runner Design

The runner keeps mutation local to prepared inputs and sink accumulation. Each scale is materialized once, cloned and prepared per implementation, executed in balanced order, then reset after the complete block.

Timing encloses payload execution and output folding. The clock stops before finish and Bench::keep, so final retention cost is excluded. Report rendering and file IO never enter the measured closure.

Every implementation receives the same configured warmup protocol before calibration. Timed blocks continue one seeded balanced order across exploratory and confirmatory phases, so each implementation occupies every execution position once per complete cycle. Repeated raw observations remain available; robust summaries use their median rather than their mean.

An implementation may provide a blocking synchronize hook for asynchronous devices. The runner invokes it immediately before and after every timed region, so queued work completes inside the measured boundary. Fixture allocation and reset follow SetupTiming; validation, event emission, final sink conversion, and report logging remain outside timing.

Oracle and relational validation execute before timing. OperationResult threads next_context across every sequence, while fixture setup/reset follows the configured lifecycle. Unsafe implementations use native subprocess workers with hard cancellation, timeout, exit status, stdout, and stderr capture. Calibration searches between the configured iteration bounds outside the measurement phase and records elapsed time, target time, and retry count. Exploratory observations are emitted separately from fixed-budget confirmatory observations.

Run summaries retain the semantic, performance, and provenance environment. Performance metadata includes CPU, device, concurrency, clock, GC, and frequency policy; comparisons should reject snapshots for which environment_compatible returns false.