benchmarking

Run the tier benchmarks via the Makefile targets, then read standings only within a single report since ranks are computed per machine and round. Absolute times never cross platforms, and headline standings come only from the quiet-machine protocol with controlled power state.

The suite compares strata against orjson, msgspec, ujson, and the stdlib across loads, dumps, load, dump, query, and search categories at three size tiers. The deciding sweep recorded 63 of 81 rows at #1, with query and search sweeping every row.

Cross-session runs on an unchanged tree have measured about 30% apart, with Low Power Mode explaining much of it. Rows within roughly 1.00-1.09x are treated as a coin-flip band that only the gated quiet-machine protocol can settle.

The rebuilt regression checker must enforce the contract thresholds of greater than 2% on median and p95 and greater than 5% on RSS. The previous tool's 5%-only threshold was part of its brokenness and must not be reproduced.

No, absolute times never cross platforms or runners. Windows runners vary between EPYC Genoa and Milan processors, swinging rows significantly with no code change, so every row must be read against the report's processor line.

The old tool crashed in main() on an AttributeError, its parser looked for outdated report headers so zero rows parsed, and a stale baseline key never matched. Lessons: test main() end-to-end, share one format definition between writer and parser, and key baselines by full filename.