How to diagnose performance issues in APL using measurement techniques?

Suppose you have an APL workspace where a particular function appears to be a bottleneck. What prerequisites (e.g., a recent APL interpreter, access to ⎕TIME or ⎕MONITOR) are needed, and what numbered steps should you follow to:

1. Instrument the code to capture start and end times for each call.
2. Run a representative workload multiple times to gather statistically significant samples.
3. Analyze the collected data (e.g., compute mean, median, outliers) to spot slow paths.
4. Verify that the observed slowdown is not due to measurement overhead or environmental noise.
5. If changes are made to improve performance, how to roll back to the original version and re‑run the benchmark to confirm the effect?