[zzylol]

Why

Since #171 and #179, the planner reads sketch accuracy from the saturation study's curves. Where the study measured nothing, the accuracy is unknown and the deployment ineligible:

• past an unsaturated curve;
• past the measured merge counts or N;
• with no merge curve for a point.

Each algorithm has a published guarantee that holds there, independent of N.

What

• New rqe_optimizer::theory::bound: each sketch's guarantee at 95% confidence, in the metric the optimizer checks it by.
  • KLL: normalized rank error 2.296 / k^0.9723 (DataSketches' 99% value), scaled to 95% by sqrt(ln 20 / ln 100), since KLL's error scales with sqrt(ln(1/δ)). It survives merges.
  • DDSketch: relative value error α.
  • HLL: 1.96 · 1.04 / sqrt(2^lg_k).
  • CMS-heap top-k (unmerged) on Zipf(θ, K): the share of the top k keys whose probability beats the (k+1)-th's by more than e/w. It needs 1 − e^−rows ≥ 95%.
• SaturationCurves::accuracy falls back to the bound for a bracketing point at or above its first N when the curves give nothing. New accuracy_with_source reports AccuracySource::{Measured, Theory}.
• These stay unknown: a point below the first N, a shape outside the grid, a sketch with no published bound, and a merged top-k answer. Merged top-k gets heap = m·k in a follow-up PR.
• Docs: rqe_sketch_deployment_v1.md eligibility.

Verification

• cargo test -p rqe-optimizer (65): bound values, the top-k rows/merge conditions, and sketches without a bound.
• clippy -D warnings and fmt are clean.

🤖 Generated with Claude Code