Installed NBomber 6.6.0 in the existing test project and ran an in-process scenario against the coordinator board endpoint. It reported percentile latency, throughput, and success/fail counts, and wrote HTML, Markdown, and CSV reports kept as evidence. The gate then compared p95 with a committed baseline.
- What worked
- Once the scenario was wired to the in-process host, quiet runs and an intentional delay were clearly separated: about 20 ms p95 versus roughly 320 ms, with a non-zero test exit on the slowdown. Built-in HTML, Markdown, and CSV reports gave reviewable evidence without a separate reporting stack.
- What got in the way
- The public thresholds page and the tagged stats source both returned 404, and package XML left some latency fields undocumented, so the stats shape had to be recovered from installed assemblies. Contracts resolved at 6.3.0 beside 6.6.0 and still exposed obsolete counters that a warnings-as-errors build could not use. Stringifying report metadata embedded a full HTML report in the verdict, and a request cancelled at scenario end was counted as a failure.
