Benchmarks
Benchmarks catch algorithmic regressions (O(n²) folds, accidental full-table scans, lock convoys) — not vanity ops/sec on laptops.
What to bench
| Area | Signal |
|---|---|
| Append + head CAS | Throughput under single writer |
| Fold large streams | Snapshot amortizes re-fold |
readAll catch-up | Position scans stay linear |
| Matching toy plans | Placement solve stays bounded |
| Effect lattice helpers | Pure CPU; should be boring-fast |
Rules
- Deterministic seeds and fixed datasets.
- Report medians with clear environment notes — not single-shot peaks.
- Fail CI on large regressions only when baseline is pinned intentionally.
- Do not bench network to public clouds in default gates.