Pin PYTHONHASHSEED in the benchmark workflow
Python randomizes string hashing per process, so dict probe sequences differ between two runs of the same code. The instrumented benchmarks are sensitive enough to pick that up as a few percent of drift on a commit that never touched the measured code path, which makes real regressions harder to spot.
Pinning PYTHONHASHSEED: 0 for the workflow removes that source of run-to-run variation.
CodSpeed doesn't call out PYTHONHASHSEED specifically, but this is the same idea as their Keep Benchmarks Deterministic guidance, applied to the interpreter rather than to the benchmark body. Their write-up on unrelated benchmark regressions covers the wider class of environment-driven drift this belongs to.
Only the workflow env changes; no package source is touched.