rqe-optimizer: hydra-kll accuracy when a query merges several windows #166

Found in review of #160 (S4b of #143).

Problem

accuracy_at (#154) reads merge_accuracy at the merge counts on either side of n = L/x, and falls back to the single-window query_accuracy when the row has none. The hydra-kll rows have an empty merge_accuracy. A fixed-size shared sketch folding n windows of every group holds about n times the items in the same grid, so a quantile RAQE with L = 10x can pass accuracy_ok_for while the true rank error exceeds the SLA.

Options

• Measure: have scripts/export_rqe_optimizer_costs.sh record hydra-kll merge accuracy at several merge counts, as rqe-optimizer: check accuracy at the query's merge count #154 does for the other families.
• Refuse until measured: make families with one_fixed_size_sketch_for_all_groups ineligible when n > 1 and merge_accuracy is empty.

Probably both: refuse now and loosen once the rows exist. Re-measure after the no-fan-out hydra-kll change, since fan-out changes the error.

🤖 Generated with Claude Code