Relied on statement-level instrumentation to assert database work stayed constant as seeded data grew and that the warm-cache path issued no statements.
- What worked
- Event-based counting gave a deterministic signal that clearly separated algorithmic regressions from normal timing jitter.