performance-profiler

# Performance Profiler ## Overview Make slow things fast — correctly and provably. This skill enforces a measure-driven loop: never optimize on a hunch, always profile to find the real hot path, fix the biggest contributor first, then re-measure to confirm the win and guard against regressions. Keywords: performance, profiling, optimization, bottleneck, latency, throughput, p99, slow, benchmark, flamegraph, CPU profile, memory leak, allocations, N+1 query, caching, big-O, complexity, hot path, regression. The cardinal rule: **measure first**. Most "obvious" optimizations target the wrong code. Profilers routinely show that 90% of time sits in a place nobody suspected. ## Workflow Follow this loop. Do not skip steps — especially step 1 and step 6. 1. **Define the goal and a metric.** Pick ONE primary metric and a target: wall-clock latency (p50/p95/p99), throughput (req/s, rows/s), CPU time, peak memory (RSS), or allocations. Write down the current value and the target. "Make it faster" is not a goal; "cut p95 from 800ms to under 200ms" is. 2. **Reproduce reliably.** Build a repeatable scenario with representative data volume. A bottleneck at 10 rows may vanish at 10M and vice-versa. Disable noise: warm caches, JIT warmup, fixed input, quiet machine, multiple runs. 3. **Measure the baseline.** Time/benchmark the whole operation before touching anything. Save the numbers. Use `scripts/bench.py` for a quick statistically sane wall-clock benchmark...