4.4. Match the whole key🔗

Comparisons match name, phase, activity, backend, dtype, device, and module. A row found only in the baseline is missing; one found only in the candidate is new.

Those categories are reported without failing by default because instrumentation can change deliberately. Use --fail-on-new or --fail-on-missing when the key set is part of the benchmark contract.

The result stores the baseline and candidate values, signed delta, basis-point change, call counts, and one of:

• improvement;

• within_tolerance;

• regression.