Optimizer: accuracy after merging n windows is unvalidated #766

Problem

Accuracy eligibility uses sketch-bench measurements of a single sketch instance. A query merges n = L/W instances, and error after merging is not validated. This applies to both the ASAPQuery optimizer and sketch-bench#129. Measurements are also taken on the profiled (synthetic Zipf) workload, not the user's data.

Change

Measure or bound accuracy as a function of merge depth (e.g. an accuracy profile keyed by config, data volume, and number of merged instances) and use it in eligibility.

Part of #753.

🤖 Generated with Claude Code