benchmarks

Overview ¶

Package benchmarks holds Harbor's V1 performance-benchmark suite (Phase 79). It is an `_test.go`-only package: the `Benchmark*` functions exercise three of the runtime's hottest seams against the REAL components — no mocks, no stubs (CLAUDE.md §13):

• engine_bench_test.go — engine envelope throughput under N concurrent runs against a single shared *engine.Engine (RFC §6.1).
• bus_bench_test.go — event-bus publish/fan-out latency as a function of subscriber count, against the real `inmem` EventBus driver wired with a real audit redactor (RFC §6.13).
• memory_bench_test.go — memory-strategy AddTurn latency for the `truncation` and `rolling_summary` strategy executors, wired with a real `inmem` StateStore + EventBus (RFC §6.6).

Run the suite with `make bench`; gate a PR against the committed baseline (`docs/perf/baseline.txt`) with `make bench-check`. The regression gate (`scripts/perf/check-regression.sh`) compares a fresh run against the baseline via `benchstat` and fails on a statistically-significant slowdown beyond 10%. See docs/plans/phase-79-performance-benchmarks.md and D-136.