eval: accuracy target sweep (90%, 95%, 99%) for AutoSketch vs. ASAP

[zzylol]

Reruns the 12 synthetic §6.3 workloads at 90% and 99% accuracy. The 95% runs are the committed ones in autosketch-vs-asap-hydra.

Changes

• autosketch_vs_asap: --target takes acc90, acc95 (alias p95) or acc99. Each means NN% accuracy in the family's own metric: error ≤ 0.10 / 0.05 / 0.01 for KLL rank, DDSketch relative value and HLL relative error, and top-k precision ≥ 0.90 / 0.95 / 0.99. The 95% behavior is unchanged.
• scripts/plot_autosketch_vs_asap_targets.py: cost vs. target figures and a summary.
• rqe-optimizer/results/autosketch-vs-asap-targets/: 24 result files, figures, and a README with the analysis.

Results (cheapest plan, CPU, vCPU). All runs have 0 RQEs dropped and 0 sanity violations.

Workload	Target	ASAP	ASAP (no roll-ups)	PerQuery	AutoSketch
mixed	90% / 95%	3.54	3.54	10.5	3,716
mixed	99%	3.85	3.85	10.9	3,757
multi-grouping	90% / 95%	0.221	0.551	1.07	10.2
multi-grouping	99%	0.627	0.979	1.49	9.94
mixed + multi-grouping	90% / 95%	3.76	4.09	11.6	3,726
mixed + multi-grouping	99%	4.47	4.82	12.3	3,767

• 90% = 95% everywhere. The 95% plans already use the cheapest measured config of each sketch, e.g. HLL lg_k=12, the smallest on the grid. A looser sweep needs smaller configs measured first.
• 99%. ASAP's cost rises 9% on mixed and 2.8× on multi-grouping. ASAP stays 2.4–2.8× below PerQuery and 16–1,000× below AutoSketch.
• Roll-ups. At 99% they save 36% on multi-grouping, against 60% at 95%.

Tests: the runner's example tests pass (10), including the new levels.

🤖 Generated with Claude Code