, causing
NodePtr::drop to skip content dropping entirely.

Also added 14 new tests for internal node borrow/merge operations,
increasing line coverage from 93% to 96%.

- test_update_separator_keys: triggers Ok(i) path when updating keys
  that match separator keys in internal nodes
- test_reverse_insert_internal_node_split: ensures split_internal()
  is called via reverse order inserts

- Remove row_id from RowVersion (8 bytes saved per version)
  Row ID is now passed separately where needed
- Remove chain_depth from VersionChainEntry (8 bytes saved)
  Depth is computed on-demand via O(k) traversal (k <= 254)
- Use NonZeroU64/NonZeroUsize for arena indices enabling niche opt
  RowIndex reduced from 24 to 16 bytes (33% smaller)
- Add V2 persistence format without redundant row_id field
  Maintains backward compatibility with V1 format via magic bytes
- Reuse existing Arc keys in bitmap/btree indexes
  Avoids allocation when value already exists in index
- Add CompactArc::from_compact_vec for zero-copy Row conversion
  Avoids intermediate Vec allocation in Row::into_arc

- Version bump to 0.2.4
- Fix import path for ConcurrentI64Map