Choose the question before the timer
Pick a gallery operation that matters: time to prepare one selected asset, total batch throughput, peak accepted payload, or cancellation completion latency. Those are different metrics. A faster checksum batch is not automatically a more responsive renderer.
Record machine, compiler, standard library, build mode, input, worker count, warmup policy, repetitions, and what the measurement includes. Keep a correctness checksum outside the timing decision so optimization does not remove the work or change its meaning unnoticed.
The companion reports peak active work as an invariant observation, not a benchmark. Its tiny workloads may be dominated by thread startup and coordination. Use the observation to verify a bound, and use an appropriately sized repeatable workload before making timing claims.
Write one measurement sentence with a metric, unit, workload, and excluded costs. Reject a second sentence that makes a broader claim than the experiment supports.