Benchmarking

Benchmarks are fixture-specific evidence, not a universal performance promise. The repository benchmark compares JSON, TSV, SQLite, and G2Lex for source and compiled bytes, cold open time, traced allocations, lookup percentiles, and sequential iteration.

python -m benchmarks.runtime_storage.benchmark \
  tests/fixtures/generic.tsv --format tsv --repetitions 1000

For binary or runtime refactors, record at least asset size, cold open/load, resident memory, lookup latency/throughput, iteration, and package size where relevant. Use deterministic fixtures and lock external sources. Network access must be an explicit opt-in; a full-source gate is preferable to silently benchmarking a partial download.

A structural improvement that causes a meaningful mmap or resident-memory regression requires a documented trade-off and should not be merged on unit tests alone. Interpret results together with fixture size, cache state, and Python version.