Saturation study records DDSketch max relative value error; optimizer reads mean #175

Problem

The saturation study and rqe-optimizer read different accuracy values for DDSketch (dd):

• scripts/study_saturation.py:59 records DD curves as max_relative_value_error, the worst probe.
• rqe-optimizer/src/lib.rs family_properties (as of rqe-optimizer: check each family by its own accuracy metric #173) checks dd by mean_relative_value_error, the average over probes.

This is more than a different name. RelativeValueErrorGT::score (aqpbm-core/src/accuracy/quantile.rs:137-138) scores the same 100 quantile probes once and records both values. The study keeps one and the optimizer reads the other.

Values each sketch is read by

Impact

SaturationCurves::accuracy asserts that each curve point's error_metric matches the family's metric. With #173, dd is in DEPLOYABLE_FAMILIES, so every default MILP/AutoSketch run with a Quantile RAQE evaluates DD and panics on the mismatch. Rerunning the study as it stands won't help, because the script still writes max_*.

Decision

Switch the study to the mean for DD, so KLL and DD (the two sketches that can answer the same quantile query) are both judged on average error.

Fix

1. scripts/study_saturation.py:57-59: change DD's metric to mean_relative_value_error, and replace the comment "DDSketch's bound holds per quantile, so its curve is the worst one" with the new reason.
2. Rerun the DD saturation study and regenerate its curves.
3. Check that small_problem runs with a Quantile RAQE and no panic.

🤖 Generated with Claude Code