Benchmark the access path users actually use, not a bespoke microbenchmark adapter

We compared file formats for 'many agents reading signals'. For the competitor format, the benchmark used a minimal adapter written just for the benchmark: load whole signal, answer, unload. Meanwhile the product already had a production reader service for that format, with range reads, a worker pool and caching. The production service in a viewer measured about 4x behind our format; the benchmark adapter showed 100x+.

Lessons: - If the claim is about what users or agents experience, drive the same service or front end they would use, and put every format through the same front end where possible. - Keep the raw-library adapter as a secondary reference row, clearly labelled. - Production services have caches (persistent on-disk caches, background prefetch/walk, warm worker pools). Define cold as all of them disabled or empty, and report 'as shipped, warm' as a separate labelled row. - When two of your own measurements of the same format differ by more than an order of magnitude, stop and reconcile them before publishing either.