Changes

0.2.5

Performance

The bitmop layer has been rewritten around ByteBuffer-based primitives and JVM intrinsics, delivering substantial performance gains across the board with no changes to the public API.

mask-offset, mask-width, and bit-count now compile to single hardware instructions (TZCNT, POPCNT) via Long/numberOfTrailingZeros and Long/bitCount, replacing the previous O(n) loops.

A new ByteBuffer abstraction (uuid->buf, buf->uuid) provides direct typed access at byte offsets via native getLong/putLong operations, eliminating manual shift/mask loops for serialization.

Representative speedups over 0.2.0:

Category	Speedup
to-byte-array	57x
to-hex-string	29x
v3 (MD5) generation	9.0x
v5 (SHA1) generation	6.0x
v8 (custom) generation	4.2x
v1 (time-based) generation	1.5x
v6 (time-based) generation	1.4x

Combined generate + serialize operations see 3-19x end-to-end improvement depending on UUID version and serialization format.

v3/v5 generation now uses a fused digest pipeline with ThreadLocal ByteBuffer reuse, ByteBuffer/wrap on digest output, and inlined version/variant bit stamping -- eliminating all intermediate allocations and var lookups. v5 is now at parity with JUG 5.2 (~260 ns vs ~254 ns).

Correctness

• get-clk-seq now uses bitmop/ldb to extract the 14-bit clock sequence directly, fixing incorrect results that occurred when Java's .clockSequence() method threw on non-v1 UUIDs.

Clock improvements

• The Gregorian monotonic clock now uses AtomicLong with compareAndSet on a packed long (50-bit millis + 14-bit seqid), replacing atom + swap! + per-call State allocation.
• v1 and v6 constructors inline the CAS loop and bit-field packing, eliminating var lookup, function dispatch, and ldb/dpb overhead on the hot path.
• The monotonic counter initial state is now seeded with a cryptographically random 10-bit value instead of zero, avoiding predictable first values after library load.
• Per-tick counter reseed uses 10-bit entropy (was 8-bit).

New UUID versions

• v6 (reordered time-based, lexically sortable)
• v7 (unix time-based, cryptographically secure, lexically sortable)
• v7nc (unix time-based, non-cryptographic, maximum throughput)
• v8 (custom / user-defined)
• Max UUID sentinel (+max+, max, max?)

v7nc uses ThreadLocalRandom and a per-thread monotonic counter instead of SecureRandom and a global AtomicLong. At ~39 ns/op, it is 1.26x faster than JUG 5.2's TimeBasedEpochGenerator (~50 ns/op).

New protocol members

• get-instant -- returns a java.util.Date for time-based UUIDs
• get-unix-time -- returns POSIX millis for v1, v6, and v7
• get-clk-seq -- returns the 14-bit clock sequence for v1 and v6
• to-hex-string -- 32-character hex encoding (no dashes)
• to-uri -- returns java.net.URI in URN format
• max? -- predicate for the max UUID
• UUIDRfc9562 protocol (with UUIDRfc4122 as a backward-compatible alias)

New constructors

• v0 / null -- null UUID constructor
• max -- max UUID constructor
• v4 two-arity form (v4 msb lsb) -- stamps version and variant bits onto caller-supplied words
• Multi-arity =, <, > comparison operators

Extended polymorphism

• as-uuid now accepts java.net.URI, byte arrays, and URN strings in addition to canonical UUID strings
• uuidable? predicate for testing coercibility

Build and packaging

• Added deps.edn for tools.deps / CLI users
• Removed .travis.yml; CI now uses GitHub Actions
• Updated to Clojure 1.12.0 and primitive-math 1.0.1
• Comprehensive test suite: 138 tests, 1.4M+ assertions, 0 failures
• Test coverage: 94% forms, 93% lines (via cloverage)

Documentation

• README rewritten to cover v6, v7, v8, max UUID, and the new performance characteristics
• Typo corrections throughout (get-timestamp, approximately, associated, requirements, consciousness, alphabetically, elapsed, duplicate, construction)