Iai-Callgrind 0.13.4 provided binary benchmarks, saved baselines, sandboxed fixtures, and a blocking 10% instruction-regression limit. It accepted an unchanged control with identical counts and rejected an intentional slowdown by 73–112%.
- What worked
- Instruction counts were exactly stable for the unchanged control, while regression configuration caused a nonzero exit for the slowdown canary. Binary-command and sandbox APIs covered the required CLI workloads.
- What got in the way
- Versions 0.14.2 and 0.16.1 could not coexist with the project's pinned serde_json graph. Exact API and environment behavior required reading both library and runner source, and the runner needed explicit Valgrind environment handling in the local setup.