v0.3.0

This release is almost entirely about performance. The client was benchmarked
head-to-head against the official Java client (jnats) across core pub/sub,
JetStream, and Object Store workloads, and the hot paths were rebuilt based on
what the profiles showed. There are no public API changes — everything
below is internal.

Performance

Core pub/sub

• The protocol parser keeps a single flat array carry instead of an immutable
  chunk queue, fixing an O(n²) scan under streaming load — 2.4× parser
  throughput and −82% allocations in JMH, ~+40% end-to-end receive (#18)
• Publishes enqueue a small write descriptor instead of building a combined
  header+payload array per message; the single writer fiber assembles batches
  into one reused buffer and flushes them in one socket write — per-publish
  allocation is now constant at 112 B regardless of payload size (#20, #21, #22)
• The receive path builds the user-facing message once, directly in the parser,
  instead of materializing an intermediate frame that gets re-wrapped (#23)
• Sends take a lock-free fast path while the connection is online (#19)
• Sockets are read in 64 KiB requests instead of fs2-io's 8 KiB default,
  cutting async read round-trips 8× on large-payload streams (#26)
• Internal bounded queues stay on cats-effect's lock-free implementation at
  any capacity. Queue.bounded silently degrades to a CAS-retry queue at
  capacities ≥ 65534, which cost >30% of client CPU under sustained load with
  large queueCapacity settings — high-capacity configs are now ~2× faster
  under CPU pressure (#24)

JetStream

• consume processes inbox messages chunk-at-a-time instead of one effect per
  message: 2.1× consume throughput (#25)

Object Store

• GET benefits from the larger socket reads and now hashes chunks without
  copying them — streaming GET reaches parity with jnats at steady state
  (~1.3 GB/s in our benchmarks) (#26)

Other changes

• New documentation site: https://thatscalaguy.github.io/fs2-nats (#17)
• README: expanded feature overview and compatibility matrix
• Apache 2.0 LICENSE file added to the repository
• Test library updates (#16)

Compatibility

• Binary compatible with 0.2.x (MiMa-checked); no source changes required
• Scala 3, cats-effect 3.6, fs2 3.13; JDK 11+

What's Changed

• docs: add sbt-typelevel-site documentation site by @ThatScalaGuy in #17
• Test library updates by @ThatScalaGuy in #16
• perf(parser): flat-array incremental decoder (fixes O(n²) control-line scan) by @ThatScalaGuy in #18
• perf(client): skip the connRef.modify CAS on the Online publish path by @ThatScalaGuy in #19
• perf(transport): reuse the writer coalescing buffer across drains by @ThatScalaGuy in #20
• perf(transport): drop the per-publish Some box from the write queue by @ThatScalaGuy in #21
• perf(publish): enqueue a write descriptor instead of a per-publish combined array by @ThatScalaGuy in #22
• perf(receive): build the user NatsMessage once on the MSG/HMSG data path by @ThatScalaGuy in #23
• perf(queues): keep hot-path queues on cats-effect's lock-free implementations at any capacity by @ThatScalaGuy in #24
• perf(jetstream): process pull-consume deliveries a chunk at a time by @ThatScalaGuy in #25
• perf(receive): 64 KiB socket reads + copy-free Object Store GET digest by @ThatScalaGuy in #26

Full Changelog: v0.2.0...v0.3.0