v0.30.0

• Benchmarks: new sourceips suite measures std/http@v1 source_ips
  against the per-endpoint connection wall — standing keep-alive
  connections (pool: dedicated) under a narrowed ephemeral port range,
  from one source IP vs several loopback addresses (peak established
  connections, connect errors). Linux CI only (it narrows the port range
  via sysctl for the suite and restores it).
• Docker: every image flavor is now its own GHCR package —
  ghcr.io/perfscale/perfscale (perfscale only), perfscale-k6,
  perfscale-jmeter, perfscale-locust, and perfscale-full (all three
  runners), each tagged X.Y.Z / X.Y / latest. Pick exactly the engine
  you need (docker pull ghcr.io/perfscale/perfscale-k6:latest) instead of
  finding the right suffix tag in one package. The old suffix tags
  (perfscale:latest-k6, perfscale:X.Y.Z-full, …) keep being published
  for the same images, so existing pulls don't break.
• std/http@v1 now transparently decodes gzip-compressed responses (Content-Encoding: gzip) — send Accept-Encoding: gzip via headers and the decoded body flows into check: and ${{ feed.body.posts }} selections as plain JSON. The max_response_body_bytes cap applies to the decoded payload.
• Custom WASM steps (RFC 009): a WASM library component can now export steps dispatchable in use: as lib/<alias>/<step> — the component receives the interpolated with: params plus the step context (VU id, iteration, run settings) and returns the standard action result (success / output / tagged logs / metrics / optional http_sample and secret_output_paths), so it folds into check:, thresholds, and the run summary like any built-in. lib/ is a reserved namespace (never shadows std/*/pro/*), perfscale lint validates lib/ references against each component's probed manifest, and the run summary's libraries: line gains step_calls / step_errors / step_p95_ms next to the token-call counters. A new process capability (triple-gated: the YAML grant, allow_library_capabilities: true, and allow_process_actions: true) gives step components host-mediated one-shot process execution with bounded output, timeout, and process-tree cleanup — never raw WASI process imports. New ABI: perfscale:step@0.1.0 (wit/step/step.wit), additive on top of the untouched perfscale:library@0.3.0.
• Benchmarks: SyTester is now a default competitor engine — whenever its
  jar and Java 21+ are available, its rows sit in the throughput,
  saturation, grpc and db tables next to k6/locust/JMeter,
  offering SyTester's maximum fixed rate (SYT_RATE, 10000 TPS) since it
  has no VU model. The grpc row uses its gRPC generator (proto set uploaded
  over the REST API), the db row its PgJDBC transport. A new APPS knob
  (bench workflow input apps, default k6 jmeter locust sytester) picks
  the competitor engines; perfscale's rows always run. Unknown names in
  SUITES or APPS now fail the run (exit 2) instead of being silently
  skipped.

What's Changed

• feat(bench): add SyTester rows to throughput, saturation, grpc and db tables by @vitalics in #23
• feat(core): custom WASM steps (RFC 009) — lib// dispatch by @vitalics in #25
• feat(http): transparent gzip response decoding by @vitalics in #26
• feat(docker): publish each image flavor as its own GHCR package by @vitalics in #22
• feat(bench): sourceips suite — source_ips vs the per-endpoint connection wall by @vitalics in #24
• Release v0.30.0 by @vitalics in #27

Full Changelog: v0.29.0...v0.30.0