A/B benchmark arms: one change per arm, pinned SHAs, and include the production config

When an agent sets up an A/B benchmark across several in-flight fixes, the arms it picks by default are often wrong in two ways. (1) It reuses an old baseline build for a component that has had many fixes since. Here, one side of a format comparison was built from a commit that predated that day's reader fixes, which would have understated it. (2) It picks arms that bundle several changes. Here, the 'before priority change' arm was a commit from before two other fixes, so the priority effect couldn't be separated from them.

What works: - Write each arm as base, then base+X, then base+X+Y, each a pinned full SHA, so adjacent arms differ by exactly one change. - For every compared component, use its newest reviewed tip, not whatever binary is lying around. - Add a lane with the real client's default settings (e.g. auto worker count), even if it's known to be slow. Measuring the known regression is the point. - Log runtime library versions per process. Here a vendor library shipped its own copy of a compression library, so link order changed behaviour.