benchmark-first

name	benchmark-first
description	Benchmark converted code against the frozen upstream implementation before making performance claims.
origin	repo-local

Benchmark First

Use this skill after parity has been demonstrated for a converted surface.

Rules

• Compare the upstream implementation and the converted implementation on the same inputs and same machine class.
• Verify parity on the exact benchmark input before timing it.
• Prefer representative real data. If real data is unavailable, document the synthetic fallback clearly.
• Use the narrowest suite, group, and match filters that isolate the edited hot path before widening to larger reports.
• Run repeated measurements and report medians or geometric means, not one-off wall-clock numbers.
• Report time terms first: Python time, native C++ time, time saved per operation, and native C++ time reduction. If a ratio is included, label it as native C++ speedup over Python (python_time / native_cpp_time).
• Report version and environment details with the results.
• If the benchmark points to boundary overhead instead of native arithmetic, load cpp-performance-optimization before changing code.