benchmark

# Benchmark Run the Perf issue playbook for measurement and repair. `review` inspects a design and its proof needs without starting a measurement. `only` explicitly narrows lanes; mark excluded lanes N/A instead of implying whole-product proof. Discover runners from the affected package, app and `benchmarks/editor`. Inspect each runner's imports, fixtures and output schema. Record its source revision, built artifact identity, actual route, action and environment. Before accepting a changed cache, index, store, subscription, scheduler, geometry owner or repeated work, name the user operation and independent scale variables. Select normal, large and stress cohorts that exercise that operation. Freeze budgets and the noise rule before reading candidate results. Run the baseline first and prove one planted slowdown fails the budget. Compare the same source-built route, fixture, options and real action. Alternate baseline and candidate samples. Report absolute time and relevant work counts beside relative changes. A baseline that refuses input or performs different work is a control, not a performance comparison. Preserve correctness and native behavior before accepting speed. A transform benchmark cannot establish trusted keyboard, caret or paint behavior. Mount and real editing are primary when the user report concerns an editor. Memory, DOM size, layout reads and subscription fan-out are supporting guards or primary metrics only when their cost is the question. Repair measurement identity, stale builds, fixture parity, sample counts, aggregation and missing guards before product code. A pre-acceptance prototype proves a design choice only. Rerun the same contract on final production bytes after the last repair. Use `pstack:benchmark-checklist` before accepting or reporting a measured result. Keep inconclusive decisions open with the next probe; never replace a measured gate with an asymptotic argument.