- Embed small scalars directly in the hash slot, avoiding value allocations for tiny entries
- Switch compression from FastLZ to LZ4 (better ratio and speed)
- get() accepts an optional $default argument (returned on miss instead of NULL)
- dump() gains an $offset parameter and reports hits, atime, embedded and c_len per entry
- Per-entry hit counters and start_time exposed in info()/dump()
- Multi-get omits missing keys (filled with $default when given) instead of false placeholders
- Field-by-field slot commits and spinlock backoff (pause/yield) reduce contention overhead
- Fixed a startup hang when values_memory_size is below the 4M segment minimum
- Fixed Windows shared memory cleanup (view unmapped exactly once)
- Fixed valgrind-reported leaks and uninitialized reads (MINFO output, compression error paths)