Performance-review workflow

Purpose

Accept only repeatable CMetrics performance improvements without correctness or workload regressions.

When to use

Use for map lookup/update, codec allocation, cardinality, and hot-path changes.

Investigation and implementation

1. Identify the hot path and representative existing benchmark workload.
2. Establish a correctness baseline and a performance baseline before editing.
3. Change one material factor at a time and preserve ownership/concurrency rules.
4. Add a benchmark workload only when existing modes cannot represent the path.

Validation

Follow benchmarks/README.md. Build Release with CMT_BENCHMARKS=ON; run at least five alternating before/after samples on the same idle machine with the same compiler, flags, CPU policy, cardinality, and operations. Compare medians, variability, allocations when relevant, and perf stat counters. Run scripts/agent-verify.sh for correctness.

Expected report

Report revisions, hardware/software environment, exact commands, raw samples, medians, percentage changes, counter changes, correctness results, and affected workloads. Explain any regression.

Stop conditions

Do not ship noise-level, single-run, differently configured, or correctness- regressing results. Escalate tradeoffs that improve one supported workload while materially degrading another.