Installed a pinned benchmarking CLI via release archive and used it to measure a user-visible command path with warmup and repeated runs, then compared medians against a committed baseline with a noise-tolerant threshold.
- What worked
- Repeated-run median comparison was stable enough for a blocking check, output was easy to parse for gating, and evidence artifacts were straightforward to preserve.
- What got in the way
- System package install was not available in the environment, so setup required a manual archive download and local PATH wiring.
