v6.0.0

• Breaking (Concurrency) — Multi-core load generation is now the default. v5 ran every load fiber on a single thread; v6 sizes the execution context to min(cpu_count, connections) OS threads and exposes it as --workers N. --workers 1 reproduces the v5 single-core path; the cpu default respects CRYSTAL_WORKERS
• Breaking (Accuracy) — Threshold gates now evaluate corrected latency by default (--latency-correction on). A run that passed --max-p99 in v5 can fail in v6 when the target could not keep up with --rate; corrected and uncorrected latency are identical when --rate is not set. Pass --latency-correction off for v5 gate semantics
• Breaking (Exit codes) — Exit codes are now a taxonomy instead of a single failure code: usage/config errors moved from 1 to 2, an unreachable target (zero HTTP responses plus at least one transport error) moved from 1 to 3, and signals now exit 130 (SIGINT) / 143 (SIGTERM) instead of 0. 1 is reserved for breached thresholds, and a breached threshold takes precedence over a signal code

Code	Meaning
0	Run finished, no threshold breached
1	A threshold was breached
2	Usage/config error
3	Target unreachable
130	Interrupted (SIGINT)
143	Terminated (SIGTERM)

• Breaking (Reporting) — Transport error categories are a fixed, normalized vocabulary instead of Crystal exception class names: dns_failure, connect_refused, connect_timeout, connect_failed, read_timeout, write_timeout, request_timeout, connection_reset, connection_closed, tls_error, proxy_error, protocol_error, other. Consumers matching on IO::TimeoutError or Socket::ConnectError must switch to the new names
• Breaking (JSON) — JSON output gained a top-level schema_version, currently 1. Every documented key is always present and is null when not applicable, so consumers can stop probing for key existence
• Breaking (CLI) — --progress now defaults to on only when stderr is a TTY, so CI logs are clean without passing --no-progress. Explicit --progress still forces it on
• Breaking (Toolchain) — Minimum Crystal is now 1.21.0 (was 1.19.0); multi-core load generation uses Fiber::ExecutionContext.default.resize
• Performance — --workers 8 -c 128 reaches a median 329,440 req/s against a SO_REUSEPORT target on loopback, versus 109,161 req/s on the single-thread --workers 1 path — a 3.0x gain (AMD Ryzen AI 9 465, 20 threads, 2-byte body, median of 5 three-second runs). --workers 16 measures slower (280,079 req/s) than --workers 8 on the same box, which is why the default is min(cpu cores, connections) rather than every core
• Performance — Response bodies are streamed through one fixed buffer per worker instead of being materialised as a String, so peak RSS no longer scales with response size: 22.9 MB at --workers 8 -c 128 with 2-byte bodies, 12.9 MB at --workers 8 -c 32 with 100 KB bodies
• Accuracy — Added coordinated-omission correction. v5 measured latency only from the moment a request was actually sent, so a stalled target made the numbers look better: in a regression test that froze the target for 1 second during -q 500 -d 5, v5 reported p99 = 0.37 ms while silently dropping 490 of 2500 scheduled requests. v6 reports the outage in corrected_latency_ms.p99 and surfaces the loss as rate.skipped_requests
• Accuracy — Added a rate block reporting requested vs attained req/s, attainment percent, scheduled and skipped requests, and schedule drift, so a rate-limited run can no longer claim a rate it never achieved
• Accuracy — Added send_delay_ms percentiles, the gap between a request's scheduled send time and its actual send time
• Accuracy — --timeout now sets the write timeout as well as connect and read; v5 left writes unbounded
• Accuracy — -q / --rate accepts fractional rates (-q 0.5)
• CI/CD — Added a full threshold matrix: --max-p50, --max-p75, --max-p90, --max-p95, --max-p999, --max-avg, and --max-latency alongside the existing --max-p99
• CI/CD — Added --min-rps N to fail a build when throughput drops below a floor
• CI/CD — Added --fail-on-rate-miss to exit 1 when the attained rate is more than 1% below the requested --rate
• CI/CD — Added repeatable per-endpoint gates via --url-threshold "PATTERN METRIC VALUE", where PATTERN is a substring matched against the target URLs and METRIC is one of max-p50 max-p75 max-p90 max-p95 max-p99 max-p999 max-avg max-latency max-fail-rate min-rps. A pattern matching no target URL is a config error (exit 2)

cryload --urls-file urls.txt -d 30 \
  --url-threshold "/api/users max-p99 120" \
  --url-threshold "/api/search max-p99 400" \
  --min-rps 2000

• CI/CD — JSON output gained a thresholds block listing every evaluated gate with its scope, metric, comparator, limit, actual value and pass state, plus a verdict block carrying the exit code and reason
• Load testing — -d/--duration, --timeout and --warmup accept duration units: 30, 30s, 2m, 1h30m, 500ms. A bare number is still seconds
• Observability — Added a latency phase breakdown (phases_ms) for dns, connect, tls, ttfb and total, so a slow run can be attributed to handshake cost versus server think time. dns/connect/tls are per connection, so with keep-alive their count is the number of connections opened
• Observability — Added a per-status breakdown (by_status) with latency percentiles per HTTP status code, so error responses can be distinguished from slow successes
• Observability — Added a per-URL breakdown (by_url) for multi-URL runs, with per-endpoint throughput, failure rate and percentiles. It is null for single-URL runs and for target lists over 1000 URLs
• Observability — Text output header gained Workers:, Latency correction: and Request timeout: lines, and the report gained Corrected Latency Percentiles, Rate, Latency Phases, Per-URL and Thresholds blocks
• Observability — CSV output appends the new v6 columns after the existing v5 columns in their original order, so header-indexed consumers keep working unchanged
• Resilience — Added --request-timeout DURATION, a total wall-clock deadline for one request including body download. It catches slow-trickle responses that never trip the per-operation --timeout and reports them as request_timeout
• Resilience — Added a file-descriptor preflight: cryload checks RLIMIT_NOFILE (POSIX only) and raises the soft limit toward the hard limit when it is below connections + 64. If it still cannot get there, it exits 2 naming the current and required limits instead of failing mid-run with connect errors
• Documentation — Rewrote docs/json-output.md for schema version 1, including every null-value rule and a v5 migration section
• Documentation — Added docs/examples.md, a cookbook of runnable commands covering request modes, bodies, auth, multi-URL, rate limiting, phase diagnosis, CI gates and exit-code handling
• Documentation — README performance table replaced with measured numbers and a methodology note; the inconsistent ~50,000 req/sec claim is gone, and a new Honest latency section explains coordinated omission
• Documentation — Updated docs/comparison.md with rows for coordinated-omission correction, latency phase breakdown, per-URL/per-status breakdown, per-endpoint gates, exit code taxonomy and total request deadline