Performance proof

The best algorithm depends on workload, constraints, data distribution, and required semantics. Optimize only a measured path.

1. Define the claim

State the user-visible performance property, representative workloads, realistic size distribution, worst credible case, required tail or memory bounds, correctness constraints, and stop condition.

Complete when: success and rejection can be decided from measurements.

2. Record a stable baseline

Capture environment, build mode, dataset, warmup, repetitions, variance, and raw results. Confirm the harness exercises the claimed path. If noise is larger than the expected effect, improve the harness before editing.

Complete when: repeated baseline runs are comparable and the workload is representative.

3. Profile the path

Use the repository's profiler, tracing, allocation, query-plan, bundle, or timing tools. Name the hot function, allocation, I/O wait, or repeated computation. State expected time and space complexity for the measured region.